Feature-engineered physics-conditioned DNN with validation-selected ridge residual correction for pore water pressure prediction in clays

Introduction Accurately predicting pore water pressure is important for evaluating effective stress, deformation, and stability in clayey soils. Methods This study develops an interpretable feature-engineered physics-conditioned DNN with validation-selected ridge residual correction using 42 complete curves from kaolin clay and Mercia Mudstone clay. Complete curves were divided into training, validation, and testing subsets. Three configurations were compared: a Physics-Conditioned DNN (PC-DNN), a feature-engineered PC-DNN (FE-PC-DNN), and the FE-PC-DNN with a candidate residual stage. Physical knowledge was used to select variables, construct hydro-mechanical interaction features, define a reference response, and bound the correction; governing equations were not enforced in the loss function. Results Validation assigned zero weight to the neural residual candidate and full weight to ridge regression. Accordingly, the reported residual correction was ridge-based, clipped to ±0.419394 kPa, and combined with a validation-bias offset of −0.099064 kPa. The ridge-corrected FE-PC-DNN achieved a test MSE of 1.58 kPa 2 , RMSE of 1.26 kPa, MAE of 0.82 kPa, and sMAPE of 8.75%. Global SHAP analysis identified soil type, axial strain, and confining pressure as the main features. A unified kernel-weighted B-spline equation, calibrated on 36 development curves and evaluated once on six locked test curves, achieved R 2 = 0.982, RMSE = 0.894 kPa, MAE = 0.716 kPa, and sMAPE = 6.03%. Discussion These results demonstrate strong internal performance within the tested two-soil domain; external validation is still required before field use.