A Bayesian evidential learning framework for safety and performance prediction in thermo-hydro-mechanical coupled deep mine geothermal systems

Deep-mine geothermal co-mining combines heat extraction with mine cooling, but intrinsic subsurface variability and thermo-hydro-mechanical (THM) coupling introduce complex uncertainties in responses such as temperature evolution and drift stability. We present a Bayesian Evidential Learning (BEL) framework that employs early simulated temperature data to forecast five-year-ahead outcomes within the same physical field (temperature to temperature) and across fields (temperature to stability), eliminating the need for costly parameter inversion. First, we sample key engineering and material parameters using Latin Hypercube Sampling (LHS) and run forward THM simulations on two-dimensional models to create a synthetic dataset. Next, Principal Component Analysis (PCA) reduces the dimensionality of inputs and outputs for efficient processing. When strong linear relationships exist, Canonical Correlation Analysis (CCA) provides Gaussian posterior estimates; when nonlinear or cross-field dependencies dominate, Mixture Density Networks (MDN) capture the more complex non-linear behavior. A PCA-based prior filtering step further reduces computation costs and prevents overestimation of uncertainty. In validation tests covering temperature-to-stability, drift cooling, and production temperature forecasts, observed test values reliably fall within the predicted posterior distributions’ 95% credible intervals, and the method achieves low continuous ranked probability scores. These findings demonstrate the framework's potential as an uncertainty forecasting tool to support decision-making in deep THM-coupled geothermal operations.