A model provides a kind of lens that reveals particular aspects of a phenomenon; but it inevitably obscures others. As a result, it can be useful to apply different models to the same phenomenon to see different things. So in this section, we'll develop a second mechanistic model to illuminate more aspects of how technological innovation develops. To see how different models can show different things, consider how classical mechanics and statistical mechanics can both describe the behavior of an ideal gas in a container. Classical mechanics describes the macroscopic behavior of the gas based on Newton's laws of motion and the principles of thermodynamics, treating the gas molecules as individual particles with well-defined positions and velocities. One of its fundamental results is the ideal gas law, $PV=nRT$, which describes the relationship between the macroscopic properties of pressure (_P_), volume (_V_), and temperature (_T_). Statistical mechanics, on the other hand, delves into the microscopic details, explaining the macroscopic behavior of the gas through the statistical distribution of particles and their interactions. It shows, for example, that the average kinetic energy of a gas molecule is directly proportional to the temperature of the gas and that the pressure is created by the momentum of the individual gas molecules crashing into the container’s walls. Those insights answer questions that would be hard to even formulate from the perspective of classical mechanics. In a similar way, different models of technology innovation bring different insights. Another approach is to examine three different levels of the dynamics driving cost (low-level mechanisms, high-level mechanisms, and policy-level mechanisms) and analyze how those dynamics interact. The questions this model will address include: - How important is government investment in research and development of technology? - Is R&D important only in the early stages of technological development, or can it speed innovation at later stages as well? - Do policy-level market-growth incentives, such as subsidies or tax credits, speed technological development? - Is it better for the government to invest in R&D or market-growth incentives, or do both have a role to play? - How big a role does learning-by-doing really play in bringing down costs, as compared to private R&D or economies of scale? - What factors have sped up or delayed the development of a particular technology? - How can we most effectively invest in particular technologies to speed their development? # Analysis: Why did PV costs plunge? Just like when we discussed the interconnected features we [[Underlying Processes#Design Complexity|model]], this second model, the multi-level drivers model, can be used to understand the options for driving innovation for any technology. It was motivated, however, by one specific question about one specific technology: Why did the cost of photovoltaics fall so dramatically? The plunge has been steeper than for nearly any other energy technology, dropping by more than 99% between 1980 and 2023. Analyzing PV technology data using Wright’s Law, we find that costs dropped by roughly 20% with every doubling of cumulative production; using Moore's Law, costs fell by about 10% every year. PV solar energy is now among the cheapest options for new electricity projects in many places on Earth. ![[Pasted image 20250406104254.png]] > [!Figure] > Adapted from Kavlak, G., McNerney, J., & Trancik, J. E. (2018). Evaluating the causes of cost reduction in photovoltaic modules. _Energy Policy, 123_, 700–710. [URL](https://hdl.handle.net/1721.1/123492) The figure above displays the cost decline of photovoltaic modules. The sustained nature and rate of cost decline in this technology has been exceptional among energy technologies. One cent’s worth of today’s solar energy would have cost US$1.40 a few decades ago, adjusting for inflation. Given the urgency of the climate crisis, understanding this stunning success to see how it might be replicated is of key importance. Moreover, this technology’s improvement rates are close to that of some other kinds of fast-improving technology, like various types of information technology. Thus understanding the drivers may begin to give us some insight into why fast-improving technologies evolve so quickly. Many different explanations have been offered for the fall in prices. Some have given credit to particular government policies, others to the accumulation of experience, others to economies of scale, and others to research and development. Some have argued for explanations more idiosyncratic to PV technology, like China’s decision to invest heavily. Many have argued for multiple explanations, placing varying weights on those they favored. However no research had systematically and definitively answered the question until a more focused approach took it on. The novel approach consisted first of recognizing that the explanations offered for the decline in PV costs operated at different scales. At a low level, individual components of the technology got cheaper because of improved manufacturing yields, conversion efficiency improvements, decreased silicon usage, etc. At a high level, these improvements were driven by research and development, learning-by-doing on the factory floor, economies of scale, etc. Finally, at the highest level, policy mechanisms drove change, including through government-funded research and market incentives such as tax credits. Those different levels meant that different explanations could be true simultaneously. For example, suppose a worker who cuts materials on the factory floor identifies a way to minimize material loss during this cutting. This innovation could result in an increased utilization of the material and lower the overall cost of the final technology. One way to explain this cost reduction is to say that the yield of the material increased (a low-level mechanism). Another way is to say that learning-by-doing drove down costs because the worker learned how to minimize material loss in the process of making the technology (i.e., learning-by-doing, a high-level mechanism). Furthermore, a public policy that spurred demand for the technology, such as a subsidy, could have led to increased production, during which the worker learned how to increase yield (a policy-level mechanism). All of these explanations are correct at the same time. > [!important] > Although there are drivers of technology innovation at many levels and implicating many actors, it is generally presumed that most people in society are not involved in driving the direction and rate of technological innovation. However, it could be argued that the market forces leading to the improvement and widespread adoption of one technology or another are an expression of the public will. ## A Step-By-Step Guide to the Model This model builds on this basic insight that technological innovation has causes at different levels. If we can break down the costs of a technology carefully and identify how they are influenced by mechanisms at each level, we can understand the drivers of changes in cost. The strategy for the model is to first come up with an equation that adds up all the costs. The costs need to be divided up in such a way that data is available to populate the equation, and of course, having lots of data over a long time is better. Populating the cost equation with that data in different time periods shows how these costs changed over time. The next step is to attribute the cost changes to low-level mechanisms, such as changes in the price of materials, increased efficiency, or other direct engineering impacts. Then, we figure out how much each of these low-level mechanisms was driven by a high-level mechanism so that, for example, a decrease in material prices from bulk purchasing is attributed to economies of scale. This allows us to quantitatively assess how much each high-level mechanism drove the change in cost. Finally, we attribute the high-level mechanisms to particular combinations of policy-level mechanisms. Economies of scale, for example, are related to market expansion policies. And because we already have a quantitative assessment of the impact of the high-level mechanisms, we can use this process to get a quantitative assessment of the role of policy-level mechanisms. A key step, then, is to map out how the explanatory levels drive one another. For the photovoltaic modules, the following diagram can be produced: ![[Pasted image 20250406105129.png]] > [!Figure] > Adapted from Trancik, J. E., & Ziegler, M. S. (2023). _Accelerating Technology Innovation: A Mechanistic Approach and Lessons for Policymakers_. Massachusetts Institute of Technology. [URL](https://dspace.mit.edu/handle/1721.1/147765) The mechanisms in the diagram above are specific to PV modules, so using this strategy for another technology would require tuning for its particularities. Similarly, it would be essential to understand how the high-level drivers affected the modified low-level mechanisms and how the policy-level drivers affected the high-level mechanisms. We’ll get into the details of how the levels relate to one another later, but the general idea can be traced in this diagram. Improvements in module efficiency, for example, are a result of a combination of public and private R&D. Public R&D is a direct result of government investment decisions, while private R&D is a result of market expansion, which is partly a consequence of government policies to encourage that expansion. Data about the ratio of public to private investment in R&D can be used to divide up attribution into those two categories. With this conceptual understanding, let’s now get into the details of how the model is implemented. ### Step 1: Identify the performance metric and performance equation The first step is to define what we hope the model will help us understand. This means identifying the performance metric we want to analyze and expressing it in terms of available data. Two broad categories of performance metrics, for example, are cost per unit service and environmental impact per unit service. In the example of the analysis of PV costs, the exact metric is cost per watt produced. Examples of environmental impact metrics include water intensity or greenhouse gas emissions per unit service. Then we create an equation that accounts for each component's contributions to that performance metric. In the general case, this is called a performance equation; in the specific case of costs, it’s also called a cost equation. It’s key to break down the components in a way that fits with the data available. So for the example of silicon-based PV costs, the three broad categories of cost are the cost of the silicon needed, the cost of the non-silicon materials (including items such as the wire, the chemicals for cell manufacturing, the glass and the frame), and the costs associated with the plant itself (including labor costs, electricity costs, maintenance, and depreciation). It’s important to capture every single cost: When it’s done correctly, the sum of the cost of the components will equal the cost of the complete modules. In developing a performance equation, you have to make decisions about how far to break down the costs, and that's largely determined by the detail of the data available, as well as the types of future interventions you may want to inform. For example, the cost equation for PV modules could be expanded to include many more details describing how the panels were manufactured, including the rate of manufacturing, the number of workers per shift, the wages paid to the workers, the electricity used by the factory, etc. However the data might not be as detailed, and even without that detail, there might be enough information about costs to quantify the impact of the low-level, high-level, and policy-level mechanisms at a reasonable level of granularity and precision. Naturally, the more data is available at shorter time intervals over a longer period of time, the more revealing the analysis will be. ### Step 2: Low-level mechanisms #### Part 1: identifying the low-level mechanisms of interest Low-level mechanisms are causes for the change in costs that concern the engineering of the technology. They reflect changes to variables in the performance (or cost) equation. The low-level mechanisms will vary between technologies, but for PV modules, these are mechanisms of interest: - Δ efficiency ( i.e., the change in the fraction of the energy in the sunlight hitting a module that is converted into electricity) - Δ non-silicon material costs (including the wires, the glass, the frame, etc.) - Δ silicon price - Δ silicon usage - Δ wafer area - Changes in factory costs. This is divided into two groups: - Δ Plant size: The change in costs that scale with the size of the factory, such as electricity usage, labor costs, maintenance, and depreciation - Δ $p_0$: The change in the cost of a factory at a fixed size - Δ yield (i.e., change in the percentage of modules that don't have to be scrapped because of production errors) Remember that low-level mechanisms are drivers of cost changes. Because we start with a cost equation, and from there we model the cost effects of changes to variables in the cost equation, the cost changes from the low-level mechanisms add up to the total change in cost. This is an important feature of the method developed here. It means that we can check whether we are capturing—at some level of granularity—all of the cost change drivers at a relatively detailed, engineering level. #### Part 2: Quantifying the effects of low-level mechanisms We now have an accounting of the costs. Next, we ask, what are the variable-level changes that drove the changes in these costs? In the case of photovoltaics, we can identify mechanisms like changes in the price of materials, engineering changes that reduced the amount of materials needed, increases in efficiency, reductions in plant-related costs from economies of scale, and more. Each of these changes affect costs through changing variables in the cost equation. These changes in variables are the low-level mechanisms. The table below lists each of the low-level mechanisms identified as being relevant, along with the particular cost components they affect. |**Low-level mechanism**|**Affected costs**| |---|---| |Δ non-silicon materials costs|Non-silicon materials costs| |Δ silicon price|Silicon costs| |Δ silicon usage|Silicon costs| |Δ wafer area|Plant size-dependent costs| |Δ plant size|Plant size-dependent costs| |Δ yield|Non-silicon materials costs<br><br>Silicon material costs<br><br>Plant size-dependent costs| |Δ efficiency|Non-silicon materials costs<br><br>Silicon material costs<br><br>Plant size-dependent costs| For each low-level mechanism, we want to understand how big an impact it had on the lowered costs (or other measure of performance). So, for example, how much did increased efficiency contribute to lower costs? How much did falling silicon prices contribute? Because multiple low-level mechanisms typically occur at once, it takes a bit of mathematical work to separate out the impacts. In the case of PV technologies, this chart summarizes the findings: ![[Pasted image 20250406110435.png]] > [!Figure] > Adapted from Kavlak, G., McNerney, J., & Trancik, J. E. (2018). Evaluating the causes of cost reduction in photovoltaic modules. _Energy Policy, 123_, 700–710. [URL](https://hdl.handle.net/1721.1/123492) We see that a variety of low-level mechanisms had a significant impact, and that those impacts changed over time. Over the entire period from 1980 to 2012, increased efficiency had the biggest impact, followed by non-silicon materials costs, silicon price, and silicon usage, in descending order. The ranking of those impacts was similar in the earlier period from 1980 to 2001, but from 2001 to 2012, plant-size related cost decreases had the biggest impact by far, followed by increased efficiency, increases in wafer area and then decreases in non-silicon materials cost. ### Step 3: High-level mechanisms Now that we know how much the low-level drivers contributed, we can ask about the high-level changes that drove them. For example, we just saw that decreases in non-silicon materials costs was the second-biggest driver of lowered PV costs over the whole period. So, what drove those decreases? Was it research and development, economies of scale, learning by doing on the factory floor, or something else? Or was it a combination of those, and if so, how big a role did each play? > [!important] > **R&D**: These are improvements made in laboratory settings (i.e. controlled environments with specific procedures) or any non-routine production activity (e.g. production line experimentation). This includes both public and private sector investment. > **Economies of Scale**: This includes volume purchasing, plant scale effects, and other scaling of suppliers. > **Learning by Doing**: These include only incremental improvements made through routine production activity. In keeping with usage throughout the literature, this is a narrow definition including only learning on the factory floor and during other routine procedures. > **Other**: The primary other high-level effect that we will consider here is called a spillover effect. This happens when another technology has impacts on the technology in question. For example, the development of silicon supplies for semiconductor manufacturing led to reduced silicon costs for photovoltaic modules, as we’ll describe in more detail below. But depending on the technology, there could be other effects that should be considered. Suppose, for example, that a new source of PV semiconductor material had been found because of investment from an industry unrelated to PV, and that this had driven down the price of the active material used in a PV panel (e.g., it had driven down the prices of that input material as compared to silicon). This did not happen, but if it had, it would be the type of thing included here. So it is important to understand that identifying the relevant high-level effects relies on a granular understanding of the technology in question. We have no direct way of quantifying the impacts of each of these mechanisms: There are no datasheets reporting them. However, this model provides a method of approximating their impact. Using knowledge of the details of the manufacturing processes of a given technology, we can assign low-level mechanisms to one or more high-level mechanisms. We can then add up the effects of the low-level mechanisms assigned to each high-level mechanism to estimate its impact. As we discuss further below, it will also be important to perform sensitivity analyses to be able to draw conclusions that are robust to various sources of uncertainty. In the case of photovoltaics, some of these impacts are clear. Improved module efficiency, silicon usage and reduced wafer area were each a result of R&D; cost savings from increased plant size were a result of economies of scale; and increased yield was a result of learning by doing. Non-silicon materials costs are a bit trickier since both R&D and economies of scale played a role. The researchers estimated that R&D and economies of scale played a roughly equal role and divided the cost savings between the two. In the case of silicon prices, the drivers varied over time. Between 1980 and 2001, the first period for which there is good data, silicon for modules often came from the semiconductor industry. The research thus considered the price change a spillover effect and assigned it to the "Other" category. During the second period (2001–2012), PV industry demand for silicon surpassed that of the semiconductor industry, and polysilicon producers scaled to meet this demand, so the price of silicon was assigned to economies of scale. Of course, in real life, there was no abrupt shift, but this represents a reasonable approximation. ==This example shows two things that hold for any technology: These assignments must be informed by a detailed knowledge of the dynamics in that particular industry, and the assignments often can only be made approximately.== Having made these assignments, the cost changes each high-level mechanism is responsible for can simply be added up across the low-level mechanisms it contributed to. So, for example, to estimate the impact of economies of scale on photovoltaics costs, we add up the impacts of increased plant size, half of that from non-silicon materials costs, and silicon prices for the second period only. The figure below shows the results for all of the high-level mechanisms: ![[Pasted image 20250406111812.png]] > [!Figure] > Adapted from Kavlak, G., McNerney, J., & Trancik, J. E. (2018). Evaluating the causes of cost reduction in photovoltaic modules. _Energy Policy, 123_, 700–710. [URL](https://hdl.handle.net/1721.1/123492) Over the entire period, public and private research and development had by far the greatest impact, with economies of scale playing a secondary role. The findings were similar during the first period from 1980 to 2001, but between 2001 and 2012, research and development played a roughly equal role with economies of scale. Learning-by-doing had a smaller effect through both time periods, partly because of its narrow definition (limited to learning effects that occur when a similar step in the manufacturing or design process of a technology is performed repeatedly). This is an example of a pattern we end up seeing in many technologies: the potential to increase human efficiency at repeated tasks tends to be limited. For this reason, the improvement of many soft technologies, at least historically, tends to be less than for hardware. ### Step 4: Policy mechanisms At the highest level, we'd like to understand the impact of government policies, which play a key role in many (though not all) technologies. For photovoltaics, we know that government policies were critical because, early in PV development, PV electricity was a hundred times more expensive than fossil fuel electricity. The biggest advantage of PV for the public is that it contributes far, far less greenhouse gas emissions per unit energy—==but that societal advantage doesn't drive most individual consumers to buy it==. Greenhouse gas emissions are an example of an externality, i.e., something affected by the technology that is not reflected in the market price. The two main categories of government policy that affected PV development were public investment in research and development and market-stimulating policies. The latter category includes many different specific policies that incentivize private sector investment in a technology. These include tax rebates, subsidies, regulations such as the requirements that automobile manufacturers meet fuel efficiency standards, or prices on externalities such as pollution. In this analysis, we allow for a third category called "Other", which captures other effects that cannot be related directly to policy. So, for example, the spillover effects we discussed earlier, where PV benefited from investments in the semiconductor industry, were not driven by policies. As with high-level mechanisms, there is no direct way to assess the impact of government policies simply by tracking available data. But again, this model allows us a way to quantitatively estimate what their impacts have been, by assigning high-level mechanisms to them and adding up the cost changes. Given that the high-level mechanisms could only be estimated, the uncertainty in the calculations for policy-level mechanisms is even higher. While keeping that uncertainty in mind, we can nonetheless draw useful and robust conclusions. In the example of photovoltaics, the four high-level mechanisms are R&D (which combines public and private investment), economies of scale, learning-by-doing, and other effects including spillovers. The data suggested that the government and the private sector made roughly equal investments in R&D in this technology, and the private investment was enabled by the expanding markets. So the researchers attributed half of the combined R&D-related cost changes to government R&D and half to market-expansion policies. Cost changes resulting from economies of scale and learning-by-doing were attributed to market expansion policies. One could wonder if government policies should really be credited for all these effects, and in some technologies, that would not be appropriate; government policies may have helped without necessarily making a decisive difference, or they may have been largely irrelevant. In the case of photovoltaics, however, it's clear that government policies were essential. As we mentioned before, in the early days, electricity from photovoltaics was a hundred times more expensive than electricity from other sources. That means that almost no one would have bought it, and without a market, private sector investment in R&D or increasing scale couldn't have happened. So in this case, attributing R&D, economies of scale, and learning-by-doing to government policies is reasonable, but this must be assessed case by case for each technology being analyzed. The figure below shows the results of the research found with economies of scale, learning-by-doing, and half of R&D (accounting for the private portion) attributed to market-stimulating policies and shown in blue, while half of R&D (accounting for the public portion) and "other" is shown in white: ![[Pasted image 20250406112251.png]] > [!Figure] > Adapted from Kavlak, G., McNerney, J., & Trancik, J. E. (2018). Evaluating the causes of cost reduction in photovoltaic modules. _Energy Policy, 123_, 700–710. [URL](https://hdl.handle.net/1721.1/123492) Market expansion policies contributed about 55% of cost reductions between 1980 and 2001, and their contribution grew to approximately 75% between 2001 and 2012. Across the full period, market expansion policies contributed nearly 60% of all observed cost reductions, suggesting that market-stimulating policies were very important in driving the cost reduction of silicon PV modules. This analysis was based on the researchers' most realistic estimates of how to attribute the low-level mechanisms to high-level mechanisms. However because there is inherent uncertainty in these estimates, the researchers also did sensitivity analyses in which they varied those attributions to cover the full range of possible scenarios, including ones that didn't seem particularly likely. Additionally, the researchers varied the contribution of public vs private R&D across all possible scenarios. These analyses generate the whisker bars shown in the figure. Based on these sensitivity analyses, the researchers showed that market-expansion policies could imaginably have contributed only half the change in cost rather than 30%, or at the maximum, a bit over 70%. But wherever it lies exactly, the fundamental conclusion is the same: both policies contributed significantly to the overall cost decline. Furthermore, examining the low-level mechanisms of cost change makes it clear that important changes occurred in both publicly funded labs and companies. Moreover, the types of improvements to low-level variables observed were quite different depending on whether they occurred in private companies or publicly funded labs. Thus, public R&D and market-stimulating policies worked in concert to support cost change, and it is unlikely that the two types of policies could have been substituted for one another to produce the same results. This is a key conclusion for policymakers and society from this analysis. While we can't travel to the past and run a counterfactual study withholding government policies to see what the difference would be, it is quite clear from this analysis that both categories of policies—public funding for R&D and market-stimulating policies—were essential for driving the improvement trend observed. ### Step 5: Forecasting The multi-level drivers model provides a powerful tool for analyzing future technology innovation scenarios, but it works in a different way from the data-driven models. You don’t simply let it run into the future. Instead, you can use it to play with scenarios targeting particular mechanisms and estimating the effects. So in the case of photovoltaics, the researchers first tested the impact on cost of improving the low-level mechanisms individually. In this scenario, they improved each low-level mechanism by 25%, except for yield, which they took to the limiting case of 100%, and plant size, which they tested by increasing by a factor of three or a factor of ten. The figure below shows the results, with increasing plant size by a factor of three shown in orange and increasing it by a factor of ten shown in blue. ![[Pasted image 20250406112804.png]] > [!Figure] > Adapted from Kavlak, G., McNerney, J., & Trancik, J. E. (2018). Evaluating the causes of cost reduction in photovoltaic modules. _Energy Policy, 123_, 700–710. [URL](https://hdl.handle.net/1721.1/123492) So increases in efficiency and plant size showed the greatest potential for decreasing cost when these variables were changed individually. In real life, it's very unlikely that only a single variable would improve. So next, the researchers tried improving all the low-level variables at once. They also grouped the improvements into the high-level mechanisms. The figure below shows their findings: ![[Pasted image 20250406112913.png]] > [!Figure] > Adapted from Kavlak, G., McNerney, J., & Trancik, J. E. (2018). Evaluating the causes of cost reduction in photovoltaic modules. _Energy Policy, 123_, 700–710. [URL](https://hdl.handle.net/1721.1/123492) Examining the high-level mechanisms, a combination of public and private R&D stands out as having the most impact, with economies of scale also contributing. Among low-level mechanisms, increases in efficiency, larger plant sizes, and decreases in non-Si material costs reduce costs the most under both scenarios. This suggests that it would be wise to target those technology characteristics, where possible, for additional research and development. Looking at a policy level, we can add the impacts of economies of scale, learning-by-doing, and the private portion of R&D to estimate the impacts of market expansion policies and consider the impacts of public R&D to estimate the impacts of public investment in R&D. We see that if plant size increases tenfold, economies of scale and learning-by-doing will together contribute about 40% of the cost decline. This suggests that while a balance of market expansion and R&D policies will pay off, additional R&D, both public and private, is especially valuable. These estimates are just a starting place. One also has to ask how feasible it is to improve these low-level mechanisms by 25% as we have in this scenario, and that depends on physical limits. Targeted expert elicitations can help provide that information. You could, for example, find an expert in the design of cells in photovoltaic modules and ask, "How much might silicon usage in photovoltaics decrease if we target research and development efforts on that for a year?" And then you could ask an expert in silicon supply chains, "How do you expect silicon prices to change over the next year?" Combining these targeted forecasts, you could then use the model to forecast module costs in a year. You’re likely to get a much more precise and accurate forecast than if you rely purely on expert elicitations without this model, asking the broad question, "What will the cost of photovoltaic modules be in a year?" # What we learn from the model Modeling the mechanisms of technological change at these different levels allows us to understand what drove or impeded technological innovation in the past, and how it might be supported in the future. We can gain valuable insights from what has worked and failed in the past in guiding our future efforts. Looking forward, we can also quantitatively model specific scenarios, to identify low-level mechanisms that can be impactful in the future for each technology, and learn about the high-level mechanisms and policies that may be needed to support them. R&D is likely going to be important for many technologies long after their initial market introduction because of the many different low-level mechanisms R&D can bring about. For technologies that will not take off in the marketplace on their own—because the service they offer addresses an externality not captured in the market price—a combination of both government funding for R&D and market-growth incentives is likely to be important to kick off innovation trajectories. We can now go back to the questions that motivated this model. ## How important is government investment in research and development of a technology? It can be very important when the technology addresses an externality, such as pollution, that is not captured in the market price. In the successful development of photovoltaics, government-funded R&D played a critical role, contributing to about 30% of the decline in PV costs. It also contributed to lithium-ion battery development, though it’s more difficult in that case to quantify the impact. ## Is R&D only important in the early stages of technological development, or can it speed innovation at later stages as well? It continued to be a key driver of cost declines for both PV (and also lithium-ion batteries) throughout their development. For photovoltaics, it was the dominant driver during the early stages of development and the second driver later, but even in the later period it accounted for more than 20% of cost declines. ## Is it better for the government to invest in R&D or market-growth incentives, or do both have a role to play? Both are important, and they support one another. ## How big a role does learning-by-doing really play in bringing down costs, as compared to private R&D or economies of scale? Learning-by-doing, as narrowly defined here and throughout the economics literature, plays a smaller role in the technologies analyzed. ## What factors sped up the development of photovoltaics? At a low manufacturing level, increase in efficiency, decrease in non-silicon materials costs, and decrease in silicon prices were the biggest drivers. At a high manufacturing level, public and private R&D played the biggest role over the entire time period, contributing more than half of the cost reduction. But during the more recent period from 2001 to 2012, economies of scale and R&D played roughly equal roles. At a policy level, market expansion policies contributed nearly 60% of the cost savings over the whole time period, and that increased to 75% from 2001 to 2012. Government-funded research and development contributed the remaining amounts. ## Which investments in PV technology are likely to have the greatest impact on costs? Research and development continues to be important and, in some locations, market-stimulating policies are still needed to drive market growth. The modeling framework presented can be used to investigate different combinations of low-level mechanisms that may be most effective at further improving the technology. # References - Eash-Gates, P., Klemun, M. M., Kavlak, G., McNerney, J., Buongiorno, J., & Trancik, J.E. (2020, November). Sources of cost overrun in nuclear power plant construction call for a new approach to engineering design. _Joule, 4_(11), 2348–2373. [URL](https://www.sciencedirect.com/science/article/pii/S254243512030458X) - Kavlak, G., McNerney, J., & Trancik, J. E. (2018). Evaluating the causes of cost reduction in photovoltaic modules. _Energy Policy, 123_, 700–710. [URL](https://hdl.handle.net/1721.1/123492) - McNerney, J., Doyne Farmer, J., Redner, S., & Trancik, J. E. (2011, May 31). Role of design complexity in technology improvement. _PNAS, 108_(22), 9008–9013. [PDF](https://www.pnas.org/doi/pdf/10.1073/pnas.1017298108) - Trancik, J. E., & Ziegler, M. S. (2023). _Accelerating Climate Innovation: A Mechanistic Approach and Lessons for Policymakers_. Massachusetts Institute of Technology. [URL](https://dspace.mit.edu/handle/1721.1/147765) - Zielger, M. S., Song, J., & Trancik, J. E. (2021). Determinants of lithium-ion battery technology cost decline. _Energy & Environmental Science, 14_, 6074–6098. [URL](https://pubs.rsc.org/en/content/articlelanding/2021/ee/d1ee01313k) - Ziegler, M. S., & Trancik, J. E. (2021). Re-examining rates of lithium-ion battery technology improvement and cost decline. _Energy & Environmental Science, 14_, 1635–1651. [URL](https://pubs.rsc.org/en/content/articlelanding/2021/ee/d0ee02681f)