Physics-guided well-family multi-expert ensemble model for in situ stress prediction using well logs

Accurate estimation of in situ stress from conventional well logs is challenging when stress measurements are available for only a few wells and blind wells differ in pressure regime, logged interval, or curve baseline. We propose a physics-guided well-family multi-expert ensemble for predicting maximum and minimum horizontal in situ stresses (SHMAX and SHMIN). Conventional logs are augmented with overburden and effective-stress proxies, pressure interactions, local curve descriptors, and well-level statistics. The model comprises Direct, depth-normalized, and Physics-residual extremely randomized trees (ExtraTrees) experts trained on 10 development wells. Development-well leave-one-well-out predictions determine the three-expert blend weights, whereas an absolute 54 MPa/1,600 m routing rule is frozen before final expert fitting and blind-well evaluation. Labels from the eight blind test wells are not used for feature construction, weight selection, expert fitting, route assignment, or prediction. In cumulative ablation, the sample-weighted average mean absolute error (MAE) decreases from 1.8437 MPa for raw logs plus ExtraTrees to 1.3776 MPa after physics-guided feature construction, 1.2564 MPa after multi-expert blending, and 0.9589 MPa after the complete fixed 54 MPa/1,600 m gate. The final SHMAX and SHMIN MAEs are 1.0284 and 0.8894 MPa, respectively; the average MAE is 0.9589 MPa, and the average is 0.9819. These results indicate complementary contributions from physics-guided representation, target-space diversification, and well-level routing; however, generalizability remains limited by the single-basin, 18-well dataset.