Modeling China’s deep temperature field with ML-based heat flow and layered heat production data Authors: Jifu He, Lei Wang, Weikai Yi, Qiuyuan Shi Published: 2026 ABSTRACT AbstractAccurate modeling of the deep crustal temperature field is critical for geothermal resource assessment and understanding lithospheric thermal evolution. Traditional approaches relying on interpolation-based terrestrial heat flow (THF) map and average radiogenic heat production often fail to capture the geological heterogeneity of the crust, particularly in data-sparse regions. In this study, we present a refined framework that integrates a machine learning–predicted THF map with vertically stratified radiogenic heat production data, applied within a one-dimensional steady-state heat conduction model. This approach allows us to compute deep temperature profiles across mainland China at multiple depths (3–10 km) with improved resolution and geological consistency. The results align well with borehole observations and reveal extensive high-temperature zones exceeding 150 °C in regions such as the Tibet–Sanjiang Orogen, North China Craton, and Yangtze Craton. Moreover, based on modeled temperatures at 3 km and 10 km, we derived a geothermal gradient map and identified eight favorable geothermal belts characterized by high vertical gradients and deep temperatures suitable for enhanced geothermal system (EGS) deployment. This study offers a scalable framework for geothermal prospecting in data-sparse regions and provides strategic insights for future deep energy exploration.Keywords: Deep temperature field, Terrestrial heat flow, Machine learning, Geothermal gradient, ChinaSubject terms: Environmental sciences, Solid Earth sciences Introduction The deep temperature field of the continental crust plays a fundamental role in controlling the distribution of geothermal resources, influencing lithospheric rheology, and driving geodynamic processes1. Accurate estimation of subsurface temperature is therefore essential for both theoretical geoscience research and practical applications such as geothermal energy exploration and deep geological engineering2,3. In the context of China, where the demand for clean and sustainable energy sources is growing rapidly, understanding the deep geothermal structure has become increasingly important. Traditionally, the deep temperature field has been inferred using one-dimensional steady-state heat conduction models, which rely on terrestrial heat flow (THF), thermal conductivity, and estimates of radiogenic heat production4,5. In earlier studies, THF values were predominantly obtained through spatial interpolation of limited measurement points. For instance, Ma et al4. and Jiang et al5. generated national-scale THF maps by interpolating 1,230 measured data points across China, which were subsequently used to calculate the deep thermal structure. Radiogenic heat production was commonly assigned according to planar zoning schemes, with each zone characterized by a single representative value. However, this methodology has notable drawbacks. First, THF observations are sparse and unevenly distributed across China (Fig. 1), necessitating interpolation techniques that may inadequately reflect geological complexity6. Second, the assignment of radiogenic heat production often assumes homogeneous crustal properties, overlooking both the vertical stratification of heat-producing layers and the lateral heterogeneity of crustal composition. These simplifications can lead to considerable uncertainties in modeling the subsurface thermal regime. Distribution of THF measurements across China (The map was generated using Surfer v27 (https://www.goldensoftware.com/). The base geographic boundary data were obtained from geoBoundaries (CC BY 4.0, https://www.geoboundaries.org/). THF data from Jiang et al5.). Recent advances in machine learning (ML) provide powerful tools to overcome long-standing limitations in THF estimation. ML models can capture complex, nonlinear relationships between THF and a wide array of geological and geophysical variables, enabling more accurate and spatially continuous predictions6–9. He et al6. developed a gradient boosting regression tree (GBRT) model trained on global THF measurements and 17 geological features, achieving strong predictive performance (R² = 0.87 on the test set). The resulting THF map of China showed markedly improved accuracy and resolution over traditional interpolation methods. This improvement is primarily evidenced by the model’s superior consistency with geological manifestations and tectonic structures. Specifically, the machine learning-based map correctly captures 84% of known high-temperature hot springs (> 75℃) within its high THF zones6, whereas traditional interpolation method10achieve a 73% capture rate. Furthermore, by integrating high-resolution geological features rather than just mathematically smoothing sparse data points, the model successfully resolves local thermal anomalies that interpolation methods inherently miss. At the same time, Wang et al11. systematically mapped crustal heat production across mainland China by integrating the global CRUST1.012 model with extensive geological data. The study stratified the crust into eight vertical layers across 19 tectonic zones, from Quaternary sediments to the upper mantle, and established the spatial extent and thickness of each. Critically, the authors compiled a robust dataset comprising 664 new measurements of radioactive heat production and ~ 1,000 records of crustal U, Th, and K concentrations. This enabled detailed layer-by-layer mapping of radiogenic heat production across the continent, offering a more geologically consistent basis for thermal modeling. In this study, we present a refined framework for estimating the deep temperature field of China by integrating a machine learning-derived THF map with stratified radiogenic heat production data, within a one-dimensional steady-state heat conduction model. This approach enables temperature calculations at multiple depths across mainland China, resulting in a high-resolution map of the subsurface thermal structure. The model reveals key features of lithospheric thermal architecture and provides essential inputs for geothermal resource assessment, deep drilling site selection, and strategic planning for geothermal energy development. More broadly, this work demonstrates the potential of machine learning to enhance geoscientific modeling in data-sparse and geologically complex regions, offering a scalable methodology applicable to other parts of the world. Methodology The novelty of this study lies in the use of the most recent THF data predicted by an ensemble model using machine learning6, combined with detailed data on heat production from crustal stratification, to reconstruct the deep temperature field of the Chinese continental lithosphere. The systematic workflow is illustrated in Fig. 2. Workflow of geothermal temperature prediction in China based on machine learning-predicted THF and layered heat production data. Conductive heat transfer dominates large scale thermal processes within the continental lithosphere13. At the scale of a tectonic province, vertical thermal variations significantly outweigh lateral heterogeneities13. Consequently, one dimensional steady state heat conduction models are extensively utilized to reconstruct deep geothermal fields4,5,14. Within this analytical framework, subsurface temperature distributions can be derived directly from borehole measurements and rock thermal properties4. The fundamental basis for this inference is the steady-state heat conduction equation, which describes the internal transfer of heat under equilibrium conditions. The general form of the steady-state heat conduction equation is given as4: where K is the thermal conductivity of the rock, W/m/K; A is the volumetric heat production rate, µW/m3. As indicated by Eq. (1), the temperature distribution T is determined by the heat production rate A and the thermal conductivity K. Under the assumption of one-dimensional steady-state heat flow, Eq. (1) simplifies to: where K is the THF, mW/m2;z is the depth, m A(z) is the heat production rate as a function of depth z, µW/m3༛K(T, z) is the thermal conductivity as a function of temperature T and depth z, W/m/K T0 is the is the temperature at the depth of the thermocline (constant-temperature zone) near the surface, ℃. In this study, T0 is approximated by the annual mean surface temperature, as the two are typically consistent in stable continental thermal regimes and high-resolution regional thermocline data are currently unavailable. In sedimentary layers at depths less than 3 km, the thermal conductivity and heat production rate can be approximated as constant within individual lithological units5, with values typically derived from measured averages for specific rock types. For simplified calculation under these conditions, the temperature at depth z can be estimated using: In contrast, for deeper crustal domains such as metamorphic and igneous rocks at depths of 3–10 km, radioactive heat-producing elements (U, Th, K) are typically enriched in the upper crust and decrease with depth. While this variation is historically approximated using an exponential decay model4, its exact mathematical formulation remains debated13,15. To minimize the substantial uncertainties associated with continuous decay assumptions, this study does not rely on the exponential model. Instead, we adopt an empirically derived, vertically stratified heat production dataset (detailed in Sect. 3.4) and implement a multi-layer recursive integration approach. The lithospheric column is discretized into six distinct stratigraphic layers. Starting from the surface boundary condition where the initial heat flow is provided by the machine learning-predicted THF map (q0 =THFML), the downward heat flow (qi) and temperature (Ti) at the bottom of each specific layer i are solved iteratively: where is the thickness of layer i; and are the layer-specific heat production and thermal conductivity, respectively; and and are the corresponding values passed from the bottom of the overlying layer. By explicitly calculating the heat flow and temperature continuity across discrete stratigraphic interfaces, this multi-layer formulation accurately captures the thermal impact of vertically heterogeneous rock properties, providing a more realistic deep geothermal reconstruction than traditional simplified models. Data According to Eq. (3), calculating the subsurface temperature distribution at different depths requires several key input parameters, including THF, rock thermal conductivity, radiogenic heat production rate, and surface temperature. A brief overview of these data sources is provided below. Terrestrial heat flow Unlike previous studies5,16,17, which relied on the THF map generated through simple interpolation, this work adopts a ML-based THF model developed by He et al18.(Fig. 3). The model reveals high THF values predominantly in the Tibet-Sanjiang Orogen, the southeastern coastal region, and the Central China Orogen, where values commonly exceed 70 mW/m² and reach above 80 mW/m² in much of the Tibet–Sanjiang area. In contrast, low THF values (< 50 mW/m²) are observed across the Tarim Craton and the Western Central Asian Orogen. Intermediate values (50–80 mW/m²) are typical of the Yangtze Craton, Eastern Central Asian Orogen, and North China Craton. Terrestrial heat flow map of China based on machine learning prediction6. WCAO: Western Central Asian Orogen, TC: Tarim Craton, CCO: Central China Orogen, TSO: Tibet-Sanjiang Orogen, YC: Yangtze Craton, WYO: Wuyi-Yunkai Orogen, NCC: North China Craton, ECAO: Eastern Central Asian Orogen. The map was generated using Surfer v27 (https://www.goldensoftware.com/). The base geographic boundary data were obtained from geoBoundaries (CC BY 4.0, https://www.geoboundaries.org/). Rock thermal conductivity As one of the key parameters influencing the subsurface temperature distribution, rock thermal conductivity is affected by factors such as mineral composition, water content, temperature, and pressure conditions. With increasing depth, thermal conductivity generally increases due to elevated pressure, but may decrease as a result of rising temperature16. Unconsolidated to weakly cemented sediments in the shallow subsurface typically exhibit relatively low thermal conductivity values, commonly ranging from 1.0 to 2.4 W/m/K 4, In contrast, Mesozoic sedimentary formations generally have higher thermal conductivity values, typically between 2.4 and 2.8 W/m/K, while Paleozoic and older sedimentary strata exhibit even higher values, often in the range of 2.8 to 3.4 W/m/K. Ma et al4. proposed depth-dependent adjustments for thermal conductivity values: 2.4–3.0 W/m/K at depths of 3–6 km, 3.0–3.4 W/m/K at 6–8 km, and 3.4–3.8 W/m/K at 8–10 km. In this study, thermal conductivity values were assigned based on the aforementioned classification: 1.5 W/m/K for Quaternary strata, 2.4 W/m/K for Cenozoic formations, 2.8 W/m/K for Mesozoic formations, 3.0 W/m/K for the crystalline basement, and 3.4 W/m/K for layers beneath the basement. Surface temperature Surface temperature data were obtained from the Global Solar Atlas, representing the annual mean air temperature measured at 2 m above ground level during the period from 1994 to 2022. Figure 4 presents the spatial distribution of the annual mean temperature across China, showing lower temperatures in the Tibetan Plateau and higher temperatures in the southeastern regions. Surface temperature map of China. The map was generated using Surfer v27 (https://www.goldensoftware.com/). The base geographic boundary data were obtained from geoBoundaries (CC BY 4.0, https://www.geoboundaries.org/). Radiogenic heat production data This study adopts the most recent findings from Wang et al11. which provide a detailed dataset of crustal heat production rates across the Chinese continental region. The dataset offers enhanced spatial resolution in two dimensions: geological-structural divisions (horizontal) and crustal layering (vertical). In contrast to previous studies that relied on averaged crustal heat production values for entire regions16or directly applied classical exponential decay models for radioactive heat-producing elements17, the present study utilizes a more refined approach based on 19 tectonic divisions (Fig. 5) and 8 vertical stratigraphic layers, including the Quaternary strata, Cenozoic strata, Mesozoic strata, crystalline basement, upper crust, middle crust, lower crust, and upper mantle. Map showing the distribution of 19 tectonic subdivisions in China. (The map was generated using the open-source vector graphics editor Inkscape (https://inkscape.org/). The tectonic boundary data and subdivision scheme were obtained from Wang et al11.). The depth boundaries of these stratigraphic layers were determined with reference to the CRUST 1.0 model, particularly its sedimentary thickness and total crustal thickness data. In this framework, the lower crust and upper mantle were not considered within the scope of this study. The Quaternary, Cenozoic, and Mesozoic strata correspond to the upper, middle, and lower sedimentary layers in the CRUST 1.0 model, respectively. The upper crust and middle crust align with the corresponding layers in CRUST 1.0, while the crystalline basement is defined as the upper 5 km of the upper crust underlying the lower sedimentary layer11. Table 1 summarizes the heat production rates assigned to each tectonic unit and stratigraphic layer in this study. Notably, the heat production values of the sedimentary layers exhibit significant variability without a clear trend, likely reflecting the diverse compositions, sources, and depositional environments of the sediments, which influence the abundance of radioactive heat-producing elements. In contrast, a gradual decrease in heat production rate is observed from the crystalline basement downward through the deeper crust, which is consistent with the commonly accepted theory of upward enrichment of radiogenic elements in the continental crust. Table 1Heat generation rate data for different tectonic subdivisions and crustal layers used in this study.(Modified from Wang et al11.).Zone No.Heat Generation Rate, µW/m3Quaternary strataCenozoic strataMesozoic strataBasement layerUpper crustMiddle crust10.751.142.971.571.200.6020.751.141.642.891.500.7030.751.142.121.631.400.6040.721.592.802.721.520.9751.331.491.501.721.010.3861.431.821.602.051.260.6071.021.431.391.600.910.5580.862.382.141.961.100.8391.022.382.141.961.260.6010---5.181.880.6011---3.011.880.6012---2.851.670.6013---3.011.670.60140.901.232.392.391.570.43150.902.202.122.331.570.72160.903.803.964.030.940.38172.212.211.620.720.940.38181.281.181.161.060.950.38191.281.180.851.310.950.38 Heat generation rate data for different tectonic subdivisions and crustal layers used in this study. (Modified from Wang et al11.). 1 – Xing’an Block, 2 – Khanka Block, 3 – Songliao Basin Block, 4 – Erqis–Xilamulun Suture Zone, 5 – Peripheral Ordos Tectonic Belt, 6 – Upper Yangtze Craton, 7 – Eastern North China Craton, 8 – Eastern Segment of the Qinling–Qilian–Kunlun Tectonic Belt, 9 – Lower Yangtze Craton, 10 – Nanling Block, 11 – Southeast Coastal Magmatic Arc, 12 – Hainan Arc–Basin System, 13 – Taiwan Arc–Basin System, 14 – Southern Tibet Block, 15 – Northern Qiangtang–Sanjiang Orogenic Belt, 16 – Central Segment of the Qinling–Qilian–Kunlun Tectonic Belt, 17 – Western Segment of the Qinling–Qilian–Kunlun Tectonic Belt, 18 – Tarim Craton, 19 – Tianshan–Junggar–Beishan Orogenic System. Results 3 km depth geothermal temperature distribution Figure 6 illustrates the geothermal temperature distribution at a depth of 3 km across the Chinese continental region. Overall, most areas exhibit temperatures ranging between 50 °C and 100 °C at this depth. Specifically, regions with temperatures below 75 °C are primarily located in the Central Asian Orogen, the eastern Tarim Craton, and the northern part of the Tibet-Sanjiang Orogen. Predicted geothermal temperature distribution at a depth of 3 km in China. WCAO: Western Central Asian Orogen, TC: Tarim Craton, CCO: Central China Orogen, TSO: Tibet-Sanjiang Orogen, YC: Yangtze Craton, WYO: Wuyi-Yunkai Orogen, NCC: North China Craton, ECAO: Eastern Central Asian Orogen. The map was generated using Surfer v27 (https://www.goldensoftware.com/). The base geographic boundary data were obtained from geoBoundaries (CC BY 4.0, https://www.geoboundaries.org/). In contrast, most of the North China Craton exhibits geothermal temperatures between 75 °C and 100 °C. However, in certain localized areas such as the Shijiazhuang-Beijing-Huhehaote corridor, temperatures fall below 75 °C, whereas along the Tianjin–Zhengzhou axis, temperatures can exceed 100 °C. Similarly, the southern part of the Tibet–Sanjiang tectonic belt, the Wuyi–Yunkai Orogenic Belt, the Yangtze Craton, and the western Tarim Craton generally show geothermal temperatures within the 75 °C to 100 °C range. Notably, localized high-temperature anomalies exceeding 100 °C are observed near Lasa and to the west of Wuhan. These anomalies correspond to regions of elevated predicted THF. 5 km depth geothermal temperature distribution Figure 7 presents the estimated terrestrial geothermal temperature distribution map of China at 5 km depth derived from this study. Compared with the thermal profile at a depth of 3 km, the temperatures at 5 km demonstrate significant elevation, predominantly ranging between 75 °C and 125 °C, while the thermal distribution pattern retains certain characteristics observed at the shallower level. Predicted geothermal temperature distribution at a depth of 5 km in China. WCAO: Western Central Asian Orogen, TC: Tarim Craton, CCO: Central China Orogen, TSO: Tibet-Sanjiang Orogen, YC: Yangtze Craton, WYO: Wuyi-Yunkai Orogen, NCC: North China Craton, ECAO: Eastern Central Asian Orogen. (The map was generated using Surfer v27 (https://www.goldensoftware.com/). The base geographic boundary data were obtained from geoBoundaries (CC BY 4.0, https://www.geoboundaries.org/). Active faults data derived from28. Hot dry rock (HDR) is generally defined as deep seated crystalline rock masses located at depths between 3 and 10 km, exhibiting temperatures exceeding 150 °C while lacking sufficient natural fluids and permeability19,20–21. The exploitation of these resources relies fundamentally on the engineering of Enhanced Geothermal Systems (EGS)22, a process where drilling costs constitute the primary economic constraint2023,. The geothermal field at a depth of 5 km is a common target for the exploration and development of HDR because it represents a strategic balance between thermal grade and economic viability1920,. While shallower intervals of 3 km often yield temperatures insufficient for commercial power generation20, the drilling expenses associated with depths of 7 km or greater increase exponentially within hard crystalline basement rocks. Current engineering limits make deep reservoir stimulation beyond 5 km economically prohibitive for widespread development1920,. Therefore, the 5 km depth serves as an optimal technical and economic window where temperatures can meet the 150 °C requirement for power generation while remaining within the limits of current engineering capabilities1920,. By applying these quantitative criteria to our modeled results, spatial analysis of Fig. 7 reveals that exceptional areas where subsurface temperatures exceed 150 °C account for approximately 7.7% of the total continental area. These localized thermal anomalies are primarily concentrated in four major regions: the southern Tibet Sanjiang Orogen, the Chengdu Chongqing Guiyang Kunming corridor within the Yangtze Craton, a linear zone extending from Tianjin through Shijiazhuang, Jinan, Zhengzhou, to Hefei within the North China Craton, and the eastern segment of the Central Asian Orogen. These areas exhibit considerable potential for HDR exploration and represent key targets for future geothermal resource assessment and strategic development. Crucially, spatial verification confirms that these identified belts are not mathematical artifacts driven by the local density of input data. This is supported by two factors. First, the input THF data exhibits strong consistency with large scale geological structures6, constraining predictions to tectonic frameworks rather than data clusters. Second, the modeled high temperature anomalies spatially correspond with major fault networks derived from the continuous fault dataset of China (Fig. 7). Areas with high fault density consistently exhibit elevated subsurface temperatures. Specifically, dense fault systems in the eastern North China Craton, active tectonic boundaries in southern Tibet, and fault networks along the Southeast Coast spatially coincide with the identified thermal belts (Fig. 7). This correlation with established fault distributions confirms geological coherence and rules out spatial bias. This robust geological coherence is further supported by actual field data. Recent HDR exploration efforts in China have achieved limited but notable results24. Representative sites include the Gonghe–Qiabuqia and Matouying reservoirs25. At the Qiabuqia GR1 borehole, temperatures exceeding 200 °C were recorded at a depth of 3,705 m, while the Matouying site reached 150 °C at 4,000 m. In addition, the Wendeng ZKCW01 borehole in Shandong Province exhibits promising thermal conditions, with projected temperatures of ~ 170 °C at 5 km depth26. Similarly, the Lijin SDGRY1 borehole is estimated to exceed 180 °C at 5120 m27. These measured and inferred HDR occurrences show strong agreement with our modeled thermal distribution (Fig. 7), where estimated temperatures consistently exceed 150 °C. This spatial correspondence provides empirical validation of the accuracy and reliability of the predicted geothermal field. 7 km depth geothermal temperature distribution Figure 8 delineates the geothermal temperature distribution at a depth of 7 km across China’s continental territory. Compared to the 5 km profile, the map reveals a significant thermal increase, with most regions exhibiting temperatures above 150 °C. Several geological domains—including the central and eastern North China Craton, the southeastern Central Asian Orogen, the Chengdu–Chongqing–Guiyang–Kunming corridor within the Yangtze Craton, the central Central Orogenic Belt, and the southern Tibet–Sanjiang Orogen—display pronounced thermal anomalies, with temperatures locally exceeding 175 °C and in some cases approaching 200 °C. Predicted geothermal temperature distribution at a depth of 7 km in China. WCAO: Western Central Asian Orogen, TC: Tarim Craton, CCO: Central China Orogen, TSO: Tibet-Sanjiang Orogen, YC: Yangtze Craton, WYO: Wuyi-Yunkai Orogen, NCC: North China Craton, ECAO: Eastern Central Asian Orogen. The map was generated using Surfer v27 (https://www.goldensoftware.com/). The base geographic boundary data were obtained from geoBoundaries (CC BY 4.0, https://www.geoboundaries.org/). These thermal patterns indicate that, with continued advancements in deep drilling technologies and decreasing operational costs, extensive high-temperature geothermal resources within the 5–7 km depth range could become increasingly accessible. The consistent temperature–depth trend observed in our model supports the technical viability of deep geothermal exploitation, highlighting these high-enthalpy zones as promising candidates for future EGS development. 10 km depth geothermal temperature distribution Figure 9 illustrates the geothermal temperature distribution at a depth of 10 km across China’s continental crust. At this depth, temperatures in most regions exceed 200 °C, with only limited zones registering intermediate values between 150 and 200 °C. Areas below 150 °C account for less than 5% of the mapped domain. This pronounced thermal stratification provides compelling evidence of China’s vast potential for deep high-enthalpy geothermal resources29. Predicted geothermal temperature distribution at a depth of 10 km in China. WCAO: Western Central Asian Orogen, TC: Tarim Craton, CCO: Central China Orogen, TSO: Tibet-Sanjiang Orogen, YC: Yangtze Craton, WYO: Wuyi-Yunkai Orogen, NCC: North China Craton, ECAO: Eastern Central Asian Orogen. The map was generated using Surfer v27 (https://www.goldensoftware.com/). The base geographic boundary data were obtained from geoBoundaries (CC BY 4.0, https://www.geoboundaries.org/). Comparative spatial analysis and model validation of deep geothermal distributions To validate the newly predicted geothermal distribution maps, we compiled continuous temperature-depth profiles from eight representative deep boreholes (Fig. 10), whose locations are illustrated in Fig. 11. This dataset includes the SK-2 well in the Songliao Basin30, three HDR exploration wells (M-1 in the Bohai Bay Basin31, GR1 in the Gonghe Basin32, and FSR1 in the Beibuwan Basin33), the Luntan1 well in the Tarim Basin34, and the DS-1, L4, and CY84 wells in the Sichuan Basin35,36–37. Comparison between predicted and measured geothermal temperature profiles (Data from30–37). Temperature field at 3 km depth based on traditional THF interpolation (The map was generated using the open-source vector graphics editor Inkscape (https://inkscape.org/), representing the general thermal trends based on data from Ma et al.4). As shown in Fig. 10, predicted vertical temperature trajectories were plotted against measured logs to evaluate depth-dependent model reliability. Overall, predictions show good agreement with measurements in conduction-dominated, stable tectonic settings38. Specifically, modeled profiles for the SK-2, Luntan1, FSR1, and Sichuan Basin wells match well with observed deep thermal gradients. While the model accurately reconstructs the regional conductive baselines across these diverse basins, pronounced positive temperature deviations are specifically observed in the deep sections of the GR1 and M-1 HDR exploration wells. The observed positive anomalies in the GR1 and M-1 wells—where measured deep temperatures significantly exceed the modeled predictions—likely reflect local geological conditions rather than modeling errors. While our hybrid framework reconstructs the regional steady-state conductive regime, the GR1 well is located in the Gonghe Basin, characterized by highly radiogenic granite intrusions39, and the M-1 well sits in the active extensional Bohai Bay Basin40. In both localized environments, upward heat transfer is driven by hydrothermal fluid advection along deep fracture networks. These dynamic factors locally violate the pure steady-state conduction assumption, causing the observed deep thermal anomalies. Quantitatively, the global Root Mean Square Error (RMSE) for all eight boreholes at 3 km depth is 14.2 °C. However, excluding these two advection-dominated outliers refines the RMSE to just 6.0 °C. This low refined error further validates the high precision of our hybrid model in capturing regional conductive baselines. Furthermore, a comparative spatial analysis was conducted against previous geothermal maps4 derived from interpolated THF data and exponential heat production models. Figure 11 shows the 3 km geothermal temperature distribution across the Chinese continental region as estimated by Ma et al.4. While the overall temperature range in this study is broadly consistent with theirs—i.e., temperatures below 100 °C in most regions—there are substantial differences in the spatial distribution. Through comparison with actual temperature measurements, it was found that the temperature estimates from this study more closely reflect observed values in many areas. First, in the western Tarim Craton, previous studies45, suggested lower values, mostly between 50 and 75 °C, with some areas even below 50 °C (Fig. 11). In contrast, our model predicts geothermal temperatures primarily in the range of 75–100 °C (Fig. 6). This higher estimation is robustly supported by actual borehole measurements. Notably, Li et al.34 reported the measured temperature profile of the Luntan1 well (Fig. 10), which records a temperature of approximately 80 °C at a depth of 3 km, exhibiting excellent agreement with our predicted temperature-depth curve. Furthermore, temperature data compiled by Liu et al41. also indicate that many boreholes in the Tarim Basin reach temperatures above 80 °C at 3 km depth. Their analysis shows that temperatures at this depth typically range between 63 and 100 °C, with an average of 82 °C. Similarly, the study by Huang et al42. on the central uplift zone of the Tarim Basin reported that most boreholes exceed 80 °C at 3 km, with some even exceeding 90 °C. Clearly, the estimates in this study are more consistent with these measured values. Second, along the Chongqing–Guiyang axis in the Yangtze Craton, previous interpolations45, estimated uniform temperatures primarily within the 75–100 °C range, with some areas even dropping below 75 °C (Fig. 11). However, our model successfully captures localized geothermal anomalies exceeding 100 °C (Fig. 6). This prediction is firmly corroborated by published borehole data from this region374344,,, which confirm the presence of high-temperature wells. Notably, the measured temperature profile of the DS-1 well37 (Fig. 10) provides direct evidence supporting these localized thermal anomalies. Additionally, in the region north of Chongqing, Ma et al4. predicted an anomalously cold background of 50–75 °C (Fig. 11), whereas our model estimates a warmer baseline of 75–100 °C (Fig. 6). This upward correction is strongly supported by the measured temperature-depth logs from the L435 and CY8436 wells (Fig. 10), which exhibit excellent agreement with our predicted vertical profiles. Furthermore, regional assessments by Liu et al44. based on field measurements indicate that temperatures in this broad area range from 60 °C to 120 °C. Their analysis highlights that—apart from Yunnan—the principal high-temperature zones are predominantly concentrated along the Chengdu–Chongqing–Guiyang corridor. This field-based pattern aligns perfectly with the spatial distribution predicted by our hybrid approach (Fig. 6). Thirdly, in the western part of the Tibet–Sanjiang tectonic belt near the Indian border, our model estimates baseline temperatures of 75–100 °C at 3 km depth, featuring localized anomalies exceeding 100 °C (Fig. 6). In contrast, earlier studies based on interpolated THF maps suggest lower temperatures of 50–75 °C for this region (Fig. 11). However, investigations by Wu et al45. on the Langjiu geothermal field (around 80.35°E, 32.35°N) report hydrothermal fluid temperatures ranging from 78 °C to 105 °C. These active geothermal manifestations strongly indicate that our elevated temperature estimates are much more realistic and successfully capture the true geothermal potential of this boundary zone. Finally, in the southeastern part of the Eastern Central Asian Orogenic Belt, west of Harbin–Changchun, our model predicts a northeast-trending band with temperatures exceeding 100 °C. Previous studies estimated temperatures in this area mostly between 75 and 100 °C, with some areas below 75 °C. Nevertheless, an extensive analysis of 1,021 oil-test temperature data points by Zhang et al46. revealed that the average temperature at 3 km depth in this region reaches up to 119.5 °C. This extreme thermal state is directly validated by the continuous measured profile of the SK-2 well30 (Fig. 10), which records approximately 110 °C at 3 km depth and exhibits remarkable agreement with our predicted vertical trajectory. The above comparisons with measured data provide validation for the reliability and superiority of the deep geothermal temperature map produced in this study. Moreover, they indirectly support the accuracy of the machine learning–based heat flow map used in this research. Given that the temperature estimations at depths of 5 km, 7 km, and 10 km were derived using the same datasets and methodology as for the 3 km results—and considering the scarcity of deep borehole measurements—no further detailed comparisons are conducted for those depths in this study. Geothermal gradient mapping and implications for EGS targeting Using the modeled temperature fields at 3 km and 10 km depth, we calculated the geothermal gradient across mainland China (Fig. 12). The gradient ranges from 0.8 to 5.4 °C/100 m, with a spatial pattern broadly consistent with that of the temperature distribution. The highest values are observed in the Tibet–Sanjiang Orogen, where extensive areas exhibit gradients exceeding 2.6 °C/100 m. Elevated gradients are also evident in the Central China Orogen, particularly in its central and western regions, with values ranging from 2.0 to 3.3 °C/100 m. In contrast, low-gradient zones are concentrated in the Western Central Asian Orogen and the Tarim Craton, where gradients mostly fall between 1.0 and 1.6 °C/100 m. Elsewhere, geothermal gradients predominantly lie within the range of 1.5 to 2.5 °C/100 m. Geothermal gradient distribution between 3 km and 10 km depth across mainland China. WCAO: Western Central Asian Orogen, TC: Tarim Craton, CCO: Central China Orogen, TSO: Tibet-Sanjiang Orogen, YC: Yangtze Craton, WYO: Wuyi-Yunkai Orogen, NCC: North China Craton, ECAO: Eastern Central Asian Orogen. The map was generated using Surfer v27 (https://www.goldensoftware.com/). The base geographic boundary data were obtained from geoBoundaries (CC BY 4.0, https://www.geoboundaries.org/). High geothermal gradients typically indicate the potential for elevated subsurface temperatures at relatively shallow depths, enhancing the accessibility of deep geothermal resources. Based on the temperature distribution at a depth of 5 km, eight high-potential geothermal belts were identified (A-A’~H-H’ in Fig. 12). Notably, these identified belts consistently align with regions of elevated geothermal gradients, demonstrating strong spatial consistency between the depth specific temperature field and the regional thermal framework presented in Fig. 12. Figure 13 presents the corresponding temperature profiles from 3 km to 10 km depth along these belts. In all cases, temperatures near 5 km depth reach or exceed 150 °C (Fig. 13), underscoring the suitability of these zones for future high-temperature geothermal exploration and development. Geothermal temperature profiles (3–10 km) along eight favorable geothermal belts in China. Discussion The reconstruction of deep thermal fields is fundamentally constrained by the accuracy of THF, which serves as the primary boundary condition for the continental lithosphere. Historically, thermal baseline models have relied on interpolated maps of THF, which often struggle to account for regional tectonic complexity or significant gaps in observation within vast and heterogeneous regions. In this study, we bridge this gap by adopting a boundary condition predicted by an ensemble model using machine learning, which is then integrated with a stratified model of radiogenic heat production. By coupling these diverse datasets, we move beyond the limitations of traditional interpolation and planar zoning to provide a more representative background conductive geotherm. However, replacing traditional interpolation with a framework driven by data necessitates a critical evaluation of the resulting spatial confidence and the applicability of the underlying physical assumptions across the continent. To quantify the reliability of these thermal predictions, we mapped the spatial uncertainty propagated directly from the input THF. To calculate the propagated standard deviation (SD), we applied an analysis of sensitivity perturbation based on the approximately linear relationship between THF and temperatures at depth within the governing equation for steady state heat conduction. Specifically, the physical model was executed twice for each spatial grid. The initial run utilized the original predicted THF to establish a baseline temperature at depth. The subsequent run utilized a thermal input perturbed by adding the local SD of the predicted THF6. The absolute difference between the temperatures derived from these two runs provides a quantitative approximation of the propagated uncertainty. We selected a depth of 5 km for this spatial assessment because it represents a geothermometric benchmark for the exploration of deep energy resources, particularly for EGS. The spatial distribution of the SD for temperature at this 5 km depth (Fig. 14) visualizes the variations in prediction confidence, which can be delineated into three categories. Low uncertainty zones (SD < 10 °C) cover approximately 46% of mainland China, primarily including the NCC, the YC, and the ECAO. These areas benefit from denser measurements of THF and relatively stable geological conditions, making them favorable targets for near term geothermal exploration. Medium uncertainty zones (SD between 10 °C and 20 °C) account for approximately 38% of the study area, encompassing the CCO, the WCAO, and the western TSO. High uncertainty zones (SD > 20 °C) are concentrated in the TSO, CCO, and the southern WCAO, with a maximum SD exceeding 40 °C in the central TSO. These larger uncertainties are primarily attributed to the sparse distribution of measurements for THF in these remote regions. Spatial distribution of the standard deviation for the predicted temperature at a depth of 5 km. The map was generated using Surfer v27 (https://www.goldensoftware.com/). The base geographic boundary data were obtained from geoBoundaries (CC BY 4.0, https://www.geoboundaries.org/). While this framework provides a conductive background, we acknowledge its physical limitations in tectonically active domains. The steady state assumption relies on thermal equilibrium. In regions like the Tibetan Plateau, this equilibrium can be disturbed by transient thermal perturbations caused by recent magmatism and rapid crustal exhumation, which affect thermal relaxation. Furthermore, by focusing on conductive heat transfer, our framework does not account for advective heat transport driven by deep hydrothermal circulation along major fault zones. Neglecting advection generally leads to localized underestimations of deep temperatures in structural zones of fluid upwelling, and corresponding overestimations in regions of fluid downwelling. However, constructing a fully coupled model with transient advection at a continental scale requires dense regional arrays of deep permeability, which are currently unavailable for China. Therefore, our results represent a macroscopic background conductive geotherm, serving as a regional baseline upon which future localized hydrogeological models can be constructed. In addition to the limitations of conduction, several structural constraints remain. First, the application of heat transfer in one dimension does not fully account for thermal anisotropy and lateral heat flow in three dimensions. Second, temperature estimations extending beyond a depth of 10 km rely on theoretical extrapolation and currently lack direct physical validation from deep boreholes. Third, the spatial coverage of radiogenic heat production data remains uneven across certain tectonic units, introducing potential modeling biases. Future research could evolve toward coupled thermal, hydrological, and mechanical models in three dimensions that incorporate deep crustal fluid flow. Furthermore, enhancing deep structural constraints using models of crustal thickness with high resolution (such as CRUST1.0) and expanding targeted measurement campaigns in under sampled western regions will be beneficial for refining the thermal architecture of the continental lithosphere. Conclusions This research establishes a refined analytical framework for reconstructing the deep thermal architecture of the Chinese continental lithosphere by integrating terrestrial heat flow predicted by machine learning with a stratified model of radiogenic heat production. By mathematically coupling the spatial structure of the crust with its thermal properties, this approach overcomes the inherent limitations of traditional interpolation and planar zoning, achieving superior geological consistency and closer alignment with observed borehole data across the 3 to 10 km depth range. The integration of data driven boundary conditions with a deterministic physical framework provides a scalable baseline for geothermal assessment in tectonically complex and data limited regions. The modeled temperature distribution reveals a clear distinction between the macroscopic thermal background and localized anomalies across the continent. At a depth of 5 km, the thermal regime is predominantly characterized by a moderate background ranging between 75 °C and 125 °C. However, exceptional regional anomalies exist where temperatures exceed 150 °C, accounting for approximately 7.7% of the total continental area. These extreme thermal environments are structurally controlled and localized within specific tectonic domains, most notably the Tibet Sanjiang Orogen and the Central China Orogen, where geothermal gradients reach as high as 5.4 °C/100 m. Based on these findings, eight favorable geothermal belts were identified as high priority targets for the development of Enhanced Geothermal Systems. As depth increases to 10 km, the thermal field continues to intensify, with modeled temperatures overwhelmingly surpassing 200 °C, further underscoring the vast potential of deep seated heat resources. Despite these advancements, the predictive accuracy of the framework is bounded by the physical assumption of steady state heat conduction. This simplified treatment does not fully account for transient thermal perturbations or advective heat transport driven by deep hydrothermal circulation, which may influence temperatures in tectonically active fault zones. Furthermore, estimations at depths exceeding 10 km require additional validation from ultra deep scientific boreholes. Future research must therefore evolve toward fully coupled thermal, hydrological, and mechanical models in three dimensions. Enhancing deep structural constraints using models of crustal thickness with high resolution will be essential for refining the thermal architecture of the lithosphere and reducing spatial uncertainties in future geothermal exploration. Author contributions Jifu He conceptualized the study, developed the methodology and model, conducted the investigation, wrote the original draft, and participated in reviewing and editing. Qiuyuan Shi performed formal analysis and visualization, and was involved in reviewing and editing the manuscript. Lei Wang perfected the methodology, created the schematic workflow (Figure 2), compiled the deep borehole data, completed the spatial comparison (Figure 10), and acquired funding. Weikai Yi performed the rigorous error propagation analysis, comprehensively optimized the Discussion section, and created the spatial uncertainty map (Figure 14). All authors read and approved the final manuscript. Funding Declaration. This work was supported by the Sinopec Key Science and Technology Project “Research on Drilling and Reservoir Stimulation Technology in the Fushenre-1 Well Area” (Grant No. P25199). Additional funding was provided by the Major Project of the National Natural Science Foundation of China and the Sinopec Key Support Project “Dynamic Damage Mechanism and Efficient Fragmentation Methods of High-Temperature Rocks” (Grant No. 52192624). Data availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.