Volume 1 photovoltaic solar energy 1 05 – historical and future cost dynamics of photovoltaic technology

1.05

Historical and Future Cost Dynamics of Photovoltaic Technology

GF Nemet and D Husmann, University of Wisconsin–Madison, Madison, WI, USA
© 2012 Elsevier Ltd. All rights reserved.

1.05.1
1.05.2
1.05.2.1
1.05.2.2
1.05.2.3
1.05.2.4
1.05.2.5
1.05.2.6
1.05.2.7
1.05.2.8
1.05.2.8.1
1.05.2.8.2
1.05.2.9
1.05.2.10
1.05.3
1.05.3.1
1.05.3.1.1
1.05.3.1.2
1.05.3.2
1.05.3.3
1.05.3.3.1
1.05.4
1.05.4.1
1.05.4.1.1

1.05.4.1.2
1.05.4.1.3
1.05.4.1.4
1.05.4.1.5
1.05.4.2
1.05.5
1.05.5.1
1.05.5.2
1.05.5.3
1.05.6
References

Introduction: Observed Reductions in the Cost of Photovoltaics
What Caused the 700Â Reduction in the Cost of PV?
Identifying Drivers of Change
R&D and Efficiency Improvements
Sequential Niche Markets
Expectations about Future Demand
Learning by Doing
Intertechnology Spillovers
Materials
Drivers Related to Supply and Demand
Industry structure
Demand shocks and rising elasticity
Quality and Product Attributes
Interactions between R&D and Experience in Production
Using Learning Curves to Predict Costs
Use of Experience Curves in Modeling and Policy
Modeling
Policy

Problems with Using Experience Curves
How Reliable Are Experience Curve Predictions?
Assessing the significance of recent deviations
Nonincremental Cost-Reducing Developments
Identifying Breakthroughs
Defining breakthrough
First pass: Expert opinion
Why patent analysis?
Backward citation analysis
Implementing backward citation analysis for PV
Results: Combining Expert Opinion and Patent Analysis
Modeling Nonincremental Changes in PV
An Approach to Modeling Nonincremental Technological Change
Results for Nonincremental Technological Change
Summary of Nonincremental Modeling
Future Progress and Development

47

48

49

49

49

50

51

51

51

52

52

52

53

54

54

54

55

55

55

57

58

61

61

61

62

62

62

63

65

67

67

68

69

69

70


1.05.1 Introduction: Observed Reductions in the Cost of Photovoltaics
The cost of photovoltaics (PV) has declined by a factor of nearly 700 since the 1950s, which is more than that for any other

energy technology in that period. At present, however, PV remains a niche electricity source and in the overwhelming
majority of situations does not compete economically with conventional sources, such as coal and gas, or even with other
renewable sources, such as wind and biomass. The extent to which the technology improves over the next few decades will
determine whether PV reaches terawatt scale and makes a meaningful contribution to reducing greenhouse gas emissions or
remains limited to niche applications.
In this chapter, we discuss the observed cost reductions in PV and the factors affecting them. We discuss the use of learning curves
for forecasting PV costs and the nonincremental changes to technology that complicate such models. We include a discussion of an
alternate forecasting methodology that incorporates R&D impacts and conclude with implications for policy and items for future
progress.
Reductions have occurred across a wide variety of components within PV systems. Foremost, the cost of PV modules have
declined from about $2700 W−1 in the 1950s to around $3 W−1 in 2006 (Figure 1) [1]. Although difficult to verify, claims that
the marginal cost of manufacturing modules is as low as $1 W−1 are now widespread [2].

Comprehensive Renewable Energy, Volume 1

doi:10.1016/B978-0-08-087872-0.00103-7

47


48

Economics and Environment

$10 000

2008$ W–1

$1000
$100

$10
$1
1950

1960

1970

1980

1990

2000

2010

2000

2010

−1

Figure 1 Cost of PV modules, 1957–2006 (2008$ W ) [1].

1000

2008$ kWh–1

100
10

1
0.1
0.01
1950

1960

1970

1980

1990

−1

Figure 2 Levelized cost of electricity generated from PV (2008$ kWh ) [1].

Most assessment of PV has focused on the evolution of the core technology, the modules that convert sunlight to electricity. As
the prices of modules have fallen, the rest of the components that comprise a PV system have accounted for an increasing share of
the overall costs. A study of balance-of-system (BOS) prices in the 1990s found that the rate of technology improvement in BOS
components has been quite similar to that of modules [3], which is rather surprising given the heterogeneous set of components
and activities that fall under the rubric of BOS. Inverters, the largest hardware component of BOS, improved more slowly than
modules; it was the noninverter costs – installation, wiring, mounting systems – that improved even more quickly. Dispersion in the
data for BOS components is higher than that for modules.
Ultimately, because technology adoption decisions are based on the cost of electricity produced, the cost of fully installed
systems is what matters most. Recent work at Lawrence Berkeley National Laboratory has documented the cost of installed systems
in the 2000s [4]. Costs have fallen from over $10 W−1 in 1998 to under $8 W−1 in 2007. Figure 2 shows the long-term decline in the
cost of electricity from PV, a factor of 800 reduction.

1.05.2 What Caused the 700Â Reduction in the Cost of PV?

A variety of factors, including government activities, have enabled the nearly 3 orders of magnitude reduction in the cost of PV over
the past six decades. Despite this achievement, the technology still remains expensive, such that widespread deployment depends on
substantial future improvement. No single determinant predominantly explains the improvement to date; R&D, economies of scale,
learning by doing, and knowledge spillovers from other technologies have all played a role in reducing system costs. Moreover,
interactions among factors that enable knowledge feedbacks – for example, between demand subsidies and R&D – have also proven
important.
But not all of the factors were important for the entire sequence of the technology’s development. Certain factors dominated for
periods. These stages have been correlated with shifts in the geographic foci of effort, both by the private sector and by the
government: in the 1970–85, R&D to improve efficiencies and manufacturing techniques; in the 1990s to early 2000s, long-term
demand programs to enable economies of scale; and in the 2000s, efforts to stimulate local learning by doing to reduce installation
costs in addition to continuous improvements through manufacturing scale [5]. As alternatives to crystallized silicon PV emerge,
efforts to improve them follow a similar cycle.


Historical and Future Cost Dynamics of Photovoltaic Technology

1.05.2.1

49

Identifying Drivers of Change

Nemet [6] sought to understand the drivers behind technical change in PV by disaggregating historical cost reductions into
observable technical factors. That study spanned the period of nascent commercialization in the mid-1970s through the early
2000s. During the 26-year period studied, there was a factor of 20 reduction in the cost of PV modules.
The results of that study point to two factors that stand out as most important: plant size accounts for 43% of the change in PV
cost and efficiency accounts for 30% of the change. Of the remaining factors, the declining cost of silicon accounts for 12%. Yield,
silicon consumption, wafer size, and polycrystalline share each have impacts of 3% or less. These observed changes are summarized
in Figure 3. The following sections discuss the sources of these observed changes. Table 1 summarizes this discussion.

1.05.2.2

R&D and Efficiency Improvements

The doubling of the average electrical conversion efficiency of PV systems since the 1970s has been important to cost reductions,
accounting for about a third of the decline in cost over time. R&D, especially public sector R&D, has been central to this change
(Figure 4). Data on the highest laboratory cell efficiencies over time show that of the 16 advances in efficiency since 1980 [7], only 6
were accomplished by firms that manufacture commercial cells. Most of the improvements were accomplished by universities, none of
which would have learned from experience with large-scale production; government and university R&D programs produced 10 of the
16 breakthroughs in cell efficiency. Almost every one of the 20 most important improvements in PV occurred during a 10-year period
between the mid-1970s and the mid-1980s [8], most of them in the United States where over a billion was invested in PV R&D during
that period [9]. Section 1.05.4 discusses a novel methodology for identifying these types of nonincremental improvements.

1.05.2.3

Sequential Niche Markets

Deployment of PV has benefited from a sequence of niche markets where users of the technology were less price-sensitive and had
strong preferences for characteristics such as reliability and performance that allowed product differentiation. Governments have
played a large role in creating or enhancing some of these niche markets. In the 1960s and 1970s, the US space program and the

$25


$25.30

2002$ W–1

$20

$15

43%

$10


30%

$5


12%

$3.68

3%

3%

Water
size

Si
used

2%

2%

5%

$0

1979
price

Si
Plant Efficiency
price
size

Yield

PolyUn2001
x-stal explained price

Figure 3 Portion of cost reduction in PV modules accounted for by each factor [1].

Table 1
Summary of effects of items on the cost of PV (1980–2001) and reasons for the change
in each factor

Item

Share change in PV module
costs attributable
(%)

Main drivers of change in each factor

Plant size
Efficiency
Silicon cost
Wafer size
Si used
Yield
Polycrystalline share
Other factors

43
30
12
3
3
2
2
5

Expected future demand and risk management
R&D, some LBD for lab-to-market
Spillover benefit from information technology industry
Strong LBD
LBD and technology spillover
Strong LBD
New process, LBD possible
Not examined

LBD, learning by doing.

50

Economics and Environment

30

400

Govt.
R&D

Efficiency (%)

300
20
Highest laboratory
200

15
10

Avg. commercial

100

R&D (2003$ million)

25

5
0

0
1940

1950

1960

1970

1980

1990

2000

Figure 4 Improvements in energy conversion efficiency of PV and US public investment in PV R&D [1].

Sales (2002$ million)

40

35.8

30
22.5

19.9

20

Satellites
Terrestrial

10
0.9
0
1974

1979

Figure 5 Shift in market from space to terrestrial applications from 1974 to 1979 [10].

Department of Defense accounted for more than half of the global market for PV (Figure 5). The high cost of electricity in space
allowed PV to be competitive even at an early stage. Subsequent markets – telecom repeater stations, off-grid homes, and especially
consumer electronics such as toys, calculators, and watches – were, importantly, independent of government decisions, and allowed
the industry to expand from the mid-1980s until the late 1990s when energy prices, and thus alternative energy, were low social
priorities. From the 1990s onward, households that had strong preferences for environmental protection, especially conspicuously,
created larger markets.

1.05.2.4

Expectations about Future Demand

Increasing demand for PV has reduced costs by enabling opportunities for economies of scale in manufacturing. Nemet [6]
assembled empirical data to populate a simple engineering-based model identifying the most important factors affecting the cost
of PV over the past three decades. The study found that three factors account for almost all of the observed cost reductions: (1) a 2
orders of magnitude increase in the size of manufacturing facilities, which provided opportunities for economies of scale (Figure 6);

(2) a doubling in the electrical conversion efficiency of commercial modules; and (3) a fall in the price of the primary input

Output per plant (MW

yr–1)

15

10

5

0
1975

1980

Figure 6 Size of PV manufacturing facilities [1].

1985

1990

1995

2000


Historical and Future Cost Dynamics of Photovoltaic Technology

51

material, purified silicon. Because investments in larger facilities take time to pay off, economies of scale depend on expectations of
future demand. As a result, public programs that reduce uncertainty by setting clear long-term expectations, such as Japan’s Sunshine
Program in the 1990s [11], are more effective at enabling scale economies than generous subsidies that can suddenly disappear,
such as California’s incentives for wind and solar in the early 1980s. Government programs with longer time horizons such as
Germany’s in the 2000s and the recently launched California Solar Initiative create similar opportunities [1, 12]. Japan’s program
was especially innovative in that it not only took a long time horizon but also set a declining subsidy such that it fell to zero after
10 years of the program. This provided not only expectations of demand but also clear expectations of future levels of subsidy.
Germany’s program is especially notable in that production there has become sufficient to create external economies of scale –
the emergence of machine tool manufacturers that now produce equipment specifically for the PV industry [13]. Similarly,
production of lower purity, thus cheaper, solar-grade silicon is now profitable because plants can be built at large scale. We also
have begun to observe some economies of scale in unit size, where large installations show much lower per-watt costs [4]. Increasing
installation size is likely to become more important as economies of scale reduce module-manufacturing costs, leaving installation
costs as an increasingly large share of system costs.
The main drivers of the change in plant size over the period were growth in expected future demand and the ability to manage
investment risk. Whether experience plays a role in enabling the shift to large facilities depends on new manufacturing problems at
larger scales and how experience may help in overcoming these problems. Examples from three PV firms indicate that limited
manufacturing experience did not preclude rapid increases in production. Mitsubishi Electric expanded from essentially zero produc­
tion in 1997 to 12 MW and as of 2000 planned to expand to 600 MW by 2012 [14]. While the firm had decades of experience in
research and satellite PV applications, its cumulative production was minimal. It began substantial manufacturing activity only with
the opening of its Iida plant and its entry into the Japanese residential PV market in 1998. Similarly, Q-Cells, a German firm, began
producing cells only in 2001 with a 12 MW line and increased production to 50 MW in only 2 years [15]. Sharp considered
construction of a 500 MW yr−1 plant in 2006, which would amount to a 10-fold expansion in the firm’s capacity in only 5 years. By
2011, it had increased capacity to 2.8 GW yr−1. Note that by mid-2011, six firms had manufacturing capacities above 2 GW yr−1. In the
rapid expansions of the past 10 years, the ability to raise capital and to take on the risk of large investments that enable construction of
large manufacturing facilities appears to have played more important roles than learning by experience in enabling cost reductions.
These results support the claim that “sometimes much of what is attributed to experience is due to scale” [16].

1.05.2.5

Learning by Doing

Learning from experience in production has played a role, albeit not a dominant one, in reducing module costs. These changes
occurred in several different processes. Experience in manufacturing led to lower defect rates and the utilization of the entire wafer area,
which increased yields. Experience was probably important in enabling the growth of larger crystals and the formation of longer
conductors from cell edges to electrical junctions; savings accrued from the ensuing larger wafer sizes. Less silicon was consumed as
experience helped improve sawing techniques so that less crystal was lost as sawdust and thinner cells could be produced. The
development of wire saws, a spillover technology from the radial tire industry, is less clearly related to experience. Learning by doing
helped the gradual shift to polycrystalline ingots. Casting of rectangular multicrystalline ingots was a new technology that partially
derives from experience with the Czochralski process for growing individual crystals. While learning by doing and experience play
more important roles in these factors, together they account for only 10% of the overall change in module cost [6].
Learning by doing plays a much more important role in reducing installation costs [17]. An important aspect of this learning is
that it is a local phenomenon, whereas module production is truly global [18]. As installation costs become a large portion of costs,
the extent of this learning by doing will be important. Whether or not the benefits of this learning are appropriable will determine
whether governments need to play a role in promoting learning investments. More generally, the global aspect of module
manufacturing suggests that interfirm and international technology spillovers are likely to be more important in module produc­
tion than in installation. Finally, learning by doing may be important in the translation of laboratory breakthroughs to commercial
products, as observed in Figure 4.

1.05.2.6

Intertechnology Spillovers

Like many technologies, PV has benefited from the adoption of innovations that originated in other industries. These include the
use of excess purified silicon from the chip-making industry, the use of wire saws from radial tires to slice multiple silicon wafers,
electronic connectors to ease installation, screen-printing techniques from lithography, as well as an array of manufacturing
techniques taken from microprocessors (for crystallized silicon) and liquid crystal displays (LCDs; for thin film).

1.05.2.7

Materials

Reductions in the cost of purified silicon were a spillover benefit from manufacturing improvements in the microprocessor industry.
Until the 2000s, the PV industry accounted for less than 15% of the world market for purified silicon [19]. Through that time, the PV
industry did not purify its own silicon but instead purchased silicon from producers whose main customers were in the much larger
microprocessor industry, where purity standards were higher. Therefore, experience in the PV industry was irrelevant to silicon cost
reductions. More recently, input costs, especially of purified silicon, increased in the mid-2000s in line with other commodity prices


52

Economics and Environment

[20]. This change has had the beneficial effect of creating strong incentives to reduce wafer thickness and find ways to conserve
materials. Global recession in 2008–09 has seen commodity prices drop. However, the benefits of lower silicon utilization remain.
Moreover, the shift to PV-specific silicon production has increased the potential for scale economies at lower levels of purity.
If lifetimes and efficiencies can be maintained at these lower levels of purity, there is strong potential for materials costs to fall
beyond the short-term business cycle effects.

1.05.2.8

Drivers Related to Supply and Demand

Changes in demand and in market structure have affected prices as well, even if they do not directly affect the technical
characteristics of the technology.

1.05.2.8.1

Industry structure

Changes in industry concentration have affected market power and have led to a changing relationship between prices and costs
over time. Market share data indicate a decline in industry concentration during this period. This change typically produces an
increase in competitiveness, a decrease in market power, and lower profit margins. For example, there were only two US firms
shipping terrestrial PV from 1970 to 1975 [21, 22]. In 1978, about 20 firms were selling modules and the top 3 firms made up 77%
of the industry [23]. By 1983, there were dozens of firms in the industry, with the largest three firms accounting for only 50% of the
megawatts sold [24].
One way to quantify this change is to use the Herfindahl–Hirschman index (HHI), which provides a way of measuring industry
concentration [25, 26]. The HHI is calculated by summing the squares of the market shares of all firms in an industry. The maximum
possible HHI is 10 000. The data show a trend to a less concentrated US market during the 1970s (Figure 7). Concentration in the
global market remained stable in the 1990s, the period for which comprehensive worldwide data are available. The increase in
international trade in PV over the last three decades indicates that the relevant scale of analysis shifted from a national market in the
earlier years to an international market today.

1.05.2.8.2

Demand shocks and rising elasticity

Demand dynamics have shifted prices in opposite directions. Foremost, the surge in subsidy programs for PV in the 2000s resembles
a demand shock, as observed in other sectors. The industry has had a difficult time adjusting quickly. The very high levels of demand
in the 2005–07 period are in part to blame for the high prices during that period. This led to a reversal in the multidecade downward
cost trajectory that can be observed in Figure 9. Some of this cost increase was due to higher materials costs [20] and the rest likely
due to higher willingness to pay as aggressive subsidy programs brought new consumers to the market for PV.
In contrast, historically, the shift from satellites to terrestrial applications affected prices because of a difference in the demand
elasticity of the two types of customers. Price data from that period provide some supporting evidence. In 1974–79, the price per
watt of PV modules for satellites was 2.5 times higher than that for terrestrial modules [10]. The impact of this price difference on
average PV prices is calculated by taking into account the change in market share mentioned below. In this period, the combination
of these price and market shifts accounts for $22 of the $28 price decline not explained by the model. Satellite customers, with their
hundreds of millions of dollars of related investments, almost certainly had a higher willingness to pay for PV panels than early
terrestrial applications such as telecom repeater sites or buoys for marine navigation. The difference in quality must account for

some of the price difference. But the difference in willingness to pay may also have led to higher differences between cost and price
for satellite than for terrestrial applications.
Another historical explanation for the change in cost is that changes in production methods occurred due to an increase in the
number of customers and the types of products they demanded. There was a shift away from a near-monopsony market in the early
1970s. Originally a single customer, the US space program, accounted for almost all sales. Conversely, in 1976, the US government

5000
US market

HHI

4000
3000

Highly
concentrated

2000
1000


Moderately
concentrated
Unconcentrated


0
1970

1975

Figure 7 Concentration in the PV industry (HHI) [1].

Global market
1980

1985

1990

1995

2000

2005


Price of PV electricity (log scale)

Historical and Future Cost Dynamics of Photovoltaic Technology

Space, communications,

navigation,

pipelines


53

Rural off-grid

Consumer
Resid. off-grid products
Green on-grid
Resid. on-grid
Current price of PV electricity

Utility scale

Market size (log scale)
Figure 8 Illustrative demand curve for PV electricity. Reproduced from Nemet GF (2009) Interim monitoring of cost dynamics for publicly supported
energy technologies. Energy Policy 37(3): 825–835 [28].

accounted for only one-third of terrestrial PV purchases [27]. With the rise of the terrestrial industry, a larger set of customers
emerged over the course of the decade. When this change in the structure of demand occurred, one result was the shift away from
producing customized modules, such as the 20 kW panels on Skylab, to producing increasingly standard products at much higher
volumes. Figure 8 provides a schematic example of what a demand curve for PV electricity might look like considering both
historical data and projections of the future opportunity. The x-axis shows the size of each market segment. The y-axis shows that the
price customers are willing to pay for PV electricity in each segment. Note the use of log–log axes. For example, in the upper left
region, the cost of electricity is very high in off-grid and remote locations, although markets are relatively small. At the other end of
the curve, markets for wholesale electricity are very large, but require PV to compete with large central station sources of electricity.

1.05.2.9

Quality and Product Attributes

Changes in quality and product attributes also affected costs. During the 1970s, the market for PV modules shifted toward terrestrial
applications. Whereas in 1974, 94% of PV systems were manufactured for applications in space, by 1979, the market share for space
had fallen to 36%. The shift led to a reduction in the quality of modules, which rendered certain characteristics nonessential,

allowing manufacturers to switch to less costly processes.
First, space and weight constraints on rockets required high-efficiency panels to maximize watts delivered per square meter. The
relaxation of this requirement for terrestrial applications enabled manufacturers to employ two important cost-saving processes
[10]. Cheaper materials were now tolerable. Modules could use the entire area of the silicon wafer – even the portions near the
edges, which tend to suffer from defects and high electrical resistivity. Also, the final assembly process could use a chemical polish to
enhance light transmission through the glass cover, rather than the more expensive ground optical finish that was required for
satellites. Second, reliability targets fell. Satellite programs, such as Vanguard and Skylab, needed satellite PV modules that would
operate reliably without maintenance, perhaps for 20 years. Terrestrial applications, on the other hand, could still be useful with
much shorter lifetimes. The rapid growth in the terrestrial market was the main driver of this change.
There are three assumptions that are commonly made when applying the experience curve model using prices rather than costs:
that margins are constant over time, that margins are close to zero with only minor perturbations, and that margins are often
negative due to forward pricing. Yet changes in demand and industry structure are important in that they erode support for these
three assumptions. Indeed, earlier work pointed out that firms’ recognition of the value of market domination, particularly during
incipient commercialization, leads to unstable pricing behavior [29]. An implication of the variation in the price–cost margin is that
industry structure affects the learning rate. In the case of an industry that becomes more competitive over time, such as PV, a
price-based experience curve overestimates the rate of technical progress.
One solution for addressing this problem would be for future work to obtain real cost data where possible. An alternative would
be to use an approach in which costs can be derived from prices and market shares using the assumptions in Cournot competition
that a firm’s profit margins decrease as the number of firms in the market increases and that a single firm’s profit margins increase
with that firm’s market share [30]. However, comparisons of competing technologies are best made on the basis of prices, not costs,
since prices reflect what a consumer faces in deciding whether to adopt a technology and which to adopt. A more general approach
would be to incorporate market dynamics into predictions of technological change: industry concentration, market power, and
changes in elasticity of demand affect prices. The HHI analysis described above shows that concentration is not stable over time,
especially if international trade is taken into account. The assumptions of perfect competition and prices equal marginal cost are too
strong in the early stages of the product life cycle when the technology is improving rapidly, industry structure is unstable, and new
types of customers are entering the market.


54

Economics and Environment

1.05.2.10

Interactions between R&D and Experience in Production

While there do appear to be periods when one factor was dominant in supporting innovation, it is also the case that the interaction
of various factors was important. In particular, Japan in the 1990s had both strong demand-side policies and support for R&D [31].
In this case, it was not so much efficiency breakthroughs and alternative cell designs that drove improvements, but rather support
from the Japanese government, via the Ministry of International Trade and Industry (MITI). This enabled coordination of
expectations and sharing of best practices, which enabled manufacturing improvements. The influence of the US federal govern­
ment may have played a similar role in enabling the 1970s breakthroughs, not just through providing resources from R&D but by
creating a sense of commitment that convinced many to work on the technical and market challenges associated with commercia­
lizing this nascent technology [32].

1.05.3 Using Learning Curves to Predict Costs
The potential for future cost reductions, combined with the magnitude of the potential impact of very inexpensive PV, has created a
strong demand for tools that enable prediction of future costs. The experience curve has emerged as the most important of these
tools. The following section surveys the use of experience curves in policy and modeling, caveats in their application, assessment of
their reliability, and the implications for policy makers [33]. While this section focuses on the historically dominant design in the
industry, crystalline silicon PV, more recent work has looked at other approaches, such as thin films [34] and organics (discussed
below). Also, note here that we use the terminology experience curve (defined above), rather than the more specific concept of the
learning curve, which, strictly defined, focuses on labor costs and within-firm knowledge accumulation.
Characterizations of technological change have identified patterns in the ways that technologies are invented, improve, and
diffuse into society [35]. Studies have described the complex nature of an innovation process in which uncertainty is inherent [36],
knowledge flows across sectors are important [37], and lags can be long [38]. Perhaps because of characteristics such as these,
theoretical work on innovation provides only a limited set of methods with which to predict changes in technology. The learning
curve model offers an exception.
Experience curves have been assembled for a wide range of technologies. While there is broad variation in the observed rates of
‘learning’, studies do provide evidence that costs almost always decline as cumulative production increases [16, 39–42]. The roots of

these microlevel observations can be traced back to early economic theories about the importance of the relationship between
specialization and trade, which were based in part on individuals developing expertise over time [41]. The notion of the experience
curve varies from the more specific formulation behind the learning curve in that it aggregates from individuals to entire industries
and from labor costs to all manufacturing costs. In the literature on experience curves, the technological ‘learning’ refers to a broad
set of improvements in the cost and performance of technologies, not strictly to the more precise notion of learning by doing [43].
An experience curve for PV modules is shown in Figure 9, that is, a double-logarithmic graph of PV module price as a function of
cumulative capacity.

1.05.3.1

Use of Experience Curves in Modeling and Policy

PV module prices (2006$ W–1)

Experience curves have become a widely used tool both for models of future energy supply and to inform the design of public
policies related to PV. For example, they provide a method for evaluating the cost-effectiveness of public policies to support new
technologies [44] and for weighing public technology investment against environmental damage costs [45]. Energy supply models
now also use learning curves to implement endogenous improvements in technology. Prior to the 1990s, technological change was
typically included as an exogenous increase in energy conversion efficiency or ignored [46]. Studies in the 1990s began to use the
learning curve to treat technology dynamically [47, 48], and since then, it has become a powerful and widely used model for

100
1976
1980
1990

10

2000

2006

1
1

10

100
1000
Cumulative capacity (MW)

10 000

Figure 9 Experience curve for PV modules, 1976–2006. Reproduced from Nemet GF (2009) Interim monitoring of cost dynamics for publicly supported
energy technologies. Energy Policy 37(3): 825–835 [28].


Historical and Future Cost Dynamics of Photovoltaic Technology

55

projecting technological change. Recent work, however, has cautioned that uncertainties in key parameters may be significant [49],
making application of the learning curve to evaluate public policies inappropriate in some cases [50].

1.05.3.1.1

Modeling

The rate and direction of future technological change in energy technologies are important sources of uncertainty in models that
assess the costs of stabilizing the climate [51]. Treatment of technology dynamics in integrated assessment models has become

increasingly sophisticated [52] as models have incorporated lessons from the economics of innovation and as increased processing
power and improved algorithms have enabled optimization of phenomena such as increasing returns, which in the past had made
computation unwieldy [53]. Yet the representation of technological change in large energy-economic models remains highly
stylized relative to the state of the art of understanding about the economics of innovation [54]. Perhaps one reason for the lag
between the research frontier for the economics of innovation and the modeling of it has to do with incompatibilities in the
methodological approaches of the two fields. On the one hand, research on the economics of innovation has tended to emphasize
uncertainty [55], cumulativeness [38], and nonergodicity [56]. The outcomes of this line of inquiry, which dates back to Schumpeter
[35], and even Marx [57], have often been characterized by richness of description, a case study approach, and, arguably, more
progress with rigorous empirical observation than with strong theoretical claims. On the other hand, optimization and simulation
models require compact quantitative estimation of parameters, with uncertainties that do not become infinite once propagated
through the model. One of the few concepts that have bridged the epistemological gap between the economics of innovation and
the integrated assessment of climate change is the experience curve. Experience curves provide a way to project future costs
conditional on the cumulative quantity of capacity produced. The resulting cost predictions are less deterministic than those
generated by temporal-based rates of technological change, but they are also not simply scenarios, internally consistent descriptions
of one possible future state of technology; they are conditional predictions.

1.05.3.1.2

Policy

Large programs and deviations from trends in cost reductions are challenging policy makers to make decisions about whether,
when, and how much to stimulate the development of energy technologies that have high external benefits. The net benefits of
subsidies and other incentives programs depend heavily on the extent to which technologies improve over time. Experience curves
provide a way for policy makers to incorporate technology dynamics into decisions that involve the future costs of technologies.
They are now used widely to inform decisions that involve billions, and even trillions, of dollars in public funding. The general
notion that learning from experience leads to cost reductions and performance improvements is well supported by a large array of
empirical studies across a variety of technologies. But the appropriateness of using experience curves to guide policy is less uniformly
acknowledged. Despite caveats in previous work, the cost projections that result from experience curves are typically used without
characterizing uncertainty in those estimates. As a result, experience curves are now employed widely to inform decisions that
involve billions of dollars in public funds. They have been used both directly – as graphical exhibits to inform debates – and

indirectly – as inputs to energy-economic models that simulate the cost of achieving environmental goals. Much of the early work to
translate the insights from experience curve studies to energy policy decisions is included in a study for the International Energy
Agency [49]. Other studies have used the tool directly to make claims about policy implications [44, 45, 58].
As mentioned above, energy-economic models that minimize the cost of energy supply now also include experience curve
relationships to include technology dynamics. Model comparison studies have found that models’ estimates of the social costs of
policy are sensitive to how technological change is characterized [51]. Working Group III of the Intergovernmental Panel on Climate
Change (IPCC) used results from a variety of energy-economic models to estimate the magnitude of economically available
greenhouse gas emissions in its Fourth Assessment Report [59]. The results of this assessment are widely used to inform national
climate change policies, as well as the architecture for the next international climate policy regime. In the 17 models they review,
some form of experience curve is used to characterize technological change in at least 10 of those models. See Table 11.15 of IPCC
[59]. Another influential report in 2006, the Stern Review on the Economics of Climate Change [60], relied heavily on experience curves
to model technological change. This report has been central to the formation of climate policy in the United Kingdom and has
played a role in debates in the United States as well, both at the federal level and in California. The International Energy Agency
relies on experience curves in its assessment of the least cost method for meeting greenhouse gas reduction targets and energy
demand for 2050 [61]. Note that the ‘learning investments’ that result from the analyses in this report are estimated in a range of
$5–8 trillion. Debates about subsidies and production requirements for ethanol also use historical experience curves as a justifica­
tion for public support of the production of biofuels [62]. At the state level, experience curves have provided one of the most
influential justifications for a $3 billion subsidy program for PV [12]. Experience curves have also been used in economic models of
the cost of meeting California’s ambitious greenhouse gas reduction targets [63]. Finally, in decisions by the 24 states that have
passed renewable portfolio standards, debates include discussions of how mandatory renewables deployment will bring down its
cost [64, 65].

1.05.3.2

Problems with Using Experience Curves

Demand among policy makers for rigorous, transparent, and reliable tools with which to predict future costs continues to be high.
Yet an array of studies, which are cited below, warn about the limitations of experience curves.

56

Economics and Environment

The learning curve model operationalizes the explanatory variable experience using a cumulative measure of production or use.
Change in cost typically represents the dependent variable and provides a measure of learning and technological improvement.
Learning curve studies have experimented with a variety of functional forms to describe the relationship between cumulative
capacity and cost [66]. The log-linear function is most common perhaps for its simplicity and generally high goodness-of-fit to
observed data. The central parameter in this learning curve model is the exponent, defining the slope of a power function, which
appears as a linear function when plotted on a log–log scale. This parameter is known as the learning coefficient (b) and can be used
to calculate the progress ratio (PR) and learning ratio (LR) as shown below. C0 and Ct are unit costs at time t = 0 and time t,
respectively, and Q0 and Qt represent cumulative outputs at time t = 0 and time t, respectively.
�
Ct ¼ C0

Qt
Q0

�– b

PR ¼ 2 – b
LR ¼ ð1 – PRÞ
Wene [67] has developed a cybernetic theory that predicts an LR of 20%. Several studies have criticized the learning curve model,
especially in its more general form as the experience curve. Dutton and Thomas [16] surveyed 108 learning curve studies and
showed a wide variation in learning rates, leading them to question the explanatory power of experience. Figure 10 combines their
learning rate data with those in a more recent survey of learning rates by McDonald and Schrattenholzer [68]. The learning rate for
PV, 0.23, lies near the mode of the distribution. Argote and Epple [69] explored this variation further and proposed four alternative
hypotheses for the observed technical improvements: economies of scale, knowledge spillovers, and two opposing factors,
organizational forgetting and employee turnover. Despite such critiques, the application of the learning curve model has persisted,
without major modifications, as a basis for predicting technical change, informing public policy, and guiding firm strategy. Below,

the advantages and limitations of using the more general version of the learning curve for such applications are outlined.
The experience curve provides an appealing model for several reasons:
1. Availability of the two empirical time series required to build an experience curve – cost and production data – facilitates testing
of the model. As a result, a rather large body of empirical studies has emerged to support the model. Compare the simplicity of
obtaining cost and production data with the difficulty of quantifying related concepts such as knowledge stocks [70] and
inventive output [71]. Still, data quality and uncertainty are infrequently explicitly addressed and as shown below can have a
large impact on results.
2. Earlier studies of the origin of technical improvements, such as in the aircraft industry [39] and shipbuilding [40], provide
narratives consistent with the theory that firms learn from past experience.
3. Studies cite the generally high goodness-of-fit of power functions to empirical data over several years, or even decades, as
validation of the model.
4. The dynamic aspect of the model – the rate of improvement adjusts to changes in the growth of production – makes the model
superior to forecasts that treat change purely as a function of time.
5. The reduction of the complex process of innovation to a single parameter, the learning rate, facilitates its inclusion in large
optimization and simulation models.
The combination of a rich body of empirical literature and the more recent applications of learning curves in predictive models
has revealed weaknesses that echo earlier critiques.

Frequency

15

10

5

0
–1

–0.5

0
PV
Learning rate

0.5

1

Figure 10 Frequency distribution of learning rates calculated in 156 learning curve studies. The learning rate for PV, 0.23, lies slightly above the mode
of the distribution. Reproduced from Nemet GF (2009) Interim monitoring of cost dynamics for publicly supported energy technologies. Energy Policy
37(3): 825–835 [28].


Historical and Future Cost Dynamics of Photovoltaic Technology

57

1. The timing of future cost reductions is highly sensitive not only to changes in the market growth rate but also to small changes in the
learning rate. Although a coefficient of determination for an experience curve of >0.95 is considered a strong validation of the
experience curve model, variation in the underlying data can lead to uncertainty about the timing of cost reductions on the scale of
decades. Two world surveys of PV prices [72, 73] produce learning rates of 0.26 and 0.17. What may appear as a minor difference
has a large effect. For example, assuming a steady industry growth rate of 15% per year, consider how long it will take for PV costs to
reach a threshold of $0.30 W−1, an estimate for competitiveness with conventional alternatives. Just the difference in the choice of
data set used produces a crossover point of 2039 for the 0.26 learning rate and 2067 for the 0.17 rate, a difference of 28 years.
McDonald and Schrattenholzer [68] show that the range of learning rates for energy technologies in general is even larger. Neij et al.
[50] find that calculations of the cost-effectiveness of public policies are sensitive to such variation. Wene [49] observes this
sensitivity as well and recommends an ongoing process of policy evaluation that continuously incorporates recent data.
2. The experience curve model gives no way to predict discontinuities in the learning rate. In the case of PV, the experience curve
switched to a lower trajectory around 1980. As a result, experience curve-based forecasts of PV in the 1970s predicted faster

technological progress than actually occurred [3]. Discontinuities present special difficulties at early stages in the life of a
technology. Early on, only a few data points define the experience curve, while at such times decisions about public support may
be most critical. Early work in economics is skeptical about the assumption that historically observed rates of learning can be
expected to continue in the future. Arrow [43] argued that that learning is subject to “sharply diminishing returns”. Looking at
studies within single plants, Hall and Howell [74] and Baloff [75] find that learning rates become essentially flat after a relatively
short amount of time – approximately 2 years in these studies. Some have suggested that as a result a cubic or logistic function
offers a more realistic functional form than a power function [76].
3. Studies that address uncertainty typically calculate uncertainties in the learning rate using the historical level of variance in the
relationship between cost and cumulative capacity. This approach ignores uncertainties and limitations in the progress of the
specific technical factors that are important in driving cost reductions [49]. For example, constraints on individual factors, such as
theoretical efficiency limits, might affect our confidence in the likelihood of future cost reductions.
4. Due to their application in planning and forecasting, emphasis has shifted away from learning curves based on employee
productivity and plant-level analysis toward experience curves aggregating industries and including all components of operating
cost. While the statistical relationships generally remain strong, the conceptual story begins to look stretched, as one must make
assumptions about the extent to which experience is shared across firms. In the strictest interpretation of the learning-by-doing
model applied to entire industries, one must assume that each firm benefits from the collective experience of all. The model
assumes homogeneous knowledge spillovers among firms.
5. The assumption that cost reductions are only determined by experience, as represented by cumulative capacity, ignores the effect
of knowledge acquired from other sources, such as from R&D or from other industries. Earlier, Sheshinski [77] wrestled with the
separation of the impact of two competing factors, investment and output. Others have addressed this limitation by incorporat­
ing additional factors such as workforce training [78], R&D [79, 80], and the interactions between R&D and diffusion [31].
Colinearity among the explanatory variables requires large and detailed data sets, the scarcity of which has so far limited
widespread application of these more sophisticated models.
6. Experience curves ignore changes in quality beyond the single dimension being analyzed [81]. The dependent variable is limited
to cost normalized by a single measure of performance – for example, hours of labor per aircraft, dollars per watt, or cents per
megabyte. Performance measures like these ignore changes in quality such as aircraft speed, reliability of power generation, and
the compactness of computer memory.

1.05.3.3

How Reliable Are Experience Curve Predictions?

Given these issues, how well do experience curves perform? Throughout the critiques, a primary concern is the issue of unacknow­
ledged uncertainty. Although studies have cautioned that policy makers must contend with discontinuities and uncertainties in
future learning rates, few do; the cost projections that result from experience curves are typically estimated without acknowledging
uncertainty. Yet a wide array of studies now have pointed to serious reservations about using experience curve projections to inform
policy decisions. Wene [49] emphasized the ways that experience curves could be used to design subsidy programs, but cautioned
about the key uncertainties in parameters because “small changes in progress ratios will change learning investments considerably”.
Concerned about the scale of this uncertainty problem, one study concluded “we do not recommend the use of experience curves to
analyze the cost effectiveness of policy measures” and recommended using multiple methods instead [82]. More recently, Neij [83]
compared experience curve projections to those based on bottom-up models, as well as expert predictions, and found that they
“agree in most cases”. However, in some cases, large uncertainties that emerge from the bottom-up analyses are “not revealed” by
experience curve studies. Rubin et al. [84] indicate that early prototypes often underestimate the costs of commercially viable
applications so that costs rise. Koomey and Hultman [85] have documented a more persistent form of this cost inflation effect for
nuclear reactors. Addressing PV specifically, Borenstein [86] argues that experience curve-based analyses do not justify government
programs because they conflate multiple effects and ignore appropriability concerns.


58

Economics and Environment

Three sources of uncertainty complicate experience curve predictions. First, there is the typical dispersion in learning rates caused
by imperfect correlations between cumulative capacity and cost. Sark [87] explores the effects of this ‘r-squared’ variation to calculate
an error around the learning rate. Inconsistencies in the chosen system boundaries, for example, geographic scope, may introduce
some of this variation. The second source has to do with whether historically observed learning rates can be expected to continue in
the future. Even in his seminal work on learning by doing, Arrow [43] argued that that learning is subject to “sharply diminishing
returns”. Looking at studies within single manufacturing facilities, Hall and Howell [74] and Baloff [75] find that learning rates
become essentially flat after a relatively short amount of time – approximately 2 years in these studies. As a result, some have
suggested that a cubic or logistic function offers a more realistic functional form than a power function [76]. The third source of

uncertainty derives from the choice of historical time period used to calculate learning rates. The timing issue captures variation in
the source data, as well as changes in the slope over time.
Studies have characterized the effects of uncertainty. A prominent study showed that calculating the error in the PR could be used
to develop a range of learning rates to use for sensitivity analysis in policy modeling [88]. Nemet [28] assessed these sources of
uncertainty using a simple and transparent model of the costs of subsidizing technologies until they are competitive with
alternatives. Those calculations include (1) the learning rate, (2) the year at which the cost of a subsidized technology approaches
a target level, and (3) the discounted cost of government subsidies needed to achieve that level. The first result from that study is that
there is a wide dispersion in learning rates depending on what time period was used. That study estimated the learning rate for PV in
each of the 253 time periods of 10 years or greater between 1976 and 2006. Figure 11 plots these learning rates by the year at which
each time series ends. For example, the values shown for 1995 include all 11 time series that end in 1995. This set of values indicates
the range of learning rates that would have been available to an analyst using experience curves to project costs in 1995. The data
begin in 1985 because that is the first year for which 10 years of historical data (1976–85) are available. The data reveal two features
about the trend in calculated learning rates: (1) there is a negative time trend, the mean of the learning rate values has decreased over
time by approximately 0.005 per year, and (2) the dispersion in learning rate values around the annual mean has increased over
time. The dispersion includes an oscillation with maxima in 1995 and 2006.
The second result from that study was estimation of the break-even year using the dispersion in learning rates described above.
The target cost for PV modules used in this example is Pa = $1 W−1 [89]. A 73% subsidy on actual 2006 prices is needed for
consumers’ costs to equal this target. Figure 12(b) shows distributions of the estimated years at which the price of PV will equal that
of this competing technology. Descriptive statistics for these distributions are shown in Table 2 for all time series and for all series
that end in 2006. The median crossover year for all series, ta = 2034, occurs 14 years earlier than the estimates using only data
through 2006, ta = 2048. Note that the dispersion has also increased with the more recent data set.
The third result from that study was to calculate the learning investment required to subsidize PV to the crossover point. The median
cost to subsidize PV is $62 billion when using all time series and $163 billion when using only the time series that end in 2006. Note
that a difference in median learning rate of 40% leads to a difference in median program costs of between a factor of 2 and 3. The
dispersion in costs has also become large; the range from the 5th percentile to the 95th percentile spans an order of magnitude.
Furthermore, notice that costs around the 95th percentile become very large, rising to the tens of trillions. Slow learning has nonlinear
effects on cost and leads to very expensive subsidy programs – even when these future costs are discounted to present values.
Figure 12 summarizes these results. Figure 12(a) shows the distribution of learning rates for all 253 periods (black columns).
The white columns show the distribution of rates using only those series that end in 2006. The latter is the data set one would expect
a contemporary planner to use. The median of the distribution of learning rates from all 253 time series (LR = 0.21) is substantially

higher than the median of the series ending in 2006 (LR = 0.15), although this difference is not significant.

1.05.3.3.1

Assessing the significance of recent deviations

One immediate policy implication of these results is that the possibility of very expensive subsidy programs makes early identification of
such a scenario important. The price escalation in the mid-2000s prompts the question, “do recently observed costs represent a

0.3

Learning rate

0.25
0.2
0.15
0.1
0.05
1975

1980

1995
1985
1990
End year of learning interval

2000

2005

Figure 11 Learning rates for PV (1976–2006) calculated for all periods ≥10 years (n = 253). Reproduced from Nemet GF (2009) Interim monitoring of
cost dynamics for publicly supported energy technologies. Energy Policy 37(3): 825–835 [28].


Historical and Future Cost Dynamics of Photovoltaic Technology

59

(a)
All series from 1976 to 2006 (n = 253)
Series ending in 2006 (n = 22)

Frequency

15

10

5

0

(b)

0

0.05

0.1

25

0.3

0.35

0.4

All series from 1976 to 2006 (n = 253)
Series ending in 2006 (n = 22)

20
Frequency

0.15
0.2
0.25
Learning rate

15
10
5
0

2010

2020

2030

2040 2050 2060
Break-even year

2070

2080

2090

(c) 30
All series from 1976 to 2006 (n = 253)

Frequency

25

Series ending in 2006 (n = 22)

20
15
10
5
0

0

100

200

300
400
Cost to subsidize to breakeven ($ billion)

>500

Figure 12 (a) Calculated learning rates for PV; (b) year at which the price of PV equals that of competing technology; (c) present value of cost to
subsidize PV until it equals the cost of competing technology. Reproduced from Nemet GF (2009) Interim monitoring of cost dynamics for publicly
supported energy technologies. Energy Policy 37(3): 825–835 [28].

significant deviation from the historical trend or does historical variation explain them?” Nemet [28] introduces two methods for
addressing this question. First, recent costs are compared to the confidence interval for the power function resulting from the dispersion
in past observations. Second, these costs are compared to the set of all possible experience curve forecasts made over time. The first
method uses straightforward statistics to examine whether recent variation fits within the confidence interval for observations around
the power function. This variation is caused by the imperfect fit of the power function to the experience curve data [87]. Here a
confidence interval is constructed for the data through 2003. This range is compared to the most recent 3 years of data, 2004, 2005, and
2006, to determine whether they fit within the range defined by projecting the experience curve for 3 years. The data from 1976 to 2003
have r2 = 0.98 and LR = 0.22. The variation around the experience curve power function using least squares yields a 95% confidence
interval around the LR of 0.22 Æ 0.01. Projecting the experience curve to the capacity reached in 2006 (E2006) yields a 95% confidence
interval of expected costs in 2006 of $1.58–$2.51. The actual value for 2006, $3.74, lies outside this range.
The second approach assumes the perspective of a policy analyst making ex ante forecasts each year, incorporating new data as
they become available. This approach assesses whether recent observations could have been projected by the set of all possible
historical forecasts. To illustrate, Figure 13 shows the predictions, over time, of the price of PV for the cumulative capacity that was
reached in 2006, E2006. The first result is that none of the 231 possible projections for 2006 would have predicted a level at or above
the actual 2006 price. Next, this method is used to project prices for the cumulative capacities reached in all years from 1986 to
2006. In Figure 14, the range in gray represents the full range of forecasts for the capacity that was reached each year. For example,
the gray range for 2006 includes all of the 231 data points portrayed in Figure 13. Actual prices each year are shown as a line with


60

Economics and Environment

Descriptive statistics for distributions of experience curve results for PV

Table 2

Cost to breakeven
($ billion)

Learning rate

Break-even year

0.25
0.21
0.14
0.03

2028
2034
2049
8

38
62
175
229

For all time series ending in 2006 (n = 22)

5th percentile
0.21
Median
0.15
95th percentile
0.08
σ
0.04

2034
2048
2082
15

59
163
2172
713

For all time series (n = 253)
5th percentile
Median
95th percentile
σ

Reproduced from Nemet GF (2006) Beyond the learning curve: Factors influencing cost reductions in photovoltaics.
Energy Policy 34(17): 3218–3232 [6]; Nemet GF (2009) Interim monitoring of cost dynamics for publicly supported
energy technologies. Energy Policy 37(3): 825–835 [28].

4

$ W–1

3

2

1

0
1975

1980

1985
1990
1995
End year of learning interval

2000

2005

Figure 13 Trend in predictions of PV prices for the capacity levels reached in 2006. The dashed line shows the actual value in 2006. Reproduced from
Nemet GF (2009) Interim monitoring of cost dynamics for publicly supported energy technologies. Energy Policy 37(3): 825–835 [28].

15

$ W–1

10

5

Actual prices
0
1975

1980

1985

1990

1995

2000

2005

Figure 14 Price projections for the cumulative capacities reached in all years from 1986 to 2006. The gray region shows the range of all forecasts for the
price of PV at the cumulative capacity reached each year. Actual prices are shown as a line with white circles [28].

white circles. The second result is that, other than two individual occurrences, the only time the actual prices have consistently fallen
outside the range of all possible learning rate-derived price forecasts was in 2004–06.
The outcome of this analysis concurs with that of the confidence interval analysis: the recent deviations in PV fall outside the
range of historical precedent. While further analysis is certainly needed to characterize the sources and persistence of these
deviations, these methods may be useful as a preliminary screen to identify near-term deviations that merit further investigation.
An important aspect to account for – given the prominent role described earlier in this chapter – is the role of niche markets. In

Historical and Future Cost Dynamics of Photovoltaic Technology

61

particular, empirical analysis of the level of willingness to pay in niche markets and the size of these markets will add insight into the
extent to which they reduce the cost of subsidy programs.
These results indicate (1) a need for policy makers to more explicitly consider uncertainty in cost projections and (2) the
importance of the development of better tools to identify the significance of near-term deviations from projections. These results
suggest that projected subsidy costs are highly sensitive to timing of the data used: both ‘when’ the forecast was made and the
‘duration’ of the historical data set used. The high dispersion in costs – and especially the skewness of the distribution toward high
values – emphasizes the importance of interim monitoring of technological improvement. These results should not be surprising
given the results above show that learning derived from experience is only one of several explanations for the cost reductions in PV.
Its role in enabling changes in the two most important factors identified in this study – plant size and module efficiency – is small
compared to those of expected future demand, risk management, R&D, and knowledge spillovers. This weak relationship suggests
careful consideration of the conditions under which one should rely on experience curves to predict technical change.
The results have two normative conclusions for policy makers. First, if policy makers are to rely on future cost projections derived
from experience curves, they need to be explicit about the reliability of predictions. Policy decisions should be made acknowledging
the observed variation in the rates of technological improvement over time. Given the current state of knowledge about what
actually causes variation in learning rates, policy makers would do well to consider learning as a stochastic process – that is, that
some aspects of the process remain unpredictable [90]. In this respect, learning is similar to the outcomes of R&D investments; they
are inherently uncertain despite improvements in understanding about R&D productivity [91]. Important further work on this topic
involves assessment of whether this uncertainty is likely to diminish over time as more observations are obtained, as suggested by
the central limit theorem. It is unclear whether the results in this study so far support such a notion. In Figure 11, it appears that the
dispersion in learning rates has increased over time. Even if such a convergence were to occur, a practical issue for policy makers
would be whether it will occur quickly enough to inform decision making and whether the cumulative capacity required for
convergence is small relative to the size of the world energy demand.
Second, devising ex ante methods to identify the significance of near-term deviations in technology cost and performance trends
is essential. How should policy makers respond to situations such as those in Figure 14 in which recent prices appear to be deviating
from the experience curve path? Are these short-term deviations driven by supply bottlenecks, or are they representations of the

lower limits on cost? Deviations make policy difficult; policy makers need to be vigilant against encountering the extremely
expensive outcomes found above. For example, debate over subsidies amounting to several billion dollars in the 2007
Independence and Security Act in the US Energy Information Administration (EIA) suggests that programs involving hundreds of
billions will be subject to scrutiny [92]. The social value that these technologies have the potential to deliver is substantial, but may
take many years to realize. As a result, policy makers may also need to defend technology support against competing social priorities
when deviations are actually short-term aberrations. How can near-term data be used to assess confidence in longer-term projec­
tions? The methods developed in this section offer some avenues for analysis, but ultimately better tools will be required.
The inclusion of experience curves in models that optimize and simulate the costs of climate policy has enhanced their realism.
Given the vast set of results showing that energy technologies improve over time, incorporating experience curves represents a
substantial improvement over omitting them and implicitly assuming a learning rate of zero. But the results summarized here
indicate that a broader set of influences than experience alone contributed to the rapid cost reductions in the past; one implication is
that experience curves overestimate the technical improvements that should be expected to accrue from deployment alone. As a
result, these findings support the efforts by modelers to explore ways of incorporating explanatory variables other than cumulative
capacity, especially when nonincremental changes occur. The following sections describe some attempts to deepen our under­
standing of the impact of other factors on PV, both historically and prospectively.

1.05.4 Nonincremental Cost-Reducing Developments
The irregular occurrence of nonincremental technological improvement accounts for much of the concern about using experience
curve cost projections to inform policy. Projecting the future costs of PV depends on a much better characterization of these changes –
including when they are likely to occur, how big an impact they will have on costs, and conversely the consequences on costs of a
sustained period without such changes. To achieve greater understanding of nonincremental changes in PV, in a recent master’s thesis,
Husmann [93] studied historical silicon PV breakthroughs under the expectation that past breakthroughs can illustrate the causes and
effects of nonincremental changes on a still-developing technology. This section describes a novel methodology for identifying
nonincremental changes (‘breakthroughs’) and provides some preliminary results. Ultimately, the identification and evaluation of
specific technical improvements will be useful for informing the type of modeling described in Section 1.05.5.

1.05.4.1

Identifying Breakthroughs

Nonincremental changes can be identified using a combination of expert opinion and patent analysis.

1.05.4.1.1

Defining breakthrough

There is no widely accepted definition of breakthrough, even among researchers who study innovations in products and technology
[94]. Among PV researchers, important achievements in the field are called many things; technical developments [8], major results


62

Economics and Environment

[95], and advances [7] are just a few of the names used in the PV literature. For the sake of simplicity, we will use breakthrough, since
it is a word recognized in both research and colloquial literature. We define breakthrough as a technical achievement that represents
a new combination of technologies. A breakthrough must either open up a new area of technology for exploration and use, or be a
specific improvement that has spread to become industry standard. Although innovation researchers often require both technical
and economic achievements in order to identify their breakthroughs, the truth is that technical achievement will not necessarily
result in economic achievement [96]. Our goal was to first identify technological breakthroughs and then determine their economic
value. However, we recognize that our methods for identifying breakthroughs contain a (sometimes strong) bias for economic
success. That bias will be discussed later in this chapter.

1.05.4.1.2

First pass: Expert opinion

Silicon solar cells have been around since the early 1950s [97], which means that some researchers today have been studying
crystalline silicon PV for their entire lives. Their extensive knowledge and experience represents a great resource, if it can only be
tapped and distilled. Fortunately, that potential has been recognized, which has led to a profusion of recent papers describing the

history of PV and, importantly, listing the most important PV breakthroughs in the authors’ opinions.
Although none of the papers was written with the specific goal of identifying breakthroughs, they all mentioned various
discoveries and technologies that have played important roles in crystalline silicon PV history. We read through six of the papers
that seemed most relevant to crystalline silicon PV [8, 95, 96, 98–100] and compiled a list of breakthroughs mentioned by the
experts. In the end, we had a list of 79 breakthroughs spanning from 1951 to 2000. We left out more recent developments, since
identification of true breakthroughs requires several years of postbreakthrough observation.
An initial problem was that the original list of breakthroughs was too long. Not only would it have been prohibitively timeconsuming to research all 79, but it also implied that a breakthrough took place every 8 months, which was not a reasonable
assumption. Though the list was thorough, if we had tried to shorten it, we would have run into problems with the implicit biases
found in expert opinion. Not only are experts more familiar with breakthroughs in their own areas of research – and thus more
likely to list them as breakthroughs – but they can also suffer from success bias, in which experts tend to associate commercial
success with technical value [101]. The combination of the six papers helped to reduce possible bias in any one researcher, but if we
had asked an expert to pick just a few breakthroughs from our list of 79, it would have immediately suffered from that researcher’s
bias – and there were not enough researchers with enough time to look through the list. Instead, we chose to use another method of
breakthrough identification to reduce the list of breakthroughs.

1.05.4.1.3

Why patent analysis?

Our second method of breakthrough identification is patent citation analysis. We decided to look at patents, because citation
frequency of a patent can serve as a proxy for a possible breakthrough. Each patent is meant to represent a novel, useful, and
nonobvious invention [102]. Not every invention is a breakthrough, but some may be – and our goal is to identify the patented
inventions that are breakthroughs. It is also true that not all inventions are patented [103], but if they become common knowledge,
then they may be incorporated into other patents. Thus, the breakthroughs that are incorporated into other patents become
discoverable. The only problem is if the breakthrough is a trade secret, because by definition it cannot be identified by someone who
does not know the secret.
Aside from trade secrets, it is likely that all the breakthroughs we are looking for can be found in at least one patent. Previous patent
studies have shown that the semiconductor industry produces a large number of patents. Since key semiconductor breakthroughs have
been patented, it makes sense that many key PV breakthroughs have also been patented. Even though trade secrets are an acknowl­
edged intellectual property (IP) protection tool used by the industry, important patents appear to outweigh important secrets.

We are also focusing on patents because they can help us avoid expert bias – the tendency of experts to consider their own areas
of expertise as important – and success bias – the tendency to equate commercial success with technical value. Any form of patent
analysis precludes the possibility of expert bias, because patent citations are typically created by several experts: the inventors and
their attorneys suggest the citations, while the patent examiner selects the citations from the list of suggestions and his or her own
knowledge of prior art [104]. This process, combined with the fact that almost every patent represents a unique combination of
inventors, attorneys, and examiners, ensures that no one expert has undue influence over the patents being analyzed – especially for
such a large group of patents as the one we analyzed.
Patent analysis can also be constructed to limit success bias. One way to do this is to differentiate between backward and forward
citations. Backward citations represent the prior art that the patent is either building off of or surpassing. It cannot be influenced by
commercial success, because it is created before an invention is even patented. Forward citations represent the number of times that
the patent serves as prior art to another patent. Not only can forward citations be heavily influenced by the commercial success of
the patent, but they can also be influenced by the prior commercial success of the inventor. Because of the success bias present in
forward citations, we have chosen a method of patent analysis that relies primarily on backward citations in identifying
breakthroughs.

1.05.4.1.4

Backward citation analysis

A full explanation and derivation of backward citation analysis is presented by Dahlin and Behrens [101], but here we give a short
overview. While forward citations represent a patent’s commercial and technological impact, backward citations represent its


Historical and Future Cost Dynamics of Photovoltaic Technology

63

technological roots. If each backward citation represents a piece of technology that was either used to create the patent or served
as a source of inspiration, then the collection of all the backward citations in one patent can represent the specific combination of
technologies that the new patent represents. And if each patent can be technologically represented by its backward citations, then

a breakthrough technology can be identified by its backward citations alone – if the right definition of breakthrough can be
found.
This method creates a mathematical definition of radicalness – which we usurp as our definition of breakthrough – that can be
applied to a list of patents and used to identify specific patents. The three components of the definition are (1) novelty – the patent
differs from previous inventions; (2) uniqueness – the patent differs from current inventions; and (3) influence – the patent needs to
have affected future inventions. Although the third component can suffer from success bias, technological success will play a bigger
role than commercial success; that is, if a particular invention works well enough to be somehow adopted by other patents, the
backward citation analysis will notice this even if the authors of the other patents were not aware of the original patent.
We considered ignoring influence, but that resulted in the identification of both successful breakthroughs and very unique failures
with no obvious way to differentiate between the two.
In order to determine if a patent fulfills these three components, the backward citation analysis compares a patent’s collection of
backward citations to another patent’s collection. If the two collections have any backward citations in common, then they are
assumed to have similar technological roots. The similarity between patents i and j can be presented mathematically as overlap score
osij, in which osij equals the number of backward citations shared by patents i and j divided by the total number of backward
citations held by patents i and j. If this overlap score is computed for all the patents being analyzed, it can reveal the similarity
between any one patent i and all of its peers j for any year t, where t can be any year before or after the patent was published. Thus,
the average annual overlap score, osti/nt, is the sum of all of the overlap scores for a patent and its peers during a particular year
divided by the number of peers: osti/nt = (∑jostij)/nt. This average annual overlap score is a measure of the similarity between a patent
and all the analyzed patents in the chosen year.
This average annual overlap score is what determines whether a patent is novel, unique, or influential. For a patent to be novel,
it cannot be similar to patents that have been published before it: for t < ti, osti/nt must be small. For a patent to be unique, it cannot
be similar to patents that were published at the same time it was published: for t = ti, osti/nt must be small. And for a patent to be
influential, it must be similar to patents that were published after it: for t > ti, osti/nt must be big. A breakthrough will satisfy all three
requirements, while an incremental invention will not. However, these are all relative measures that will depend on the total
number of patents being analyzed and the variation between patents. It is important to recognize that a breakthrough identified
when looking at only PV patents may not appear to be a breakthrough when compared to patents covering, say, the World Wide
Web or the steam engine.
Backward citation analysis can also be prone to biases of its own. If a patent does not have any backward citations, it
shares no backward citations with any other patents and fails the influence component. This can happen occasionally if an
inventor is trying to be strategic, but it is discouraged by patent examiners and therefore uncommon. Another problem can

result if two patents had the same inventor (e.g., patents 4510674 and 4510675), because that inventor is much more likely
to list the exact same backward citations, making the two patents appear very similar. However, this can be mitigated if a
large number of patents are used in the analysis, because the greater number of patents tends to reduce the impact of any
one overlap score. Lastly, the backward citation analysis tends to recognize early breakthroughs more easily than later
breakthroughs, because early breakthroughs necessarily have a smaller overlap score for the few years of patents listed before
them, while later patents have a smaller overlap score for the few years of patents listed after them. This is a problem specific
to our use of backward citation analysis (not backward citation analysis in general) and will be addressed at length later in
this chapter.

1.05.4.1.5

Implementing backward citation analysis for PV

In order to accurately identify crystalline silicon PV breakthroughs using backward citation analysis, we needed a list of every
crystalline silicon PV patent. In all practicality, compiling such a list is impossible. If patents are given cryptic description – such as
labeling a PV patent as a semiconductor improvement – they can evade common search terms. Furthermore, not every PV patent
is relevant, since many relate to amorphous silicon, cadmium telluride, or some other form of PV. The only way to know if a
patent is relevant to crystalline silicon PV is if a PV expert reads through the entire patent, which would require hundreds of manhours.
Instead, we relied on a source that identified the crystalline silicon PV patents for us: NREL’s (National Renewable Energy
Laboratory) publication of U.S. Photovoltaic Patents from 1951 to 1993. This is a list of 1651 PV patents that was compiled by
searching for patents in the subclasses ‘Photoelectric’, ‘Testing’, and ‘Applications’ under ‘Batteries, Thermoelectric and Photoelectric’
or containing the word ‘photovoltaic(s)’ or ‘solar cell(s)’ [105]. Since the search probably relied on specific patent classes [106],
chances are that it missed most ‘solar cell’ patents outside of those classes; one omission (patent 4661200) has already been
identified.
Despite the likelihood of other missed PV patents, the benefits of using NREL [105] outweigh the costs, because the patents are
listed by category. It appears that at least one PV expert (Thomas Basso of NREL) read through the patents and was able to identify
the single-crystal silicon cells, amorphous silicon cells, III–V cells, and so on [106], thus providing the lists of crystalline silicon PV
patents that we require. Out of 17 categories of patents, we used 6: single-crystal silicon cells, polycrystalline and ribbon silicon cells,
cell components, cell enhancement techniques, materials production and processes, and flat-plate collectors. Our goal was to look

64

Economics and Environment

at the patents that have affected module design as well as silicon-specific developments, because even generic module changes have
had an important effect on the steady improvement of crystalline silicon PV.
The 1993 cutoff date was imposed on our analysis by NREL [105], since the list was never updated after that year. This necessarily
implies that our analysis will be historical, since the breakthroughs that we select for study are likely either no longer in use or
considered ‘common knowledge’ in the PV field.
While going through the list of patents in NREL [105], we selectively removed or changed some of the patents to be analyzed.
Design patents, which can be identified by the ‘D’ in front of the patent number, were considered irrelevant. They cover ornamental
appearances and thus do not represent a technological advance. Defensive publications or technical disclosures, which can be
identified by the ‘T’ in front of the patent number, were not considered patents and were not included. And reissued patents, which
can be identified by the ‘RE’ in front of the patent number, were changed to the original patent number. This prevented a reissued
patent from being compared to its original patent, since the two would necessarily have a high overlap score.
After changing or removing all of the design, defensive, and reissued patents, we ended up with a list of 1651 patents that were
ready to be analyzed, not including repeated patents. The NREL source included the same patent on multiple lists if it counted for
multiple categories. As we mentioned earlier, the composition of the list makes a difference in the breakthrough identification.
If each list were analyzed separately, it would guarantee that we would find breakthroughs for each category of module design,
because we can systematically take the top 10% of the results from each category. On the other hand, one complete list of all the
patents could discover if one category somehow contained more breakthroughs – perhaps through more intensive study or greater
innovation – but on the other hand, smaller lists by category could result in greater differentiation in scores, making it easier to
separate out breakthroughs. Furthermore, it would guarantee that we would have results from every category of module design. In
the end, we tried to balance these concerns by analyzing the categories both separately and combined in one list.
To conduct such a computation-intensive analysis, we used a very basic unpublished software program written by Dr.
Harlan Husmann. The inputs to the program were lists of patents, the years when each patent was published, and the list of
backward citations for each patent. The program calculated the average annual overlap score for each patent each year. The
default level of the patent then calculates two scores for each patent: the ‘before’ score and the ‘after’ score. The ‘before’ score is
the sum of each average annual overlap score from the first year to the year when the patent was published – that is, it

represents the novelty and uniqueness of each patent (lower ‘before’ scores represent higher novelty and uniqueness). The
‘after’ score is the sum of each average annual overlap score from the year after the patent was published to the last year – that
is, it represents the influence of each patent (higher ‘after’ scores represent higher influence). The advanced level does the same,
but also lists the patents that overlap each other.
These two scores are where our analysis diverges from the backward citation analysis described above. That analysis only looks at
the 5 years before the patent is published and the 5 years after publication, which is how they avoid biasing their results in favor of
early patents. We cannot do this, because PV patents can lie around for years before being incorporated in the field’s body of
knowledge.
Once the ‘before’ and ‘after’ scores are calculated, we can eliminate the incremental patents by taking the ratio of ‘after’ scores to
‘before’ scores. The higher the ratio, the less likely the patent is to be incremental. However, this does not remove the failed patents,
because if a failure has no overlaps before publication and a tiny overlap after publication, its ratio is still infinity. Therefore, to
remove the failures, we then go through the high ratios and select for the high ratios that also have high ‘after’ scores.
There are no specific numbers that count as high ratios and ‘after’ scores, because the backward citation analysis is relative.
Instead, our goal was to look at approximately the top 10% of high-scoring patents, since that would reduce the 1651 patents to a
more manageable group of about 160 patents that could actually be read in a reasonable amount of time. Therefore, we went
through the results for each list, first selecting the top ∼50% of high ratios, then reducing it to the top ∼10% with high ratios and high
‘after’ scores. Then we read through the abstracts of the top 10%, checking that each patent is actually relevant to crystalline silicon
PV and not some other type of PV. This entire process is summarized in Table 3.
Although we analyzed a total of 1651 patents, the sum for the category lists is 1986 patents because many of the patents
qualified for multiple categories. The ratio cutoff and ‘after’ cutoff values emphasize the relativity of the overlap scores. For
example, a ratio cutoff of 2 resulted in the selection of 55% of the patents in the ‘all patents’ list, while a ratio cutoff of 5 resulted
in a 55% selection in the ‘cell enhancements’ list. While the main goal of each ratio cutoff choice was in trying to end up with
∼50% of the patents in the first round of selections and ∼10% in the second round, the cutoff values were heavily influenced by
our need to pick values that did not result in too high percentages of qualifying patents and to keep both values relatively high
(i.e., a ratio of at least 2 and an ‘after’ value of at least 0.01). Once the cutoff values are applied and the ∼10% point is reached, we
checked that each patent was relevant – that it applied to silicon PV cells and was not purely decorative or similarly incon­
sequential. After removing the irrelevant and repeated patents in the multiple lists, we had a list of 113 patents culled from the
multiple categories. The list of 103 relevant patents selected from the list of ‘all patents’ had 35 patents in common with the list
of 113 patents, so our approach resulted in a total of 181 patents that qualified as breakthroughs under the backward citation
analysis.

As mentioned earlier, our backward citation analysis is prone to its own biases – most importantly, the fact that the overlap score
favors early breakthroughs over later breakthroughs. This can be seen in Figure 15, which shows, respectively, the percent and
number of silicon cell patents selected by the backward citation analysis versus the year in which the patents were issued. It is
possible that the first few years may have high percentages because the technology was new and each patent truly represented a
breakthrough. However, the last year certainly had no selected patents because the ‘after’ values were zero, and it is more than likely


Historical and Future Cost Dynamics of Photovoltaic Technology

Table 3

65

Summary of the patent selection process in backward citation analysis

Patent list from U.S.
Photovoltaic Patents

Number of
patents in the list

Ratio
cutoff

Percent making
the cutoff

‘After’
cutoff

Percent making
the cutoff

Number of remaining
relevant patents

All patents
Silicon cells
Polycrystalline silicon and
ribbon silicon cells
Cell components
Cell enhancements
Material processes
Flat-plate collectors

1651
251
99

2
5
3

55
32
43

0.02
0.01
0.07

6
16
11

103
26
10

420
75
713
428

2
5
2
2

56
55
59
32

0.03
0.06
0.03
0.03

10

12
7
10

32
4
19
34

100

90

80

Percent

70

60

50

40
30
20
10
0

1950

1960

1970


1980

1990

Year
Figure 15 Percent of silicon cell patents selected by the backward citation analysis in the year in which the patents were issued [93].

that the last few years of patent selections also suffered from artificially low ‘after’ values. For the purposes of our study, this means
that the range of years we are studying is slightly less than the range of years presented in the NREL document.

1.05.4.2

Results: Combining Expert Opinion and Patent Analysis

An ideal breakthrough selection process would combine an expert’s proficiency born of years of practice with a computer’s strict
impartiality – and that is what we aim to do by combining the expert’s list of breakthroughs with the patent list of breakthroughs.
These two lists are not particularly well matched: the expert‘s list contains the names of breakthroughs, while the patent list contains
the patent numbers of breakthroughs. Furthermore, the expert‘s list can be unclear in its meanings, especially if different researchers
use different names for the same breakthrough or the same name for different breakthroughs. Likewise, an inventor often describes a
new process or invention using names and terms that do not become standard in the literature – not to mention the fact that a
patent number alone contains no indication of what technology the patent represents.
In order to combine these two very different lists, we have become interpreters of sorts by comparing experts’ descriptions
of breakthroughs and inventors’ summaries of their inventions. There was no strict methodology – we simply read through
each patent on the list, looking for processes or technologies that matched any of the breakthroughs described in the experts’

papers. In general, we relied more heavily on the patents identified in the various categories of module design, because they
had already been examined and sorted by experienced PV researcher Tom Basso [106]. While reading through patents, we
looked for keywords (e.g., ‘boron’ in the case of boron diffusion), and while reading through research papers, we looked for
references to previous research that could help elucidate the author’s meaning behind a breakthrough name. For the most
part, matching patents to breakthroughs was an iterative process that could always be improved with more time and
expertise. However, we felt confident with our results after we had identified 39 patents that were connected to 23 break­
throughs (see Table 4).
The lists of patents and breakthroughs were not a one-to-one match. Most of the experts’ breakthroughs had no patents that
appeared related to them, whereas a few breakthroughs had multiple connected patents – up to six patents in one case. This is not
surprising; although we call breakthroughs nonincremental, the truth is that most breakthroughs were developed over time, even if
that interval of time was shorter than most breakthroughs. Furthermore, patents tend to emphasize incremental changes, since each
change can earn the inventor intellectual property and money. We accommodate this reality by underscoring the fact that the 23
breakthroughs from the experts’ list are the final breakthroughs that we select – not the patents. This way, our final list of


66

Economics and Environment

Table 4

List of selected breakthroughs and the patents that connect to them

Breakthrough

Patent (year)

Breakthrough description

Boron diffusion

P–N junctions created out of boron rather than lithium increase cell
efficiency from 4% to 6%.
Many thin contact fingers reduce the losses experienced by the
current as it travels through the cell.

Antireflective coatings

2794846 (1957)
3015590 (1962)
2862160 (1958)
2919299 (1959)
3040416 (1962)
3046324 (1962)
3450568 (1969)
3493437 (1970)
3533850 (1970)

Better, thinner top junctions

3811954 (1974)

Tunneling metal–insulator–semiconductor
contact

3928073 (1975)

Contacts (grid and fingers)

Laminating cells to glass with polyvinyl butyral (PVB)

4104084 (1978)
3990100 (1976)
4086102 (1978)
4171997 (1979)
4009054 (1977)

Unexposed silicone rubber

4057439 (1977)

Aluminum-based pastes

4086102 (1978)

Screen printing

4105471 (1978)

Hydrogen plasma passivation

4113514 (1978)
4321420 (1982)
4322253 (1982)
4557037 (1985)
4154625 (1979)

Oxide surface passivation

Pulse annealing

Metallization

Quasi-square wafers
Reduced metallization resistance

4235644 (1980)
4348546 (1982)
4361718 (1982)
4356141 (1982)

Metal–insulator N–P (MINP) cell

4395583 (1983)
4694115 (1987)
4404422 (1983)

Ribbon on sacrificial growth plate

4478880 (1984)

Ethylene–vinyl acetate (EVA) laminate

4499658 (1985)

Plasma deposition of SiN passivation

4640001 (1987)

Low-contact resistance with anti-reflective films

4643913 (1987)

Reactive-ion etching
Passivated emitter solar cell (PESC)

4664748 (1987)
4667058 (1987)
4589191 (1986)

Microgrooving

4626613 (1986)

Transparent layer of antireflective material on top of the cell reduces
reflection and increases the amount of sunlight that is absorbed by
the cell.
Shallower P–N junctions increase the cell response to shorter
wavelengths and thus the amount of sunlight the cell could convert.
A thin insulator placed between the semiconductor layer and the metal
contacts reduces recombination losses and allows more electrons
to reach the contacts.
Silicon dioxide attached to broken silicon bonds on the surface of the
cell increases cell efficiency.
Replacing the layer of exposed silicone with glass allows PV modules
to weather the elements better.
Replacing the layer of exposed silicone with glass allows PV modules
to weather the elements better.
Al-based pastes used at the rear of the cell helped remove impurities
and give the modules better weatherization.
By printing contacts on the top and bottom of the cells, PV modules

can be made more easily, reducing costs.
Hydrogen atoms attach to broken silicon bonds throughout the cell
and increase cell efficiency.

Pulse annealing creates larger silicon crystals, improving the quality of
the silicon at a lower cost.
Metallization uses thin films of metal to create contacts, reducing the
amount of metal used and thus the cost.
Square wafers use space and silicon more efficiently than circular
wafers.
Reduced resistance between the metal contacts and silicon decreases
losses and allows more electrons to reach the contacts.
Combining metal–insulator–semiconductor contacts with shallow N–
P junctions, increasing efficiency.
Thin ribbons of silicon are continuously produced with less waste
silicon, reducing costs.
EVA doesn’t yellow with exposure to sunlight as quickly as PVB,
allowing the PV module to last longer.
Thin films of silicon can be made by heating silicon nitride until the
nitrogen escapes, reducing the cost and amount of silicon used.
Silver pastes lowered the contact resistance, reducing recombination
losses and increasing efficiency.
Highly reactive ions texture the surface of the cell, reducing reflective
losses.
Contacts are made through slits in the top oxide layer, increasing
contact passivation and thus efficiency.
Selective surface etching of microgrooves reduces reflectivity,
reduces resistance losses, and is easier to work with than etched
pyramids.

Historical and Future Cost Dynamics of Photovoltaic Technology

67

breakthroughs is not beholden to the single point of time represented by a patent; it is associated more with idea behind the patent.
Although we cannot prove that this is true in our backward citation analysis, we assume it is true for the purposes of our analysis.
One example that is illustrative of the usefulness of backward citation analysis is the case of patent number 4165241, which we
had identified because its name included the words ‘printed contact’. The backward citation analysis had not identified the patent as
a possible breakthrough because its ratio value was 0.1429 – far below the cutoff. This was suspect because one of its inventors, John
Yerkes, was a recognized PV researcher and the patent summary seemed perfect for the ‘screen-printing’ breakthrough. However,
after further investigation, we discovered that this patent was a forward citation for patent 4105471 by the same inventor. The
backward citation analysis had not only identified this patent as a breakthrough, but identified the patent 3 times – in the list of ‘all
patents’ and two different categories of module design. This example does not mean that backward citation analysis will always lead
to a breakthrough or the origin of a breakthrough, but it does give us confidence that it has some proficiency in tracing research
developments in patents.
These results show the promise of a rigorous, transparent, and replicable methodology for identifying nonincremental cost
reductions in PV. The identification of a specific list of the most important technical improvement provides an avenue for
determining the factors that enabled these breakthroughs to occur – as well as to assess the impact they have had. Ultimately,
this much more specific characterization of technological change will be helpful for structuring and populating models that predict
future changes, such as the one described next.

1.05.5 Modeling Nonincremental Changes in PV
Given the importance of this array of nonincremental changes, how can modeling take them into account? The previous section
shows a set of nonincremental technical advances within crystalline silicon PV. The transition to new technological generations in
PV is likely to have greater impacts, limiting the reliability of learning curve projections in describing them. PV based on crystalline
silicon has remained the overwhelmingly dominant technology for three decades, despite the advantages of thin-film technologies
that use less raw material and are more amenable to mass manufacturing techniques. Policy related to PV in the longer term must
address subsequent generations of PV technology, such as those based purely on organic materials. In a recent study, Nemet and
Baker [107] combined an expert elicitation and a bottom-up manufacturing cost model to compare the effects of R&D and demand

subsidies. They modeled the effects of these policy instruments on the future costs of a low-carbon energy technology that is not
currently commercially available, namely, organic PV. That study found that production-related effects on technological advance –
learning by doing and economies of scale – are not as critical to the long-term potential for cost reduction in organic PV as the
investment in and success of R&D.
One example of a new technological generation is purely organic PV. It is particularly intriguing because of characteristics that
distinguish it from the current generation of PV, which consists of cells made from crystallized silicon. Purely organic PV use a thin
film of organic semiconductor material for photon conversion. Because they do not require a glass substrate, organic PV cells can be
manufactured on highly flexible material, leaving open the possibility of a much wider range of applications. These manufacturing
techniques are more amenable to automation and high throughput because they involve chemical rather than mechanical
production processes. That they also require only a thin layer of light-absorbing PV material, rather than a crystal structure,
means that the amount of input materials needed is very low. The combination of highly automated ‘reel-to-reel’ manufacturing
processes and small materials consumption gives organic PV its most appealing distinguishing characteristics – the potential for very
low manufacturing costs [108]. However, organic PV are not currently manufactured on a commercial scale. Moreover, the current
models have very low efficiency, with the highest being around 5% in laboratory conditions [109]; this compares to about 15%
efficiency for silicon-based solar cells. Finally, organic materials are susceptible to degradation in sunlight, leading to concerns about
the lifetimes of these cells.
In this case of a new generation of technology like organic PV, policy can impact future cost in multiple ways. First,
technology-push policies, such as direct government-sponsored R&D, can increase the likelihood of achieving technical break­
throughs. Nemet and Baker assume that government R&D has an impact on two technical characteristics of organic solar cells:
(1) their electrical conversion efficiency and (2) their lifetime. Second, demand-pull policies, such as demand subsidies, increase
demand for organic PV and thus create opportunities for cost reductions through economies of scale and learning by doing. That
model focuses on these two avenues of technical change. Note, however, another potential impact: particularly at later stages of
technology development, demand-pull policies may stimulate private sector R&D through the promise of a larger, less risky market.

1.05.5.1

An Approach to Modeling Nonincremental Technological Change

Nemet and Baker [106] developed the following methodology, taking the perspective that the combination of expert
elicitation with a bottom-up manufacturing cost model provides a promising avenue for more robustly understanding future

technology costs. Adoption subsidies are modeled as having an impact on cost by enabling economies of scale through
increasing demand. R&D investment outcomes are modeled using the results of the expert elicitation. They use this schema
to evaluate the uncertain impact of combinations of R&D investments and subsidies on the cost of electricity over time. The
central question in that study was how R&D investment policies interact with demand subsidy policies to impact the cost of
electricity from PV.


68

Economics and Environment

This model considers the effects of two demand-pull instruments: demand subsidies and carbon prices. The authors model
subsidies as a decision variable and treat carbon prices as an exogenous sensitivity. (The reason for the focus on subsidies as the
primary demand-pull decision variable in this model is that they can be designed to exclusively support organic PV, whereas carbon
prices enhance demand for low-carbon technologies in general.) In order to assess the effectiveness with which technology policy
can induce technical change in organic PV, they determined how the specific policies – investment in R&D and demand subsidies –
affect technology improvements.

1.05.5.2

Results for Nonincremental Technological Change

They simulate efforts by the government to fund R&D and subsidize demand at three levels of policy intensity each, calculate the
costs of PV electricity in 2040 and 2050 under the nine combinations of government technology programs (low R&D is $15
million yr−1 for 10 years, high R&D is $80 million yr−1 for 10 years; low subsidy is 20¢ kWh−1 for 5 years, high subsidy is 25¢ kWh−1
declining to 5¢ kWh−1 over 20 years). While both subsidies and successful R&D programs reduce costs, the effect of successful R&D
on cost in 2050 is an order of magnitude larger than the effect of subsidies. Subsidies are relatively more effective in 2040 than in
2050, but the effect of successful R&D is still much larger, despite the fact that in their model only half of the benefits of R&D arrive
by then. Even the highest subsidy levels do not achieve cost-effective organic PV without successful R&D. The cost of PV without
successful R&D never falls below 16¢ kWh−1, far from the target level of 4¢ kWh−1. Note also the counterintuitive result that, under

successful R&D programs, the high- and low-subsidy programs produce costs in 2050 that are slightly higher than without the
subsidy program. This result occurs because the subsidy programs shift a substantial amount of PV production to earlier years;
without subsidies, almost all of the demand for PV electricity in 2050 is met by production between 2040 and 2050. Consequently,
without subsidies, the scale of manufacturing plants in 2050 reaches a larger, more efficient scale and the cost in 2050 is lower. The
curves in Figure 16 show the path of cost reductions over time and the relationships among the policy combinations. The three
subsidy curves in Figures 16(a)–16(c) are much more similar to each other than the three R&D curves in Figures 16(d)–16(f).
Interestingly, the relative effectiveness of successful R&D and subsidies does not change under varying assumptions about
storage and carbon prices; under all four scenarios, R&D success has a greater effect on cost reductions than do subsidies in 2050.
High carbon prices do enhance the relative impact of subsidies and free storage increases the relative impact of R&D success, but in
both cases the effects are small. Furthermore, sensitivity analysis shows that the two main claims are robust to uncertainty in the data
used to populate the model. First, the analysis supports the claim that the base case set of assumptions represents an upper bound

$1

(b)
$1

(c)
$1

10c

10c

10c

1c

1c

(a)

1c
20

No subsidy

Low subsidy

High subsidy

4c goal
30

40

50

20

30

40

(e)

(d)

20

50

$1

$1

10c

10c

10c

1c

1c

1c
20

30

40

50

40

50

30

40

50

(f)

$1

No R&D

Low R&D

High R&D

4c goal

30

20

30

40

50

20

Figure 16 Impact of subsidies and R&D on cost per kWh of PV electricity: (a) no R&D; (b) low R&D; (c) high R&D; (d) no subsidy; (e) low subsidy; and

(f) high subsidy. Low and high R&D cases are conditional on program goals for efficiency and lifetime being reached. Costs are on a log scale [106].
Reproduced from Nemet GF (2006) Beyond the learning curve: Factors influencing cost reductions in photovoltaics. Energy Policy 34(17): 3218–3232 [6].


Historical and Future Cost Dynamics of Photovoltaic Technology

69

on the effectiveness of a subsidy program. Second, all alternative scenarios support the finding that the cost-reducing effect of
successful R&D is larger than the effects of subsidies.
Another result that is consistent across scenarios is that subsidizing a large demand for PV before the benefits of R&D arrive can
be expensive. Under the base case assumptions, no carbon tax and no free storage, the net present social cost of subsidies is
$5 billion for the low-subsidy program and $80 billion for the high-subsidy program. These values are in line with recent estimates
of the cost of subsidizing the current generation of PV [110]. Note that they are considerably higher than the R&D amounts we have
considered, which have a net present value of $0.1 billion and $0.7 billion. Also, the cost of each subsidy program increases as
demand for solar electricity increases. For example, in the presence of a $1000 carbon tax, the cost of the low-subsidy program rises
to $30 billion and that of the high-subsidy program rises to $3 trillion. Given the wide range of subsidy program costs, it may be
useful in future work to use this model to optimize the timing and level of subsidies – especially given various assumptions about
carbon prices and storage technology. The outcomes here are similar to those discussed above on sensitivity of learning curve
predictions to expensive outlier cases.
Despite the possibility of expensive outcomes, a case can still be made for subsidies. A subsidy with no R&D is ‘less risky’, since it
avoids the worst case of a very high electricity cost, and the expected cost is lower as well. This result implies that if (1) the goal were
simply to achieve as low an electricity cost as possible and (2) the two programs had equal costs, the ‘no R&D/high-subsidy
program’ would be strictly preferred by all risk averters. Note, however, that neither of these conditions necessarily holds. These
results imply that the value of subsidies is that they provide a hedge against the possibility that breakthroughs in technical change
fail to take place. In a choice under uncertainty framework, subsidies provide a benefit in reducing risk.

1.05.5.3

Summary of Nonincremental Modeling

The Nemet and Baker study found that (1) successful R&D enables PV to achieve a cost target of 4¢ kWh−1, (2) the cost of PV does
not reach the target when only subsidies – and not R&D – are implemented, and (3) production-related effects on technological
advance – learning by doing and economies of scale – are not as critical to the long-term potential for cost reduction in organic PV as
the investment in and success of R&D. Those results were insensitive to the intensity of either type of program, the level of a carbon
price, the availability of storage technology, and uncertainty in the main parameters used in the model. The central policy
implication of those results is that governments must find a way to engender this R&D, whether it is funded by the government
itself or by the private sector in response to changing demand conditions. In fact, one might argue that the key question policy
makers face, with regard to PV development, is how to encourage this R&D, rather than how to support economies of scale and
learning by doing.
To be sure, that study found that a case could still be made for subsidies through analysis of risk. Because of the possibility of
R&D failure, the benefits of subsidies stochastically dominate those of R&D. In the event of R&D failure, subsidies make the costs of
PV much lower than they would otherwise be, albeit not at levels close to the target. The importance of subsidies as a hedge against
inherently uncertain R&D programs depends on the value that society places on the availability of a low-carbon energy source that is
moderately inexpensive – that is, unlikely to be competitive with all other technologies but perhaps inexpensive enough to be
deployed at a large-enough scale to diversify energy supply.
Much work needs to be done to develop this and other methods for modeling discontinuous evolution of PV technology and the
impacts on costs. Given the importance of historical nonincremental changes described in the previous section, as well as the
possibility of even more significant changes in the future, the need for better models is great.

1.05.6 Future Progress and Development
The insights gained from assessing cost development in PV technology have some direct policy implications; they also clarify needs
for modeling.
As a result of the assessment of the sources of historical cost reductions as well as modeling future costs, we now have several
design criteria for public policies intended to reduce the costs of PV. We need an array of supporting policy instruments: for
example, R&D, demand subsidies, and encouragement of intersectoral spillovers. Timing matters; making good decisions about
when to switch from a focus on R&D to a more capital-intensive investment in wide-scale deployment is crucial. Multiyear demand
subsidies are important because the benefits of demand subsidies come from expectations about markets in the long term.
Intermittent demand supports may actually be worse than no support. R&D support also needs a long-term commitment; budgets
or grants that span multiple years are important; supporting policies, such as Japan’s Sunshine Program or US Project Independence,

that demonstrated commitment by making this area of work a serious national priority were also successful. Niche markets have
been crucial, especially when government support was lacking. The success of new technological generations may require renewed
R&D support even while markets for the existing technology are expanding; new problems need to be worked and postdeployment
experience needs to feed back into subsequent R&D decisions.
These insights imply a strong need for better tools with which to understand technological change in PV. Much is at stake, in
terms of both the public’s financial resources used to fund these programs and the environmental impacts these programs are
designed to mediate. These decisions are too important – and mistakes too expensive – to rely on simple heuristics that mask large


70

Economics and Environment

uncertainties and that are easily ignored. Promising developments exist. An important analytical improvement has certainly been
the inclusion of explicit treatment of learning uncertainty in modeling [111–113]. Estimating technology costs through the
summation of bottom-up characterization of technology dynamics in individual components provides an appealing alternative
in that sources of uncertainty can be identified more precisely [114]. The combination of such bottom-up models with experience
curves and expert opinion provides a method that is more robust to bias within any single method [83]. As discussed above, an
alternative use of bottom-up methods is to integrate them with expert elicitation into a single model that represents both
incremental and nonincremental technical change [107]. This integration will help account for the introduction of new technolo­
gical generations, which seems especially likely in the case of PV. Improving the accuracy and precision of models such as these is an
important research endeavor. Identifying the specific nonincremental changes that occurred, as described by Husman [115].,
provides a methodology for connecting them to causes and impacts; this provides a potential means by which to supplement
expert opinion on the effectiveness of future research investments. The policy implications described above reveal important
decisions, such as timing and resource allocation between R&D and subsidies. Developments in modeling future costs have the
potential to improve the efficiency and efficacy of these programs in fully realizing the very large potential societal benefits of
widespread PV deployment.

References
[1] Nemet GF (2007) Policy and Innovation in Low-Carbon Energy Technologies. PhD Dissertation, University of California.

[2] FS (2009) First Solar Passes $1 Per Watt Industry Milestone (25 Feb.). Tempe, AZ: First Solar.
[3] Schaeffer GJ, Seebregts AJ, Beurskens LWM, et al. (2004) Learning from the Sun. Analysis of the use of experience curves for energy policy purposes: The case of photovoltaic
power. Final Report of the PHOTEX Project, Petten, The Netherlands. ECN Renewable Energy in the Built Environment.
[4] Wiser R, Barbose G, et al. (2009) Tracking the Sun: The Installed Cost of Photovoltaics in the U.S. from 1998–2007. Berkeley, CA: Lawrence Berkeley National Laboratory.
[5] GEA (2011) The energy technology innovation system. In: Nakicenovic N (ed.) The Global Energy Assessment (GEA). Cambridge: Cambridge University Press.
[6] Nemet GF (2006) Beyond the learning curve: Factors influencing cost reductions in photovoltaics. Energy Policy 34(17): 3218–3232.
[7] Surek T (2003) Progress in U.S. photovoltaics: Looking back 30 years and looking ahead 20. In: Proceedings of the 3rd World Conference on Photovoltaic Energy Conversion,
May 11–18, Osaka, Japan.
[8] Green MA (2005) Silicon photovoltaic modules: A brief history of the first 50 years. Progress in Photovoltaics: Research and Applications 13(5): 447–455.
[9] Taylor M, Nemet G, et al. (2007) Government Actions and Innovation in Clean Energy Technologies: The Cases of Photovoltaic Cells, Solar Thermal Electric Power, and Solar
Water Heating, CEC-500-2007-012. Sacramento, CA: California Energy Commission.
[10] Moore RM (1982) Czochralski silicon solar cell modules: Present cost and future prospects. Solar Cells 5: 313–329.
[11] Shum KL and Watanabe C (2007) Photovoltaic deployment strategy in Japan and the USA – An institutional appraisal. Energy Policy 35(2): 1186–1195.
[12] Peevey MR and Malcolm K (2006) Interim Order Adopting Policies and Funding for the California Solar Initiative. California Public Utilities Commission, Rulemaking

04-03-017.

[13] Neuhoff K, Nemet G, et al. (2007) The Role of the Supply Chain in Innovation: The Example of Photovoltaic Cells. Cambridge: University of Cambridge – Electricity Policy
Research Group.
[14] Jaeger-Waldau A (2004) PV Status Report 2004: Research, Solar Cell Production and Market Implementation of Photovoltaics. Ispra, Italy: European Commission, DG JRC,
Institute for Environment and Sustainability, Renewable Energies Unit.
[15] Maycock PD (2005) PV technology, performance, cost 1995–2010. Williamsburg, VA: PV Energy Systems.
[16] Dutton JM and Thomas A (1984) Treating progress functions as a managerial opportunity. Academy of Management Review 9(2): 235–247.
[17] van Benthem A, Gillingham K, et al. (2008) Learning-by-doing and the optimal solar policy in California. The Energy Journal 29(3): 131.
[18] Shum KL and Watanabe C (2008) Toward a local learning (innovation) model of solar photovoltaic development. Energy Policy 36(2): 508–521.
[19] Menanteau P (2000) Learning from variety and competition between technological options for generating photovoltaic electricity. Technological Forecasting and Social Change
63(1): 63–80.
[20] Yu CF, van Sark WGJHM, et al. (2011) Unraveling the photovoltaic technology learning curve by incorporation of input price changes and scale effects. Renewable and

Sustainable Energy Reviews 15: 324–337.

[21] Maycock PD and Stirewalt EN (1985) A Guide to the Photovoltaic Revolution. Emmaus, PA: Rodale Press.
[22] Wolf M (1974) Historic development of photovoltaic power generation. In Helmut R Loesch (ed.): International Conference on Photovoltaic Power Generation., pp. 49–65.
Hamburg, Germany: Deutsche Gesellschaft fuer Luft- und Raumfahrt.
[23] Roessner DJ (1982) Government–industry relationships in technology commercialization: The case of photovoltaics. Solar Cells 5(2): 101–134.
[24] Maycock PD (1984) U.S.–Japanese competition for the world photovoltaic market. In: Ebinger CK and Morse RA (eds.) U.S.–Japanese Energy Relations: Cooperation and
Competition, pp. 207–217. London: Westview Press.
[25] Herfindahl OC (1950) Concentration in the US Steel Industry. New York: Columbia University.
[26] Hirschman AO (1945) National Power and the Structure of Foreign Trade. Berkeley; Los Angeles: University of California Press.
[27] Costello D and Rappaport P (1980) The technological and economic development of photovoltaics. Annual Review of Energy 5(1): 335–356.
[28] Nemet GF (2009) Interim monitoring of cost dynamics for publicly supported energy technologies. Energy Policy 37(3): 825–835.
[29] BCG (1972) Perspectives on Experience. Boston, MA: The Boston Consulting Group.
[30] Irwin DA and Klenow PJ (1994) Learning-by-doing spillovers in the semiconductor industry. Journal of Political Economy 102(6): 1200–1227.
[31] Watanabe C, Wakabayashi K, et al. (2000) Industrial dynamism and the creation of a ‘‘virtuous cycle’’ between R&D, market growth and price reduction – The case of

photovoltaic power generation (PV) development in Japan. Technovation 20: 299–312.

[32] Laird FN (2001) Solar Energy, Technology Policy, and Institutional Values. New York: Cambridge University Press.
[33] van Sark W, Alsema E, et al. (2010) General aspects and caveats of experience curve analysis. In: Junginger M, van Sark W, and Faaij A (eds.) Technological Learning in the
Energy Sector: Lessons for Policy, Industry and Science. Cheltenham, UK: Edward Elgar.
[34] Hoffmann W and Pellkofer T (in press, corrected proof) Thin films in photovoltaics: Technologies and perspectives. Thin Solid Films. />
tsf.2011.04.146.

[35] Schumpeter JA (1947) Capitalism, Socialism, and Democracy. New York; London: Harper.
[36] Freeman C (1994) The economics of technical change. Cambridge Journal of Economics 18(5): 463–514.
[37] Mowery DC and Rosenberg N (1998) Paths of Innovation: Technological Change in 20th-Century America. Cambridge: Cambridge University Press.


Historical and Future Cost Dynamics of Photovoltaic Technology

[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]

[68]
[69]
[70]
[71]
[72]
[73]
[74]
[75]
[76]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]
[85]
[86]
[87]
[88]
[89]
[90]
[91]
[92]
[93]
[94]
[95]

71

Rosenberg N (1994) Exploring the Black Box: Technology, Economics, and History. Cambridge: Cambridge University Press.
Alchian A (1963) Reliability of progress curves in airframe production. Econometrica 31(4): 679–693.
Rapping L (1965) Learning and World War II production functions. The Review of Economic Statistics 47(1): 81–86.
Smith A (1776) Of the division of labor. In: An Inquiry into the Nature and the Causes of the Wealth of Nations. Chapter 1. London: Methuen and Co.
Wright TP (1936) Factors affecting the costs of airplanes. Journal of the Aeronautical Sciences 3: 122–128.
Arrow K (1962) The economic implications of learning by doing. The Review of Economic Studies 29(3): 155–173.
Duke RD and Kammen DM (1999) The economics of energy market transformation initiatives. The Energy Journal 20(4): 15–64.
van der Zwaan B and Rabl A (2003) Prospects for PV: A learning curve analysis. Solar Energy 74: 19–31.
Azar C and Dowlatabadi H (1999) A review of technical change in assessment of climate policy. Annual Review of Energy and the Environment 24(1): 513–544.
Grubler A, Nakicenovic N, et al. (1999) Dynamics of energy technologies and global change. Energy Policy 27: 247–280.
Williams RH and Terzian G (1993) A Benefit–Cost Analysis of Accelerated Development of Photovoltaic Technology. Princeton, NJ: The Center for Energy and Environmental
Studies, Princeton University.
Wene C-O (2000) Experience Curves for Technology Policy. Paris: International Energy Agency.
Neij L, Andersen PD, et al. (2003) The use of experience curves for assessing energy policy programs. In: Proceedings of the EU/IEA Workshop on Experience Curves: A Tool for
Energy Policy Analysis and Design, 22–24 January. Paris, France.
Edenhofer O, Lessmann K, et al. (2006) Induced technological change: Exploring its implications for the economics of atmospheric stabilization. Synthesis report from the
Innovation Modeling Comparison Project. Special Issue: Endogenous Technological Change and the Economics of Atmospheric Stabilisation. The Energy Journal 27: 57–108.
Grubb M, Koehler J, et al. (2002) Induced technical change in energy and environmental modeling: Analytic approaches and policy implications. Annual Review of Energy and
the Environment 27: 271–308.
Messner S (1997) Endogenized technological learning in an energy systems model. Journal of Evolutionary Economics 7(3): 291–313.
Nordhaus WD (2002) Modeling induced innovation in climate change policy. In: Grubler A, Nakicenovic N, and Nordhaus WD (eds.) Technological Change and the
Environment, pp. 182–209. Washington, DC: Resources for the Future; International Institute for Applied Systems Analysis.
Freeman C and Louca F (2001) As Time Goes By: From the Industrial Revolutions to the Information Revolution. Oxford: Oxford University Press.
Arthur WB (2006) Out-of-equilibrium economics and agent-based modeling. In: Judd K and Tesfatsion L (eds.) Handbook of Computational Economics, Vol. 2: Agent-Based
Computational Economics, pp. 1551–1564. Elsevier; North-Holland: Amsterdam.
Marx K (1867) Machinery and modern industry. In: Capital. Harmondsworth, Middlesex, UK.
Junginger M, van Sark W, et al. (2010) Technological Learning in the Energy Sector: Lessons for Policy, Industry and Science. Cheltenham, UK: Edward Elgar.
IPCC (2007) Climate Change 2007: Mitigation. Contribution of Working Group III to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change.
Cambridge; New York: Cambridge University Press.

Stern N (2006) Stern Review on the Economics of Climate Change. Cambridge: Cambridge University Press.
IEA (2008) Deployment and technology learning. In: Energy Technology Perspectives. Paris: International Energy Agency.
Goldemberg J, Coelho ST, et al. (2004) How adequate policies can push renewables. Energy Policy 32(9): 1141–1146.
CPUC (2009) California Solar Initiative Annual Program Assessment, California Public Utilities Commission (CPUC), June, 2009.
CPUC (2003) Decision 03-06-071, Order Initiating Implementation of the Senate Bill 1078 Renewable Portfolio Standard Program. California Public Utilities Commission (CPUC).
Sher B (2002) Senate Bill 1078: The Renewables Portfolio Standard of California – Chapter 516, Statutes of 2002, State of California.
Yelle LE (1979) The learning curve: Historical review and comprehensive survey. Decision Science 10: 302–328.
Wene CO (2007) Technology learning systems as non-trivial machines. Kybernetes 36(3–4): 348–363.
McDonald A and Schrattenholzer L (2001) Learning rates for energy technologies. Energy Policy 29: 255–261.
Argote L and Epple D (1990) Learning curves in manufacturing. Science 247(4945): 920–924.
Romer PM (1990) Endogenous technological change. The Journal of Political Economy 98(5): S71–S102.
Hall B and Mairesse J (2006) Empirical studies of innovation in the knowledge-driven economy. Economics of Innovation and New Technology 15: 289–299.
Maycock PD (2002) The World Photovoltaic Market. Warrenton, VA: PV Energy Systems.
Strategies Unlimited (2003) Photovoltaic Five-Year Market Forecast, 2002–2007. Mountain View, CA: Strategies Unlimited.
Hall G and Howell S (1985) The experience curve from the economist’s perspective. Strategic Management Journal 6(3): 197–212.
Baloff N (1966) Learning curve – Some controversial issues. Journal of Industrial Economics 14(3): 275–282.
Carlson JG (1973) Cubic learning curves: Precision tool for labor estimating. Manufacturing Engineering and Management 71(5): 22–25.
Sheshinski E (1967) Tests of the learning by doing hypothesis. Review of Economics and Statistics 49(4): 568–578.
Adler PS and Clark KB (1991) Behind the learning curve: A sketch of the learning process. Management Science 37(3): 267–281.
Buonanno P, Carraro C, et al. (2003) Endogenous induced technical change and the costs of Kyoto. Resource and Energy Economics 25(1): 11–34.
Miketa A and Schrattenholzer L (2004) Experiments with a methodology to model the role of R&D expenditures in energy technology learning processes: First results. Energy
Policy 32: 1679–1692.
Thompson P (2001) How much did the liberty shipbuilders learn? New evidence for an old case study. Journal of Political Economy 109(1): 103–137.
Neij L and Astrand K (2006) Outcome indicators for the evaluation of energy policy instruments and technical change. Energy Policy 34(17): 2662–2676.
Neij L (2008) Cost development of future technologies for power generation – A study based on experience curves and complementary bottom-up assessments. Energy Policy
36(6): 2200–2211.
Rubin ES, Antes M, et al. (2005) Estimating Future Trends in the Cost of CO2 Capture Technologies. Pittsburgh, PA: International Energy Agency Greenhouse Gas R&D
Programme (IEAGHG).
Koomey J and Hultman NE (2007) A reactor-level analysis of busbar costs for U.S. nuclear plants, 1970–2005. Energy Policy 35(11): 5630–5642.
Borenstein S (2008) The Market Value and Cost of Solar Photovoltaic Electricity Production. Berkeley, CA: University of California Energy Institute’s Center for the Study of

Energy Markets.
van Sark WGJHM (2008) Introducing errors in progress ratios determined from experience curves. Technological Forecasting and Social Change 75(3): 405–415.
van Sark W, Alsema EA, et al. (2008) Accuracy of progress ratios determined from experience curves: The case of crystalline silicon photovoltaic module technology
development. Progress in Photovoltaics 16(5): 441–453.
SEIA (2004) Our Solar Power Future: The Photovoltaics Industry Roadmap Through 2030 and Beyond. Washington, DC: Solar Energy Industries Association.
Gritsevskyi A and Nakicenovic N (2000) Modeling uncertainty of induced technological change. Energy Policy 28(13): 907–921.
Baker E and Adu-Bonnah K (2008) Investment in risky R&D programs in the face of climate uncertainty. Energy Economics 30(2): 465–486.
Schnapp R (2008) Federal Financial Interventions and Subsidies in Energy Markets 2007. Washington, DC: Energy Information Administration – Office of Coal, Nuclear, Electric
and Alternate Fuels.
Husmann D (2011) Identification of Breakthroughs in Photovoltaics. MS Thesis, University of Wisconsin.
Garcia R and Calantone R (2002) A critical look at technological innovation typology and innovativeness terminology: A literature review. Journal of Product Innovation
Management 19(2): 110–132.
Saitoh T (2003) 30 Years of progress in crystalline silicon solar cells. In: Proceedings of the 3rd World Conference on Photovoltaic Energy Conversion, pp. A23–A28. Osaka,
Japan, 18 May 2003.