Jifu He

Geothermal heat flow (GHF) is essential for quantifying lithospheric thermal conditions and supporting geothermal resource assessment. However, direct GHF measurements remain sparse and unevenly distributed due to drilling limitations and high acquisition costs. While machine learning offers a scalable alternative for GHF estimation, purely data-driven models often struggle to capture complex geological heterogeneity, and their performance tends to plateau as additional geophysical features are introduced. This study introduces a physics-guided clustered Gradient Boosting Regression Trees (GBRT) ensemble framework. Nineteen geological and geophysical features were incorporated, including two newly developed analytical GHF features derived from thermal equilibrium and conduction principles as physics-derived features. The clustered ensemble method significantly enhances prediction accuracy compared with a baseline GBRT model. Ensemble performance plateaus at eight model units, yielding an R ² of 92.01% and reductions in MAE and RMSE of 22.72% and 22.37%, respectively. The analytical features show the highest importance scores, demonstrating that embedding physics-guided constraints improves interpretability and reliability. Finally, an updated GHF map of China is generated, offering improved alignment with geothermal manifestations such as high-temperature springs. The proposed framework advances the use of ensemble and physics-guided learning for robust GHF estimation in data-scarce regions.

Accurate 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.

Reservoir temperature is a critical factor influencing the formation and production of coal, coalbed methane (CBM), and geothermal heat flow (GHF), with GHF serving as a key indicator of subsurface thermal conditions. In this study, an extreme gradient boosting (XGB) model is developed to predict GHF in China using 12 geological and geophysical features. The model is trained on both global and China-specific data sets and evaluated using MAE, RMSE, and R ² metrics. Feature importance is assessed via weight, gain, coverage, and SHAP values. Results show that the China-based model performs better in low-heat-flow areas (R ² = 0.73), while the global model is more accurate in high-heat-flow regions (R ² = 0.65). Based on predictions, heat flow maps were generated for China and Shanxi Province and compared to kriging-interpolated maps. The model-generated maps exhibit smoother spatial trends and stronger correlations with measured data, demonstrating improved consistency and geological coherence. Statistical analysis reveals that GHF in medium- to high-rank coal regions, favorable for CBM generation, is significantly higher than in low-rank regions. In Shanxi, coal-bearing areas also show elevated GHF compared to noncoal regions, reinforcing the link between thermal conditions and CBM potential. Feature importance analysis highlights crustal structure and volcanic activity (e.g., distance to volcanoes, distance to mid-ocean ridges, and Moho depth) as dominant GHF controls. The XGB model demonstrates robust predictive capability, offering valuable guidance for geothermal resource exploration and CBM development, particularly in data-scarce regions.

Compared to natural and shale gas, studies on predicting production specific to coalbed methane (CBM) are still relatively limited, and mainly use decline curve methods such as Arps, Stretched Exponential Decline Model, and Duong’s model. In recent years, machine learning (ML) methods applied to CBM production prediction have focused on the significant data characteristics of production, achieving more accurate predictions. However, throughout the application process, these models require a large amount of data for training and can only achieve accurate forecasts over a short period, such as 30 days. This study constructs a hybrid ML model by integrating a long short-term memory (LSTM) network and Transformer architecture. The model is trained using the mean absolute error (MAE) loss function, optimized using the Adam optimizer, and finally evaluated using metrics such as MAE, root mean square error (RMSE), and R squared (R²) scores. The results show that the LSTM-Attention (LSTM-A) hybrid model based on small training datasets can accurately capture the CBM production trend and is superior to traditional methods and the LSTM model regarding prediction accuracy and effective prediction time interval. The methodologies established and the results obtained in this study are of great significance to accurately predict CBM production. It is also helpful to better understand the mechanisms of CBM production.

Geothermal heat flux (GHF) serves as a key parameter in geothermal exploration but is sparsely measured worldwide because of technological challenges and high costs. Traditional interpolation and empirical methods for GHF prediction often provide limited accuracy. Unlike traditional methods, machine-learning methods can integrate multiple geologic factors affecting GHF and capture their nonlinear relationships, leading to more accurate predictions. Here, we develop a gradient boosting regression tree (GBRT) model to predict GHF across China. We screen and choose a comprehensive set of 16 geologic and geophysical features that affect GHF as the input variables of the GBRT model. A global data set comprising 36,325 GHF data points is divided into training and test sets. The training set is used to develop the model, and the test set is used to evaluate its performance. The GBRT model established using the global data set exhibits great performance in predicting GHF in China, achieving an R-squared value of 87% on the test set, which is among the highest reported for GHF prediction. The predictions are validated with geologic features and high-temperature hot spring data, confirming their accuracy. The newly generated GHF map of China reveals previously unidentified high GHF regions (“sweet spots”), providing valuable targets for future geothermal exploration in China.