Machine learning model for predicting geothermal in Xianning, China

• A small sample data set of 12 geological features was collected, and the model adaptation of multi-source heterogeneous data was completed. • The prediction accuracy of four machine learning models was compared, and a regression model suitable for geothermal resource distribution prediction was selected. • The models were stacked and combined, and a model with random forest and multi-layer perceptron superposition was used to further improve the prediction accuracy of a single model. • The distribution and classification of potential geothermal resources in Xianning City were predicted, and a geothermal resource prediction map was drawn. Geothermal energy is a renewable and efficient green energy source crucial for sustainable development. This study aims to predict geothermal resource potential in the Xianning area of Hubei Province, China, where 13 geothermal fields are irregularly distributed. Twelve geological features—including heat flow, geothermal gradient, aeromagnetic data, Curie depth, and isotopic recharge elevation—were integrated with 176 known geothermal data points (80% for training, 20% for testing) and 38,872 prediction points. Four machine learning models—Linear Regression (LR), Deep Neural Network (DNN), Multilayer Perceptron (MLP), and Random Forest (RF)—were compared. Results show that RF achieved the best performance (R 2 ≈ 0.997, error ≈2.1%), while LR performed the poorest (error ≈21%). A stacking ensemble model combining RF and MLP further improved prediction accuracy (R 2 ≈ 0.998, error ≈1.4%). Geothermal potential maps generated by the models consistently identified northwestern Xianning as the most promising zone, followed by the southeast, with central and eastern areas showing lower potential. The study demonstrates that machine learning, particularly ensemble methods, can effectively predict geothermal resources using limited and heterogeneous data, providing a reliable tool for regional geothermal exploration.