STAR★Methods

Key resources table

REAGENT or RESOURCE | SOURCE | IDENTIFIER
Deposited data

Geothermal power plant database (514 plants, 206 with cost data) | This study | Available on request to T.S.S.
Bloomberg New Energy Finance (BNEF) geothermal data | Bloomberg New Energy Finance | https://about.bnef.com/
Country update reports from geothermal world/continental congresses | Various | https://worldgeothermal.org/geothermal-data/conference-paper-database
ThinkGeoEnergy portal data | ThinkGeoEnergy ehf. | https://www.thinkgeoenergy.com/
IMF exchange rate data | International Monetary Fund | https://data.imf.org/
Federal Reserve exchange rate data | Federal Reserve System | https://fred.stlouisfed.org/
US Consumer Price Index | U.S. Bureau of Labor Statistics | https://www.bls.gov/cpi/
Analysis code | This study (GitHub) | GitHub: https://github.com/fmueller93/GeothermalLearning2025; Zenodo: https://doi.org/10.5281/zenodo.21776541

Software and algorithms

Excel | Microsoft | https://www.microsoft.com/en-us/microsoft-365/excel
MAXQDA | Verbi Software | https://www.maxqda.com/
Python v3.14 | Python Software Foundation | https://www.python.org/
pandas v2.3 | PyPI | https://pypi.org/project/pandas/
NumPy v2.4 | PyPI | https://pypi.org/project/numpy/; RRID:SCR_008633
openpyxl v3.1 | PyPI | https://pypi.org/project/openpyxl/
statsmodels v0.14 | PyPI | https://pypi.org/project/statsmodels/; RRID:SCR_016074
SciPy v1.17 | PyPI | https://pypi.org/project/scipy/; RRID:SCR_008058
Matplotlib v3.10 | PyPI | https://pypi.org/project/matplotlib/; RRID:SCR_008624
GeoPandas v1.1 | PyPI | https://pypi.org/project/geopandas/

Other

Interview guide/questionnaire | This study | Table S9
Expert rating framework for technology characteristics | Malhotra and Schmidt32; Sievert et al.33 | https://doi.org/10.1016/J.JOULE.2020.09.004; https://doi.org/10.1016/J.JOULE.2024.02.005

Method details

Quantitative data and analysis

Approach and scope

We study geothermal power plants built since 1950 (when the first modern plants came online) and put into operations before 1st January 2025.35,70 To our best knowledge, all global electricity-producing plants that came online since 1950 are included in the data, including later-decommissioned plants. In total, there are 514 plants that are or have been operational. In addition, the database covers 439 plants that are planned, abandoned, or are under construction. 19 of the abandoned projects were drilled to a large part, so that we count them for experience and cumulative capacity (see below).

Data sources and verification

Initial data came from multiple sources, most importantly Bloomberg New Energy Finance (BNEF), the most recent country update reports71 from the geothermal world or continental congresses, the global portal ThinkGeoEnergy,72 and national portals. In the second step, this data was consolidated, and we verified the completeness of the list of plants. Third, we cross-checked and filled missing values, using resources from national and local governments, industry organizations, and operators and suppliers. Furthermore, industry contacts, many of them made during several industry congresses, were used to find resources and verify data. Specific data points, especially outliers and important data gaps were filled during interviews (see below) and through individual requests per mail.

Projects considered for the calculation of cumulative capacity

Cumulative capacity includes all projects that came online since 1950. The initial capacity in 1950 is assumed to be at 300 MW and fully located in Italy.35,70 The cumulatively deployed national and global capacity is calculated from all projects until the start of the year in which a plant became operational. Thus, all projects that came online within one year are assigned the same cumulative capacity. Included in the cumulative capacity are all projects that finalized their drilling phase, even if the plant did not become operational. The underlying reasoning is that experience is gained even if a project does not go operational, as the operators gather experience about drilling the local geology and local hydrogeology, which is decisive for project costs. Projects that were abandoned before a significant amount of drilling was done, including projects that did not drill more than a dedicated exploration well, are not included in the cumulative capacity.

Converting currencies and deflating

Values are given in USD2024. Whenever the input data was presented in another currency, we converted it to USD for the respective year using annual exchange rates from the International Monetary Fund73 or the Federal Reserve System.74,75 Subsequently, nominal values were converted to real (2024) USD by applying the US Consumer Price Index.76

In the regression models that adjust for exchange rate fluctuations, we introduce a correction factor that normalizes investment costs across different currency areas to a consistent base year exchange rate.38

Cost data sample

Cost data used for the regression models is available for 206 out of the 514 operational or formerly operational plants (40%). Cost data coverage tends to be higher for more recent projects and is scarce for projects built before 2000 and also varies by country (see Table S2). We do not expect a systematic bias on the cost data from these limitations.

One plant is considered an outlier and manually removed from the dataset (Bad Blumau plant in Austria). This plant is primarily used for heat generation and has less than 1 MW electrical capacity, thus exceeding a specific cost of 100 USD/W.

Experience curve model

We use the well-established29,33,38 formulawhere Ct is the plant cost in year t, C0 is the plant cost in year 0, xt is the cumulative capacity in year t, x0 is the cumulative capacity in year 0, and b is the rate of cost reduction.13

Ct=C0(xtx0)b,

The ER, i.e., the percentage cost reduction from a doubling of cumulative capacity, is then given by

ER=1−2b

For our empirical estimation of the ER, we can rewrite the first equation in its logarithmic form:where α = log(C0), β = b, Dt is a vector of control variables, and ϵt is the error term.

log(Ct)=α+β·log(xt)+δDt+ϵt,

Regression model

Two model specifications (Models 2 and 4) exclude small projects (with less than 5 MW output), as a skewing of the results was expected due to the increased popularity of such plants (Figure 2A) in recent years and the suspected negative scale effect, i.e., the expectation that these projects would be expensive.19 Models 3 and 4 correct for exchange rate fluctuations, which can be suspected to skew the observed ERs due to the long time horizon of observations and many plants in soft-currency countries.38 All models use ordinary least squares regressions with heteroscedasticity-robust standard errors.

Qualitative data and analysis

Expert selection

The expert selection considered three aspects. First, we selected experts with a broad view of the industry and actor network. Second, we aimed at covering all relevant parts of the value chain. Third, we looked for experts to broaden and balance the geographic distribution of experts.

An initial selection of experts was found through industry reports and publications. Experts were approached by the authors at eight industry congresses and five site visits between October 2023 and October 2025, over LinkedIn, and by e-mail.

Experts were contacted on a rolling basis. Due to contacts from the network and direct referrals, the response rate to our requests was high. Only 20 experts that we contacted did not respond to our interview request. We continued recruiting experts until no additional relevant insights were observed.77 An anonymized list of interviewees is provided in Table S8.

Interview structure

The first interviews were conducted in an explorative way to refine a questionnaire for subsequent in-depth interviews. Based on their expertise, experts were selected either for long interviews covering the entire questionnaire, for short interviews covering only specific aspects of the questionnaire, or for interviews to classify the future technologies EGS and CLGS (see below). Interviewees were sent an interview guide in advance that allowed them to prepare for the interview, see Table S9. The interview guide provided background information on the research project and explained the objectives of the interview. Interviews were held under the Chatham House rule, which states that “participants are free to use the information received, but neither the identity nor the affiliation of the speaker(s) [.] may be revealed.”78 In total, 51 interviews were conducted between October 2023 and May 2025. Short interviews took between 10 and 20 min; long interviews took between 50 and 120 min. Each interview was preceded by desk research about the interviewee’s expertise and geothermal plants in which they were involved. Interviews were conducted online or in person, by one or two researchers, in English or German. The discussions were recorded and automatically transcribed. In one case where the interviewee did not consent to recording and for some short interviews which were held at conference venues the interviewers took detailed individual notes.

Interview coding

Interview transcripts and notes were analyzed using software for computer-assisted qualitative analysis (MAXQDA 24).79 The analysis followed a systematic labeling of each statement. In total, 280 segments were labeled with one or more of 25 different codes and consolidated subsequently.

Approach to derive technology characteristic ratings for future technologies

Eleven qualified experts were asked to assess existing and future geothermal technologies regarding technology-inherent characteristics following ref.28 The experts were asked to rate hydrothermal, EGS and CLGS in terms of their project design complexity and need for customization on a scale from 1 to 6, which, for both metrics, had other technologies given as references. The ratings were later converted to a range from 1 to 3. For technology characteristics, we introduced one variation from the original framework by Malhotra and Schmidt: Instead of “design complexity”, we use “project design complexity” in this study, to account for the fact that many of the challenges associated with geothermal energy have to do with project management and are not limited to engineering design.

Value and validity of expert elicitation to estimate ERs

A promising approach to estimate experience rates for emerging technologies leverages expert elicitations not to predict future costs directly but to assess technology-inherent characteristics that are empirically linked to cost reduction potential. This approach proves more reliable than asking experts directly for cost estimation, a task at which experts performed poorly, especially in nascent technologies.80,81 Meng et al. (2021) demonstrate that model-based projections using experience curves have outperformed expert elicitations in forecasting energy technology costs,18 yet constructing experience curves for novel technologies is hampered by limited deployment data. Sievert et al. (2024) address this gap by asking experts to rate the design complexity and customization needs of technology components, which can be judged reliably by experts,80 and then mapping these ratings onto empirically grounded experience rate distributions following the technology typology of Malhotra and Schmidt (2020). Crucially, Sievert et al. (2024) validated this method ex-post against observed cost trajectories for concentrated solar power, onshore wind, and solar photovoltaics, with projected cost ranges overlapping 83–100% with actual costs. This suggests that expert-assessed technology characteristics offer a robust basis for deriving experience rates where historical deployment data are insufficient, as is the case for EGS and CLGS.

Quantification and statistical analysis

All statistical analyses were performed using Python and the packages listed in the key resources table above. The analysis code is available at https://github.com/fmueller93/GeothermalLearning2025.git.

Global and country-level experience rates were estimated using ordinary least squares (OLS) regression on the log-linearized experience curve model. All regression models use heteroscedasticity-robust standard errors (White standard errors). The sample size for the global models is n = 206 plants with cost data (or n = 173 for models excluding plants below 5 MW). For country-level models, sample sizes vary by country and are reported in Table S4. In each case, n represents the number of individual geothermal power plants.

Confidence intervals (CIs) for the experience rate estimates are reported at the 90% level (Figures 2 and 3; Tables S3 and S4). Statistical significance of the experience rate (i.e., whether the learning parameter b differs from zero) is assessed based on whether the CI excludes zero. Detailed regression parameters, including coefficients, standard errors, p-values, and R2 values, are reported in Tables S3 (global models) and S4 (country-level models).

For the qualitative analysis, interview data were coded using MAXQDA 24. A total of 280 coded segments were labeled with 25 codes. Descriptive statistics of expert ratings for technology characteristics (project design complexity and need for customization) are reported as means with individual ratings shown in Figures 4D–4F and detailed statistics in Table S6.

One outlier (Bad Blumau plant, Austria) was excluded from the cost dataset based on predefined criteria: the plant is primarily used for heat generation, has less than 1 MW electrical capacity, and exhibits a specific cost exceeding 100 USD/W, which is more than an order of magnitude above other plants.

Additional resources

Analysis code repository: The Python code used for all quantitative analyses in this study is publicly available at https://github.com/fmueller93/GeothermalLearning2025.git.