Abstract: The water-rock heat exchange behavior in rock fractures is considered a critical factor in geothermal engineering design during the research and utilization of hot dry rocks.Accurate prediction of outlet water temperature in fractures is essential to significantly reduce engineering costs and energy consumption.A multi-field triaxial experimental system was employed to perform convective heat transfer experiments on U50mm×100mm granite fracture samples under varying environmental temperatures and volumetric flow rates.A dataset of seepage-heat transfer experimental data was established,and the parameters of a Long Short-Term Memory-Multi-Layer Perceptron (LSTM-MLP) combined neural network were optimized using the Grey Wolf Optimization (GWO) algorithm.Temporal dependencies in the seepage-heat transfer process were captured by the LSTM network,while nonlinear features were extracted by the MLP,allowing the advantages of both networks to be combined.The GWO algorithm was applied to avoid local optima effectively and to ensure optimal parameter configuration for the model.Four input parameters,including environmental temperature,inlet temperature,volumetric flow rate,and fracture aperture,were selected for predicting the outlet water temperature.Model performance was evaluated using three common statistical metrics,and the temporal correlation of the seepage-heat transfer process was analyzed.The results indicate that the GWO-LSTM-MLP model achieved the highest prediction accuracy compared to machine learning models used for geothermal production prediction in the past five years (R²=0.989; RMSE=1.238; MAE=0.922).The application of GWO significantly improved the model’s performance,increasing R² by 5.3%,reducing RMSE by 54.37%,and decreasing MAE by 60.53%.The steady-state outlet temperature of the seepage process was accurately predicted,with a maximum absolute error of 0.8912℃ and a percentage error of 1.338%.