Results

Deployment of geothermal power globally

Geothermal power has been used since the early 1900s.35 As of mid-2025, projects exist in more than 60 countries (Figure 1A), yet geothermal accounts for only 0.5% of global renewable electricity generation.4 Recent growth has been modest, with just 0.2 GW of new capacity added in 2024 (Figure 1B), compared to 114 GW in wind turbines and 452 GW in solar PV.36

Figure 1: Global distribution map and historical deployment of geothermal power plantsMap of geothermal power plants (A), historical deployment globally (B), in the top ten countries by cumulative deployment (C–L), and in all remaining countries combined (M). See Figure S1 for cumulative deployment.

Many large plants are in geologically highly active regions, especially around the Pacific Ring of Fire (Figure 1A). Multiple halted projects (pre-operation) illustrate the development challenges associated with geothermal deployment beyond the most advantageous regions.

Global deployment varies strongly over the years (Figure 1B) as few large plants in geologically highly active regions dominate capacity additions. While deployment generally increased over the last decades, there has been a downward trend in capacity additions since a peak in 2019, when the newly installed capacity was close to 1 GW. At the national level, deployment is even more irregular (Figures 1C–1M).

Examples illustrate the variations in deployment. In the United States, capacity additions peaked in the 1980s, when multiple big fields strongly expanded production (Figure 1C). Recent capacity additions are lower, as they mostly focus on the edges of existing fields, less attractive new fields, and the repowering of old, partly depleted fields. Türkiye experienced a large geothermal boom in the 2010s, boosted by subsidies. After subsidies were phased out, deployment plummeted (Figure 1F).37 In other countries, deployment patterns were mostly driven by site availability and policy changes as well (Note S1).

Estimating and explaining global experience rates

Figure 2A displays all 206 geothermal plants with cost data, plotted against cumulative deployment, which is based on all 514 projects (see STAR Methods for details). We estimate a global ER of −20% [90% confidence interval (CI): −41.2% to −2.6%] (Model 1 in Figure 2B). The wide CIs stem from the fact that we use all projects with all their site-specific conditions. Notably, many early plants had low installed costs around 3 USD2024/W, whereas several recent plants exhibit significantly higher costs, some exceeding 10 USD2024/W. When we exclude small plants (i.e., the 16% of the sample with an electrical capacity below 5 MW), the estimated ER is not much affected, however (Model 2).

Figure 2: Plant costs and global experience curves of geothermal powerPlant costs in 2024 USD and global experience curve of Model 1 (A) and ERs with alternative models (B). Model 1 regresses over all plants. Model 2 includes only plants with capacities of at least 5 MW. Models 3 and 4 mirror Models 1 and 2, respectively, but adjust for exchange rate distortions. Detailed parameters for all models can be found in Table S3. In (A), global cumulative capacity per plant is assigned per the capacity at the beginning of the year in which the plant became operational. In this, no mutual learning is assumed between plants becoming operational within the same year and thus constructed in parallel.

In estimating global ER, adjusting for exchange rate fluctuations can help to avoid biased ER estimates (see STAR Methods).38 The exchange rate-adjusted ER in Model 3 (−14%) is slightly higher than in Model 1 as the effect of US-Dollar appreciation against softer currencies is accounted for. The currency depreciation effect mostly affects small plants, which is why Model 4 shows less sensitivity to currency corrections and its ER (−19%) is close to Model 2. Models 3 and 4 have wider CIs than Models 1 and 2 due to the larger impact of some soft currencies’ fluctuations against the US Dollar.

Observers might note some statistical peculiarities in our data and results. First, there is a large spread between the cheapest and most expensive projects. We explain more about the techno-economic reasons for these variations below and note that cost variations of about one order of magnitude in a single year are the norm in all large-scale renewable energy technologies.36 Second, as a consequence of these fluctuations and the fact that we use plant-level data, the observed confidence intervals are wider than in most studies that use yearly averages. Third, the wide confidence intervals in some models cross the zero-line, facilitated by ERs that are closer to zero than in some other technologies. In the literature, there are many technologies without a CI clearly in the positive or negative range. Depending on the study and model, these include offshore wind, onshore wind, hydropower, nuclear power, concentrating solar power, bioenergy, and coal and gas generation.19,20 Many other studies do not include confidence intervals, but aggregate over different time intervals or regions and also show both positive and negative ERs for many technologies.13,21 Only solar PV has its CI entirely in the positive range in all studies. This is due to solar PV’s generally high ER, e.g., +34% according to Yao et al. (2021), and its globally standardized nature, which limits cost fluctuations.

While the experience curves quantify how geothermal costs have evolved, they cannot reveal why learning has stalled globally and why it varies so strongly across countries. The drivers behind these patterns are not contained in the cost data itself but are distributed among the practitioners who develop the plants. We therefore draw on 51 expert interviews. Their value lies less in uncovering individual barriers that might already be familiar with geothermal specialists, but rather in making them analytically useful: We separate drivers that systematically and globally suppress learning from those that act only locally, at specific times, or for specific technologies; we group them into broader factors; and we phrase these factors so they transfer to other technologies and to the wider scholarship on technological learning. This yields six factors that explain the observed global cost dynamics (Table 1). These relate to different ER determinants described in the literature.

Factor | Explanation | Interview quote | Effect
Diminishing productivity of new sites | Easily accessible high-quality resources with high temperature gradients are rare. Thus, project developers must successively turn to less productive or more costly sites | “The best resources have been captured early. [Today,] we do much more well redrilling and plants at the edges of existing fields [where yield is less and more uncertain]. Also, the wells become more complex.” (Interview 9) | Increases cost of later plants
Related determinant:
Resource effect
Project design complexity | Degree of interdependence between project phases and actors. Higher complexity increases project risks and coordination needs and slows learning-by-doing | “[Projects are] complex, with many specialists needed. […] and, we need many iterations.” (Interview 7) | Slows learning due to need for many interactions between actors
Related determinant:
Design complexity
Varying hydro-geology | Hydrogeology can strongly vary between sites, thus requiring re-assessments of geology and different knowledge to drill and to operate reservoirs | “[There are] very few economies of scale […]. One can learn very little from other geothermal projects — unless they are right on your doorstep.” (Interview 7) | Slows learning due to customization and lower scaling of the exact same setup (no repetition)
Related determinant:
Need for customization
Tacit nature of drilling knowledge | Much of the knowledge required for projects is tacit. This knowledge resides in drilling teams and project managers and is hard to transfer | “[It takes] two years for a drilling crew to become fully working. […] any successful project requires experts who understand the local geology” (Interview 39) | Constrains knowledge to local sites and teams
Related determinant:
Knowledge transfer
Regulatory density and variance | Regulation varies significantly in place and time. Reasons for this lie in differing seismic risks, vulnerability of water reserves, and population density. Regulatory density has increased | “Regulations, permitting and subsidy are more complicated than geology. […] There are differences on a communal level […] and there are new surprises every day.” (Interview 7) | Increases cost and slows learning due to increased project design complexity and customization
Related determinant:
Need for customization
Knowledge transfer from oil and gas industry | Existing oil and gas knowledge positively impacts geothermal projects. This includes drilling know-how, equipment and personnel availability, familiarity of regulators with drilling, public acceptance, and data availability concerning the local geology | “[In our region] there used to be good knowledge, but it was lost with [oil and gas firm name] that disintegrated. They were the last ones.” (Interview 33) | Decreases cost without a need for geothermal investment
Related determinant:
Knowledge transfer

Diminishing productivity of new sites can drive up costs from project to project, as e.g., surface-near heat reservoirs are exploited first, thus countering the experience curve effect. In the literature, this factor is described as “resource effect.” A plot of the available resources against their (increasing) cost is known as the resource supply curve, and such curves can be found in the literature.39,40,41 The most important aspect of site productivity in geothermal is temperature, since higher temperatures have a double-positive effect of increasing energy of the fluid and higher energy extraction efficiency (from Carnot efficiency). We find that resource temperatures are declining globally and in all important countries (Figure S2). Next to this, increasing resource depth also plays a role in increasing project costs, as does the decreasing “drillability” of the rock and the decreasing permeability of the targeted formation. Furthermore, increasing remoteness of sites and decreasing knowledge about underground properties can increase costs of new sites. The diminishing productivity of new sites is also relevant in other renewable energy technologies, such as hydropower, where most suited dam sites are already exploited.42 In contrast, resource depletion plays only a minor role in wind power, where wind-rich sites remain widely available, and is barely relevant in solar PV.7,43,44

Project design complexity hinders fast improvements and learning-by-doing due to long improvement cycles within projects. The iterative nature of projects with many actors increases coordination cost, the risk of bottlenecks, and the cost of experimentation, as interviewees pointed out. In the literature, high design complexity is named as a reason for the low experience rates of technologies such as nuclear or carbon capture and storage plants.30 In contrast, the project design complexity of solar PV plants is much lower and experience rates much higher.32

Varying hydrogeology, or more broadly, spatial heterogeneity in the physical environment, slows learning according to our interviews. In geothermal, every reservoir differs in terms of rock permeability, fluid chemistry, and thermal gradient, necessitating bespoke solutions. This factor drives the need for customization, which has been associated with lower experience rates. It is also observed in hydropower, where dam design is highly site-specific, and biomass, where the fuel’s physical properties vary across locations.32 Wind energy also faces site-specific differences in wind profiles and ground conditions, but standardized turbine classes can accommodate most of this variation.45 In solar PV, geography has almost no bearing on core system design, enabling very high standardization and scale economies.

The tacit nature of drilling knowledge is largely unique to geothermal and unparalleled in other forms of renewable energy. According to our interviews, much of the expertise resides within localized drilling crews and cannot be easily codified or transferred. This factor sharply contrasts with installation and operational procedures in wind and solar PV, which are much less tacit and well-documented and thus more easily replicated across sites.21 Literature has shown that knowledge transfer between actors is highly relevant for learning in renewables.46,47,48

Regulatory density and variance also impede geothermal cost reductions, as many interviewees pointed out. Environmental and permitting requirements vary widely across jurisdictions and change over time, particularly with respect to subsurface risks such as induced seismicity and groundwater protection. Regulatory variance as a driver of the need for customization and thus reduced experience rates has been described in the literature.32,49 However, while regulatory variance exists to some extent for wind power, the implications differ: In wind, compliance typically involves adapting peripheral systems (e.g., siting or noise control), whereas in geothermal, regulation affects core project design and thus has a major impact on cost.

Knowledge transfer from the oil and gas industry largely benefits geothermal projects but is also important to other clean energy technologies.50 For example, offshore wind benefits from such spillovers in offshore construction techniques and marine project management.51 Importantly, our interviewees stress that geothermal power is also negatively affected when the related oil and gas industry declines (locally).52,53

The combination of the mentioned factors explains the negative historical global ERs. The varying hydrogeology and high regulatory density and variance increase the need to customize each project. Increased customization and (project) design complexity have been shown to negatively correlate with ERs.32 That said, these constraints primarily affect subsurface activities. Surface plant components, such as turbines and heat exchangers, are less site-specific and can more benefit from global knowledge accumulation and established supply chains. Importantly, the variance in hydrogeology and the regulatory variance should be lower at the local scale. Furthermore, tacit knowledge is more easily transferred at a local level. Together, these factors imply higher ERs at the local level.

Local learning dynamics

Shifting from the global to the national level, we find considerable variation in local ERs. Figures 3A–3G present experience curves for the seven countries with the highest number of plants with available cost data, including the six top countries by cumulative deployment and Germany, which has many small projects. We categorize countries into three groups based on their ERs: the United States with a strong positive ER (+18%), New Zealand and Indonesia with strongly negative ERs (−53% and −36%), and countries with near-zero ERs, concretely the Philippines, Türkiye, Kenya, and Germany. Our interviews help explain the differences in ERs as follows.

Figure 3: Country-level costs and experience curves of geothermal power(A–G) Country-level experience rates. Costs are in 2024 USD per W; deployment is given in MW. Vertical axis: Plant costs in 2024 USD per W. Horizontal axis: National deployment in MW. Indicative years are shown for the first project per decade in each country. Details about data coverage and model parameters are in Table S4.

In the United States (ER: +18%), factors driving up cost or limiting experience effects are less pronounced than in other countries, which explains the high ER: Projects are often located in large and geologically homogeneous structures, limiting the effect of diminishing productivity of new sites and varying hydrogeology. At the same time, regulation is less dense and relatively stable, particularly in sparsely populated desert regions, and many projects are viable without subsidies, limiting exposure to changing support policies in the first place. Additionally, strong cross-industry knowledge transfer from the oil and gas industry supports exploration, drilling, and project management and reduces cost without having to invest to gain these experiences within the geothermal industry. Nevertheless, projects and costs are heterogeneous, explaining the wide CI, even though there is a 70% confidence that the ER is positive (Table S4). New Zealand’s negative ER (−53%) is primarily driven by the diminishing productivity of new sites: While early plants were exploiting near-surface reservoirs, newer projects are located in more difficult geological settings, increasing drilling complexity, depth, and thus, costs. Also, in Indonesia (ER: −36%), the diminishing productivity of new sites played out as newer projects face worse geological conditions and higher development costs. On top of that, projects moved into more remote, logistically challenging regions requiring additional infrastructure investments. Türkiye (ER: −9%) and Kenya (ER: +3%) show comparably stable costs due to the high concentration of the projects in place and time. In the Philippines (ER: −7%), recent projects fall into two categories: Some in existing fields are cheaper than older ones due to learning effects, while others in new, less explored regions are costlier, resulting in an overall ER near zero. Germany shows a positive ER (+3%), primarily due to the now-overcome high costs of early, experimental projects. More recent projects, often in the same reservoirs and under more stable policy conditions, benefited from learning effects, despite challenges from regulatory variance. Nevertheless, absolute costs remain high due to dense regulation and complex project design.

Learning potential of future geothermal technologies

Our results show that global learning in hydrothermal geothermal power has been limited, with strong variation across countries. This can be explained—among others—by a high need to customize projects and a scarcity of good sites, which implies a rather steep resource supply curve. Emerging technologies raise hopes among industry and investors for faster cost declines as they promise greater standardization8 and a flatter resource supply curve, but their limited deployment precludes empirical estimates of ERs. Here, we assess the learning potential of these technologies with an arguably differing maturity using a recently successfully implemented approach: expert ratings of the technologies’ two inherent characteristics (1) need for customization and (2) project design complexity.32,33,34 Technologies with low customization needs are better suited to replication and scaling. Meanwhile, complexity can slow learning, driven by factors such as tacit knowledge and system integration demands. Specifically, we had experts rate EGS and CLGS’ characteristics along with hydrothermal’s characteristics as a baseline for comparison. A lower complexity and need for customization rating of EGS and CLGS compared to hydrothermal would signify that these technologies can learn faster than hydrothermal, while higher ratings would predict slower learning.

For hydrothermal technology (see Figure 4A for a schematic), expert ratings confirm the above qualitative finding that hydrothermal systems are rated as having considerable project design complexity (2.0) and considerable customization needs (2.4), with the greatest variation in ratings among the three technologies (Figure 4D). This heterogeneity reflects the importance of local context in shaping expert perceptions, particularly differences in geology and regulatory density. For emerging geothermal technologies to achieve higher and more consistent ERs, they would need to reduce dependence on local subsurface conditions and regulatory adaptation. Minimizing customization needs, especially in regions with heterogeneous hydrogeology, is a critical prerequisite for scalable cost reductions.

Figure 4: Current and future geothermal technologies and expert assessments of their complexity and need for customizationTechnical sketches of existing hydrothermal systems (A) and future enhanced (B) and closed-loop systems (C) along expert assessments of complexity and need for customization of the technologies (D–F). N = number of experts that rated each technology. Large circles represent average expert ratings; small circles represent individual ratings. High complexity reflects intricate components and interdependencies, while high customization captures the need to adapt designs to local geological or regulatory conditions. Detailed rating statistics and supporting interview quotes are in Tables S6 and S7, respectively.

The first emerging technology, EGS, enhances reservoirs by creating artificial fractures using oil and gas industry techniques (Figure 4B). These systems are typically constructed with two or more horizontal wells connected by a network of fracked faults. Experts rated the customization needs of EGS slightly lower than hydrothermal’s (2.4, Figure 4E). Two counteracting dynamics are at play: resource uncertainty is lower, as natural permeability is no longer required, but customization demands rise due to the need to adapt technologies such as fast drilling and fracking to site-specific geology. For instance, achieving competitive drilling speeds depends on optimizing the interaction between drill bit, drilling mud, and operational parameters such as weight on bit, requiring experienced rig crews familiar with local conditions. Experts rated the complexity of EGS higher than hydrothermal systems, at 2.5. In addition to drilling, maintaining adequate long-term permeability adds further technological challenges. Regulatory complexity also increases in densely populated areas, for example, in Central Europe, where fracking-related seismic events have triggered public skepticism and political resistance.54

The second technology, CLGS, uses artificially constructed heat exchangers with numerous thin, sealed wells and does not rely on natural fractures or underground water (Figure 4C). As there is no direct contact between the working fluid and the rock, chemical interactions are eliminated. The system relies on an extensive network of narrow wells, increasing total drilling distance by a factor of more than 10 compared to hydrothermal systems (see Note S2).

Experts rated CLGS slightly lower in terms of customization needs (2.2, Figure 4F). This lower value stems from its adaptability to user-specific demands, for example, by adjusting power output or temperature via wellfield design, and from avoiding natural flow dependency. However, advanced drilling technologies still require local adaptation. CLGS’ project design complexity was rated at 2.4, slightly lower than EGS, but higher than hydrothermal. The system’s complexity arises from the need for long horizontal drilling, effective fracture sealing, and highly efficient drilling practices to remain economically viable. That said, CLGS eliminates certain complexities: Sealed systems avoid fluid-rock interactions, and thermosyphon-driven circulation avoids the need for pumps, which is a common reliability burden in hydrothermal systems.27

In summary, EGS and CLGS exhibit inherent characteristics that suggest learning may not accelerate drastically on a global scale. Some of the barriers that limited hydrothermal learning can be mitigated. For example, EGS and even more so CLGS reduce reliance on natural water flow and chemistry and local rock properties, which suggests a flatter resource supply curve and faster learning. However, EGS and even more so CLGS come with increased challenges to project design complexity. Thus, EGS and CLGS are unlikely to enable rapid global learning and have experience rates that are still similar to hydrothermal. However, similar to the past learning of hydrothermal technology in the USA, higher local learning is possible, especially in regions with homogeneous hydrogeology and where cross-industry knowledge spillovers can facilitate effective management of unavoidable project complexity.32