Artificial Intelligence-Driven Identification of Favorable Geothermal Sites Based on Radioactive Heat Production: Case Study from Western Türkiye

Abstract

In recent years, the exploration and utilization of geothermal energy have received growing attention as a sustainable alternative to conventional energy sources. Reliable, data-driven identification of geothermal reservoirs, particularly in crystalline basement terrains, is crucial for reducing exploration uncertainties and costs. In such geological settings, magnetic susceptibility, radioactive heat production, and seismic wave characteristics play a vital role in evaluating geothermal energy potential. Building on this foundation, our study integrates in situ and laboratory measurements, collected using advanced sensors from spatially diverse locations, with statistical and unsupervised artificial intelligence (AI) clustering models. This integrated framework improves the effectiveness and reliability of identifying clusters of potential geothermal sites. We applied this methodology to the migmatitic gneisses within the Simav Basin in western Türkiye. Among the statistical and AI-based models evaluated, Density-Based Spatial Clustering of Applications with Noise and Autoencoder-Based Deep Clustering identified the most promising and spatially confined subregions with high geothermal production potential. The potential geothermal sites identified by the AI models align closely with those identified by statistical models and show strong agreement with independent datasets, including existing drilling locations, thermal springs, and the distribution of major earthquake epicenters in the region.

1. Introduction

Despite its vast potential, geothermal energy currently accounts for only 0.34% of global energy production [1]. Nevertheless, it holds considerable promise as a sustainable and renewable energy source, due to its low environmental impact, high efficiency, and capacity to reduce dependence on fossil fuels [2]. By harnessing the Earth’s internal heat from beneath the crust, geothermal energy enables a wide range of economically viable applications, including electricity generation, residential heating, and industrial processing [3,4].

Geothermal energy production is primarily driven by the temperature gradient within the Earth’s crust, which results from radioactive heat production (RHP) due to the decay of radioactive isotopes. Short-lived isotopes such as ²⁶Al, ³⁶Cl, and ⁶⁰Fe were dominant sources of radioactive heat during the early stages of Earth’s formation. In contrast, long-lived isotopes, including ²³⁵U, ²³⁸U, ²³²Th, and ⁴⁰K, have become the primary contributors to sustained heat generation over geological timescales. Tectonic activity in geodynamically active regions further modulates the heat flux within the crust, with a significant portion of this thermal energy originating from the ongoing radioactive decay of these long-lived isotopes [5]. In particular, the decay of uranium (U) and thorium (Th) isotopes within these rocks produces heat that can be harnessed for geothermal energy [6].

The geothermal exploration process has an integrated approach involving geology, geochemistry, hydrogeology, and geophysics. Field investigations are conducted to analyze the structural framework and petrographic characteristics of the region, supported by detailed geological mapping. Hydrogeologic studies aim to define aquifer systems and elucidate the subsurface flow mechanisms of geothermal fluids, including their interactions with bedrock. Geochemical analyses of water and rock samples from surface and borehole sources help characterize the hydrochemical composition and origin of geothermal fluids, providing insight into the reservoir, cap rock, and heat source. Geophysical investigations incorporate both surface and borehole techniques to delineate the spatial extent of geothermal systems. Surface methods, including gravity, magnetic, electrical, seismic, and electromagnetic surveys, are used to identify the cap rock, reservoir zones, and fluid pathways. Borehole geophysics is used to assess formation properties and temperature profiles. More recently, data from geological and geophysical surveys have been used to develop synthetic subsurface models, enhancing the understanding and visualization of geothermal systems [7,8,9]. Additionally, a two-dimensional electrical conductivity model derived from magnetotelluric data has been utilized to reveal electrically anisotropic features associated with tectonic deformation and magmatic intrusions. This model has identified a deeper hydrothermal system above a magmatic structure, providing valuable insights for future geothermal exploration [10].

In recent years, the integration of artificial intelligence (AI) into geothermal energy exploration has garnered significant attention for its potential to improve the efficiency and accuracy of geothermal resource assessment. AI offers innovative solutions for processing and interpreting complex datasets, thereby improving the efficiency and accuracy of geothermal energy evaluations. AlGaiar et al. [9] investigated the role of AI in geothermal resource exploration, highlighting its potential to improve efficiency, data interpretation, and prediction accuracy, key factors in advancing sustainable renewable energy solutions. Duan et al. [11] presented an AI-based approach for clustering and classifying geothermal reservoirs in the Ying-Qiong Basin, China. By applying AI algorithms to geological and geophysical data, the authors identified patterns and grouped reservoirs with similar characteristics. Their study demonstrated the potential of AI to automate and refine the classification process, providing valuable insights for more effective reservoir management and resource utilization.

The novelty and primary contribution of this study lie in integrating spatially variable magnetic susceptibility, seismic, and geochemical data, while employing radiogenic heat production (RHP) as a key indicator for identifying geothermally favorable sites. This integration is a necessary step, as the critical importance of seismic and geochemical data in assessing geothermal potential has been well-documented [12,13]. In addition, RHP is closely linked to the concentration of heat-producing elements, primarily U, Th, and K, which are significant contributors to subsurface thermal anomalies. Accordingly, RHP-based methods have been widely employed for the identification of geothermal heat sources and the preliminary assessment of geothermal energy potential [14,15,16,17,18,19,20,21,22]. In this study, we also introduce a novel modeling framework that uses AI-based and statistical clustering algorithms to enhance prediction accuracy in locating potential geothermal-rich zones. The predictive performance of these models is tested on Precambrian-age migmatitic gneisses of the northern Menderes Massif, located within the Simav Basin in western Türkiye. Furthermore, we examine the spatial relationships between AI-predicted geothermal sites and earthquake epicenters to explore potential links between tectonic activity and the localization of geothermal resources.

2. Study Area

The study area is located in the Simav Basin in the northern Menderes Massif of western Anatolia, Türkiye (Figure 1). This region forms part of the Alpine–Himalayan orogenic belt and has undergone significant extensional tectonics from the late Neogene to the present. This setting provides a geodynamically active environment shaped by complex structural deformation and magmatic activity. Within this tectonic framework, the Demirci, Simav, and Dursunbey basins in the northeastern Menderes Massif have been the focus of extensive multidisciplinary research due to their pull-apart basin structures, magmatic–metamorphic evolution [23,24,25,26,27,28], and rich geothermal resources.

The Simav region and its surroundings have emerged as one of Türkiye’s prominent geothermal energy zones. Numerous studies, including those by [29,30], have evaluated the hydrogeochemical properties and reservoir potential of the area’s thermal fluids. Geothermal resources in Simav are actively used for residential and greenhouse heating, as well as electricity generation. The heat potential of the Simav geothermal field is 1.3 × 1014 kJ, and the electricity generation potential is 26 MW, with an average probability of 50%, while the most plausible value has been calculated as 19 MW [30].

Figure 1. (a) The location of the study area on a regional map, modified from [31,32]. (b) A 1:100,000-scale geological map of the Kütahya J21 and J22 quadrangles, modified from [33,34]. The base elevation data are derived from the Terra Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global Digital Elevation Model (GDEM), Version 3, covering the N39E028 and N39E029 grids, with a spatial resolution of approximately 30 m. (c) Study area in the Simav Basin.

Figure 1. (a) The location of the study area on a regional map, modified from [31,32]. (b) A 1:100,000-scale geological map of the Kütahya J21 and J22 quadrangles, modified from [33,34]. The base elevation data are derived from the Terra Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global Digital Elevation Model (GDEM), Version 3, covering the N39E028 and N39E029 grids, with a spatial resolution of approximately 30 m. (c) Study area in the Simav Basin.

The geological map and column section of the study area are presented in Figure 1 and Figure 2, respectively. The geothermal energy resources in the Simav Graben, where the study area is located, are associated with lithologies that possess potential for heat generation, in addition to featuring permeable reservoir rocks and impermeable cap rocks essential for geothermal system development [35]. This study focuses on the migmatitic gneisses of the Kalkan Formation, which constitute the basement rocks of the Menderes Massif and exhibit significant potential for RHP. The formation displays depth-dependent brittle and ductile deformation features [36,37], indicating complex tectonometamorphic processes. Normal faults within the basin, along with intersecting transfer faults, are believed to facilitate the upward migration of geothermal fluids. The most lithologically suitable reservoir rocks for hosting geothermal fluids within the basin include the limestones of the Budağan Formation, marbles of the Balıkbaşı Formation, and quartzite layers within the Simav Metamorphites (Figure 2).

The Budağan Limestone Formation contains interbedded banded schist layers, reflecting a complex depositional and metamorphic history [35,39,40]. The Balıkbaşı Formation is notable for its extensive fracturing and the development of micro- to macro-scale dissolution features, including karstic voids, which significantly enhance its capacity as a geothermal reservoir [34]. Additionally, the quartzite layers within the Simav Metamorphites demonstrate favorable reservoir characteristics due to their structural integrity and permeability [34,35].

The Simav Graben, which cuts across the northern Menderes basins, trends predominantly in a north–south direction and is structurally defined by high-angle normal faults [41]. In the context of geothermal fluid transport within the graben, zones where transfer faults intersect with these high-angle normal faults are of particular significance, as they frequently serve as key conduits for fluid migration. Fault focal mechanism analyses of ongoing seismicity north of Simav further underscore the importance of regions where strike-slip and normal faults converge. These structurally complex zones are likely to enhance the permeability and facilitate the upward migration of geothermal fluids, making them critical targets for exploration.

The Simav Graben is located on the active Simav fault zone, which was formed as a result of the extensional tectonic regime in the northern Menderes Massif region. The approximately east–west-trending and north-trending Simav active fault cuts north of the Demirci and Selendi Basins in the NE-SW trending northern Menderes Massif, which was opened in the early Miocene [26,41,42,43].

As discussed above, the geothermal systems in the Simav Basin are closely linked to active fault zones. These faults act as conduits for the upward migration of deep hydrothermal fluids. The presence of elevated heat flow and surface hot springs near fault intersections strongly support this structural control. Notably, the increase in seismic activity since 2000 indicates ongoing tectonic deformation, which further facilitates fluid circulation within the geothermal reservoirs [44].

A notable example of this tectonic–geothermal interaction was observed during the Mw = 5.8 earthquake that occurred on 19 May 2011 in the Simav region. Post-earthquake geophysical and hydrogeological observations revealed significant changes in the temperature and discharge rates of several geothermal springs, suggesting a dynamic link between seismicity and geothermal fluid pathways [45].

3. Materials and Methods

3.1. Sampling and Analytical Methods

Field sampling was conducted on the migmatitic gneisses within the J21 and J22 quadrangles on the 1:25,000-scale geological map (Figure 1). A total of 90 rock samples were collected from various locations within the Kalkan Formation. These samples underwent geochemical analysis to determine the concentrations of major oxides, including Si${\mathrm{O}}_2$, A${\mathrm{l}}_2$${\mathrm{O}}_3$, ${\mathrm{K}}_2$O, CaO, MgO, FeO, F${\mathrm{e}}_2$${\mathrm{O}}_3$, Ti${\mathrm{O}}_2$, ${\mathrm{P}}_2$${\mathrm{O}}_5$, and MnO. The following elements are present in the compound: C${\mathrm{r}}_2$${\mathrm{O}}_3$, BaO, SrO, and trace elements, including Rb, Ni, Nb, Y, Ta, V, W, U, Th, Zr, Pb, Cu, Zn, Mo, Co, Cd, Re, Au, Ag, Hg, Sn, As, Sb, Cl, Se, and S. Descriptive statistics of the measured geochemical variables are presented in Appendix A. Magnetic susceptibility measurements were also conducted on these rock samples at both low and high frequencies, resulting in mass-dependent and frequency-dependent magnetic susceptibility values.

Seismic data for the study area between 1 January 1900 and 27 March 2025 were acquired from the Boğaziçi University Kandilli Observatory and Earthquake Research Institute (KOERI) Earthquake Query System (http://www.koeri.boun.edu.tr/sismo/zeqdb/, accessed on 27 March 2025). The earthquake magnitude and frequency distribution is as follows: 0–1 (4306 events), 1–2 (2434 events), 2–3 (2909 events), 3–4 (599 events), 4–5 (104 events), and 5–6 (6 events). Notably, no seismic events exceeding magnitude 6 were recorded in the region during the specified period.

Additionally, the designated study area includes 59 geothermal boreholes, as reported by the General Directorate of Mineral Research and Exploration. Of these, approximately 10 are classified as high-temperature geothermal wells, while the remaining wells fall into the medium- to low-temperature categories.

3.2. Laboratory Analyses

Laboratory analyses consisted primarily of magnetic susceptibility measurements and assessments of radioactive heat production.

3.2.1. Magnetic Susceptibility Measurement

Magnetic susceptibility is a well-established geophysical technique used to evaluate the magnetic properties of rocks and sediments. It quantifies the extent to which a material becomes magnetized in response to an external magnetic field, providing valuable insights into the mineralogical composition of geological materials, particularly the presence and concentration of magnetic minerals such as magnetite. For instance, rocks enriched in ferromagnetic minerals like magnetite typically exhibit high magnetic susceptibility, while those with low iron oxide content tend to show much lower values. In this study, magnetic susceptibility is expressed in SI units (${\mathrm{m}}^3$/kg). These measurements offer a cost-effective, non-destructive method for characterizing the magnetic and compositional properties of subsurface materials, making them particularly useful in geological and geothermal investigations [46].

In this study, magnetic susceptibility measurements were performed using the MS2B sensor of the Bartington Instruments system. Rock samples were first crushed and ground to a grain size of approximately 75 $\mathsf{\mu }$m using a tungsten carbide vibratory grinder to ensure homogeneity and measurement consistency. For measurement purposes, samples were shaped into cylindrical forms with a diameter of approximately 10 mm and a length ranging between 10–20 mm. For measurement, the powdered samples were formed into cylindrical shapes with a diameter of approximately 10 mm and a length ranging from 10 to 20 mm. Prior to measurement, all samples were carefully cleaned and oven-dried to remove any moisture or contaminants that could potentially affect the results [46]. The prepared samples were then placed into 10 cc plastic pots and inserted into the sensor’s measurement chamber. Readings were recorded at both low (0.47 kHz) and high (4.7 kHz) frequencies, with a sensitivity of 0.1 SI units, to allow for detection of frequency-dependent susceptibility. To calculate the mass-specific magnetic susceptibility, the measured magnetic susceptibility values ($\chi$) were normalized by the corresponding sample densities ($\rho$), following the methodology described by Dearing et al. [46]:

$$\begin{array}{cc}\hfill {\chi }_{LF}& =\chi /\rho ,\hfill \\ \hfill {\chi }_{HF}& =\chi /\rho \hfill \end{array}$$

(1)

The percent frequency-dependent magnetic susceptibility, ${\chi }_{FD}$%, is derived from the values obtained from Equation (1) through

$$\begin{array}{c}\hfill {\chi }_{FD}\%=[({\chi }_{LF}-{\chi }_{HF})/{\chi }_{LF}]\times 100\end{array}$$

(2)

The mass-specific and frequency-dependent magnetic susceptibility, ${\chi }_{FD}$, is calculated using the values obtained from Equation (1), along with the density:

$$\begin{array}{c}\hfill {\chi }_{FD}=({\chi }_{LF}-{\chi }_{HF})/\rho .\end{array}$$

(3)

3.2.2. Radioactive Heat Production (RHP)

Radioactive heat generation is the process by which heat energy is released through the decay of radioactive isotopes (e.g., uranium (U), thorium (Th), potassium (K)) in rock-forming minerals [13,47]. This form of heat production plays a critical role in assessing the geothermal energy potential of a region.

In this study, RHP potential was determined using X-ray fluorescence, XRF, measurements. The activity concentrations of radioactive isotopes present in the sample can be derived from these elemental values, which in turn are used to estimate the RHP of the samples. The spatial distributions of U, Th, and K are presented in Figure 3.

As shown in the figure, the eastern part of the Simav district exhibits higher concentrations and greater variability in U, Th, and K compared to the western region, which is dominated by migmatitic rocks. Notably, Th concentrations are significantly elevated in the eastern region, while they remain consistently low in the western part of the study area.

The RHP potential at the sampling sites can be estimated using the isotope concentrations and the correlative relationships given in Equation (4). In these expressions, RHP represents the amount of heat energy released per unit volume of the sample as a result of the radioactive decay of isotopes. Based on data from 90 sampling locations, these relationships yielded RHP values with a coefficient of determination, $R^2$, of 1.00 and root mean square error, RMSE, values ranging from 0.01 to 0.04 $\mathsf{\mu }$Wk${\mathrm{g}}^{-1}$. These results indicate that any of the listed expressions can be reliably applied to estimate RHP in the study region.

RHP	$=\rho \times \left[\left(9.7C_U\right)+\left(2.7C_{Th}\right)+\left(3.6C_K\right)\right]\times {10}^{-5},$	[14]	(4)
	$=\rho \times \left[\left(9.77C_U\right)+\left(2.63C_{Th}\right)+\left(3.4C_K\right)\right]\times {10}^{-5},$	[15]
	$=\rho \times \left[\left(9.67C_U\right)+\left(2.63C_{Th}\right)+\left(3.5C_K\right)\right]\times {10}^{-5},$	[16]
	$=\rho \times \left[\left(9.52C_U\right)+\left(2.56C_{Th}\right)+\left(3.48C_K\right)\right]\times {10}^{-5},$	[17]
	$=\rho \times \left[\left(0.097C_U\right)+\left(0.026C_{Th}\right)+\left(0.036C_K\right)\right]\times {10}^{-3},$	[18,19]
	$=\rho \times \left[\left(0.0957C_U\right)+\left(0.0256C_{Th}\right)+\left(0.0348C_K\right)\right]\times {10}^{-3},$	[17,18]
	$=\rho \times \left[\left(0.0718C_U\right)+\left(0.193C_{Th}\right)+\left(0.262C_K\right)\right]\times 1.325,$	[19,20]

3.3. Cluster Analysis

The RHP estimates obtained using Equation (4) do not account for the effects of magnetic susceptibility and seismic velocity, both of which directly influence RHP, as discussed in Section 3.2.2. To address this limitation, we applied three statistical models and three unsupervised artificial intelligence (AI) models to identify clusters of the most favorable sites for geothermal energy exploration in the study region. These models used magnetic susceptibility, seismic velocity, and geochemical variables as input features for clustering. A list of geochemical variables used in the analysis is provided in Section 3.1. All features were normalized. This is a critical step, as some algorithms, such as KMeans, are sensitive to the scale of the data. Normalization also mitigates the risk of features with larger ranges dominating the clustering process. The target variable is RHP, which does not require standardization.

3.3.1. Cluster Analysis Using Statistical Models

For the clustering analysis, we used three statistical models, including the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) [48], the Gaussian Mixture Model (GMM) [49], and the Agglomerative Clustering Model (ACM) [49]. These models were selected to capture a range of clustering behaviors—DBSCAN for its ability to detect arbitrary-shaped clusters and outliers, GMM for its probabilistic approach to overlapping clusters, and ACM for its hierarchical structure that does not require a predefined number of clusters.

DBSCAN is a density-based clustering algorithm that does not require the user to predefine the number of clusters. Unlike partitioning algorithms such as k-means, DBSCAN can identify clusters of arbitrary shape and size, which is particularly advantageous when working with complex geospatial datasets or data with non-uniform density distributions. The algorithm groups together points that are closely packed, i.e., those in high-density regions, while labeling points in low-density areas as noise or outliers. This ability to differentiate between dense clusters and sparse regions makes DBSCAN well-suited for identifying natural groupings in geophysical and geological data, where meaningful structures often do not conform to simple geometric boundaries.

GMM adopts a probabilistic framework that allows data points to belong to multiple clusters with varying degrees of membership or probabilities, unlike k-means, which assigns each piece of data exclusively to a single cluster. This soft clustering approach provides greater flexibility, making GMM particularly effective for identifying complex, overlapping clusters with different shapes, sizes, and densities. GMM assumes that the data is generated from a mixture of several Gaussian (normal) distributions, each representing a cluster. The model implements the Expectation-Maximization algorithm to iteratively estimate the parameters, including means, covariances, and mixing coefficients, of the underlying Gaussian distributions, enabling more nuanced clustering compared to hard-assignment methods like k-means. This probabilistic nature also allows GMM to model uncertainty in cluster assignments, making it especially useful in applications where data points do not clearly belong to one cluster or where cluster boundaries are ambiguous. As a result, GMM offers a more nuanced and flexible clustering solution compared to hard-assignment methods like k-means.

ACM is a hierarchical clustering technique that builds a nested cluster structure by iteratively merging the closest pairs of data points or existing clusters based on a defined linkage criterion (e.g., single, complete, or average linkage). Unlike partitioning methods such as k-means, ACM does not require the number of clusters to be specified in advance. One of the key strengths of agglomerative clustering is its flexibility in identifying clusters with complex shapes, sizes, and densities, as it makes no assumptions about the underlying cluster geometry. This makes it particularly suitable for datasets where cluster boundaries are non-linear or irregular.

For these statistical models, we applied Principal Component Analysis (PCA) to reduce the dimensionality of the feature set while preserving as much variance as possible. PCA achieves this by projecting the data onto a new set of orthogonal axes—principal components—that represent the directions of greatest variance in the data, providing a simplified yet informative representation of the dataset. In our case, the feature space was reduced to two principal components. All statistical models were subsequently applied to the PCA-transformed variables to enhance interpretability and computational efficiency.

We then applied the Elbow method to determine the optimal number of clusters, ensuring a consistent cluster count was used across all model results. The elbow point, identified in the inertia plot, marks where the rate of decrease in inertia begins to level off. In our analysis, this method indicated that four clusters provided an optimal balance between model complexity and explanatory power. The identified optimal number of clusters was subsequently used as an input parameter for both the statistical and AI-based clustering models.

3.3.2. Cluster Analysis Using Unsupervised Artificial Models

For the clustering analysis, we used three AI-based methods, including Autoencoder-based Deep Clustering (ABDC) [50], Self-Organizing Maps (SOM) [51], and Deep Embedded Clustering [52], to detect complex patterns and clusters that traditional algorithms like k-means often fail to capture, particularly in cases involving non-spherical or highly overlapping clusters.

ABDC integrates unsupervised learning with deep neural networks to extract meaningful compact feature representation from high-dimensional data. In this framework, an autoencoder is first trained to compress the input into a low-dimensional latent space that captures the underlying structure and key patterns of the dataset. Once the latent representation is learned, the k-means algorithm is applied to this reduced feature space to identify initial cluster assignments. The model iteratively refines the learned features and the cluster assignments, by minimizing both reconstruction error and clustering loss. This feedback loop continuously aligns the learned feature space with the clustering task, leading to more coherent and accurate cluster formation. As a result, the approach is particularly effective for complex, nonlinear data where traditional clustering algorithms may struggle.

SOM is an artificial neural network model, particularly well-suited for clustering and visualizing high-dimensional data. It projects complex data onto a low-dimensional, typically two-dimensional, grid, while preserving the topological relationships of the input space. The algorithm organizes input data based on similarity, so that neighboring nodes (or neurons) on the grid respond similarly to similar inputs, effectively mapping intricate data patterns into an interpretable spatial layout. This spatial organization enables SOM to reveal intrinsic structures and relationships within complex datasets, making them highly effective for exploratory data analysis and pattern recognition. Additionally, SOM are robust to noise and capable of capturing nonlinear relationships, allowing them to identify subtle clusters and data transitions.

DEC is an advanced AI-based clustering method that integrates deep learning with unsupervised clustering. DEC begins by training an autoencoder to learn a compact, low-dimensional latent representation of the input data, effectively capturing complex and nonlinear patterns that traditional feature extraction methods may miss. Unlike conventional clustering algorithms that operate on raw data or fixed features, DEC simultaneously refines both the feature space and the cluster assignments through an iterative optimization process. DEC uses a soft assignment approach based on the Student’s t-distribution, which assigns data points to clusters probabilistically rather than through hard assignments. This probabilistic approach allows the model to express uncertainty and subtle cluster memberships, improving robustness and flexibility. Through repeated iterations, DEC minimizes a clustering loss function that encourages the latent feature space to align more closely with the evolving cluster structure. This joint learning framework enables DEC to excel at clustering high-dimensional and complex datasets where conventional clustering methods often struggle.

3.4. Performance Metrics for Cluster Analysis

We used the Silhouette Score and Davies–Bouldin Index to evaluate the quality of clustering results. The Silhouette Score measures how similar a data point is to its own cluster (cohesion) compared to other clusters (separation). It lies in the range of −1 ≤ Silhouette Score ≤ 1, where negative values indicate possible misclassification, 0 suggested overlapping clusters, and values closer to 1 reflect well-separated and compact clusters. The Davies–Bouldin Index evaluates clustering quality by measuring the average similarity between each cluster. Its value is always zero or positive, where lower values represent tight, well-separated clusters, while higher values represent loose clustering.

4. Discussion and Results

4.1. Regional Distribution of Radioactive Elements

During the formation of metamorphic rocks, variations in environmental conditions, such as pressure, temperature, and fluid activity, can lead to changes in the concentrations of naturally occurring radioactive elements. These variations are closely linked to the composition of the original source rock that formed the migmatite, as well as the specific metamorphic processes and stages it underwent. As a result, the diversity and abundance of radioelements can vary considerably, providing valuable geochemical indicators of a region’s geothermal energy potential. Migmatites, in particular, typically form near the base of the crust through partial melting, a process that facilitates the mobilization and redistribution of minerals, including radioactive elements, within the rock.

In the study region, spatial variations in the concentrations of K (wt.%), Th (ppm), and U (ppm) indicate that areas with elevated K and Th levels in migmatitic rocks are closely associated with zones of known geothermal activity. K concentrations range from 0.002 to 8.164 wt.%, Th values from 1 to 65 ppm, and U concentrations from 0.50 to 15 ppm (Figure 3). These variations were observed in migmatite samples collected in the vicinity of geothermal wells drilled by the General Directorate of Mineral Research and Exploration in the Simav region. The observed spatial association between high radioelement concentrations in migmatitic gneisses and geothermal activity suggests a strong geothermal energy potential in the area. These findings support the use of radioelement geochemistry as a valuable tool in geothermal exploration, particularly in regions where migmatitic and other high-grade metamorphic rocks are present.

4.2. Regional Distribution of Magnetic Susceptibility

The expressions provided in Equation (4), and hence the resultant geothermal potential assessment, do not account for the influence of magnetic susceptibility, specifically, low- and high-frequency-dependent and mass-specific magnetic susceptibility, despite the relevance of these parameters in evaluating the region’s RHP potential. To address this, we analyzed magnetic susceptibility from laboratory measurements of rock samples collected across the study area. As shown in Figure 4, the green, red, and purple regions represent areas with high magnetic susceptibility, while yellow and blue regions indicate lower values. Notably, zones with elevated magnetic susceptibility correspond to mineral-rich areas and locations where geothermal energy is actively utilized. This is likely due to the dissolution of minerals in the presence of hot water and their subsequent interactions under elevated temperatures. Elevated magnetic susceptibility values observed in areas of hydrothermal alteration suggest that geothermal processes enhance mineral transformations. These findings contribute to a more comprehensive understanding of the geothermal potential of the study area.

4.3. Regional Radioactive Heat Production

Radioactive heat production (RHP), a key indicator of geothermal energy potential, was estimated using the first expression in Equation (4), as differences among the available formulations were determined to be negligible. The estimation was based on the measured spatial distribution of K, U, and Th concentrations. The resulting spatial distribution of RHP is presented in Figure 5, highlighting zones of elevated radiogenic heat generation. RHP values in the study area range from 0.291 to 6.784 µWk${\mathrm{g}}^{-1}$. RHP values exceeding 2.0 µWk${\mathrm{g}}^{-1}$ are generally considered indicative of significant heat producing rocks [17,19]. In this context, migmatites with RHP values exceeding this threshold are predominantly concentrated in the eastern part of the region, in close proximity to seismically active zones. This spatial association suggests that these migmatites represent the principal heat-producing lithology in the area that contributes to the geothermal energy potential of the region. The combination of high RHP and tectonic activity further underscores the geothermal energy potential of this part of the study area.

Moreover, RHP-based assessment of geothermal energy potential in Equation (4) does not account for the effect of the primary seismic wave (P-wave) velocity, $V_p$. However, the relationship between $V_p$ and RHP, as demonstrated by Rybach and Buntebarth [19], provides valuable insights for both geothermal system identification and the interpretation of seismotectonic processes. Low $V_p$ values are typically associated with increased fracture density, high porosity, and advanced hydrothermal alteration, features commonly found in high-RHP geothermal regions and seismically active fault zones. In geothermal environments, elevated heat flow and hydrothermal alteration reduce the elastic properties of rocks, resulting in lower $V_p$ values and seismic wave anomalies. These low $V_p$ zones also influence the mechanical behavior of fault systems, potentially facilitating earthquake occurrence. Therefore, the frequent seismic activity observed in geothermal areas with low $V_p$ values may reflect elevated geothermal potential and active tectonic deformation that facilitates the upward migration of hydrothermal fluids. Consequently, $V_p$ anomalies serve as valuable geophysical indicators for delineating brittle zones in the crust and assessing geothermal energy potential, as they are often associated with regions of high thermal output and increased seismicity.

4.4. Cluster Analysis-Based Identification of Favorable Geothermal Energy Zones

Statistical and unsupervised AI models allow for the integration of all measured chemical elements, beyond U, Th, and K, along with magnetic susceptibility and P-wave velocity data, and RHP estimates, enabling a more effective identification of areas with geothermal energy potential. We used the commonly applied k-means clustering method [53] as the baseline. However, it resulted in a Silhouette Score of −0.112 and a Davies–Bouldin Index of 6.129, indicating poor clustering quality and significant misclassification. Consequently, the k-means method was excluded from further consideration.

Clustering results for the analysis of potential geothermal sites from each statistical and AI-model are shown in Figure 6, and the corresponding clustering metrics are presented in

Table 1. Comparison of clustering models based on evaluation metrics.

	Model
Metric	DBSCAN	GMM	ACM	ABDC	SOM	DEC
Silhouette Score	0.717	0.433	0.445	0.340	0.177	0.134
Davies–Bouldin Index	0.392	0.732	0.684	1.333	1.680	2.447

. When earthquake epicenters, production drilling sites, and the Eynal hot spring formed by hot water redirected from nearby drilling are overlaid with the clustering results, overlapping zones of the clusters designate the most promising geothermal energy locations.

Among the statistical models evaluated, DBSCAN demonstrated the strongest predictive performance by identifying a narrow, well-defined wedge-shaped zone formed by overlapping clusters. This zone encompasses the majority of the production drilling sites as well as thermal springs, which are sustained by discharge from production wells.

Among the AI models, ABDC emerged as the most promising unsupervised AI model. It identified a relatively compact overlapping cluster zone, also enclosing the drilling sites near the seismically active portion of the study area. Notably, while DBSCAN excluded the Ahmetli region as a potential geothermal zone, ABDC included it within its prediction. Future exploratory drilling in the Ahmetli area could serve as a valuable validation step to assess and compare the predictive reliability of DBSCAN and ABDC.

Overall, the potential geothermal zones identified based on RHP values calculated using Equation (4), derived from U, TH, and K concentrations, are generally consistent with the results obtained from DBSCAN and ABDC models. However, the statistical and AI-based models delineated more localized and narrowly defined zones. These narrower zones result from the inclusion of additional variables, especially magnetic susceptibility measures and seismic velocities, that were not considered in the RHP formulations but were integrated into the clustering analyses.

5. Conclusions

Correlative relationships, given in Equation (4), based on the concentrations of radioactive elements, U, Th, and K, and radioactive heat production (RHP) were used to identify potential geothermal sites across the study area. This analysis revealed three distinct zones within the Simav Basin: western, central, and eastern regions. Migmatitic gneisses from the eastern region, where geothermal energy is actively exploited, exhibited RHP values exceeding 2.0 µWk${\mathrm{g}}^{-1}$, whereas the central and western regions showed values below this threshold. Thus, field and laboratory analyses indicate that the eastern region has a higher potential for geothermal energy production. Notably, this area lies in proximity to seismically active zones, where recent earthquake magnitudes have exceeded 4.

Statistical and unsupervised artificial intelligence (AI) models enabled the integration of magnetic susceptibility and seismic characteristics, and the concentrations of radioactive and other chemical elements. Among the statistical methods, the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) model effectively delineated a narrow, well-defined wedge-shaped zone, formed by overlapping clusters, that marks the most favorable subregion for further geothermal exploration and potential production. Similarly, the Autoencoder-Based Deep Clustering (ABDC) model emerged as the most promising AI model for narrowing the search area for geothermal resources. Notably, the subregions identified by both DBSCAN and ABDC encompass the existing geothermal production wells and thermal springs, which were not included as model features but served as independent validation of the model outcomes. Although these identified zones are more spatially constrained, they align well with geothermally favorable areas previously identified through field investigations and laboratory analyses.

Our results demonstrate that, unlike traditional correlative methods, statistical and unsupervised AI models can incorporate magnetic and seismic data to more effectively narrow down potential geothermal exploration zones, providing time and cost savings. In both modeling approaches, the zones formed by the overlap of all clusters consistently highlighted the most favorable subregions for geothermal energy production.The results obtained from the AI models closely matched those from the statistical models, indicating that both methodologies can complement one another and should be jointly considered to enhance robustness. Future drilling of exploration wells and expanded geochemical sampling across the study area will be essential for validating the clustering performance of each approach.