Embedding geotechnical knowledge into physics-informed neural learning for robust marine clay compressibility prediction

Accurate estimation of the compression index Cc is important for assessing the settlement behavior of marine clay. However, conventional consolidation tests are time-consuming, while empirical correlations may have limited applicability across regions. This study proposes a knowledge-guided robust physics-informed neural network (KGR-PINN) for predicting marine clay Cc. The model integrates physical monotonicity constraints, empirical-correlation-based soft guidance, and region-robust training within a unified framework. A database of 1,000 samples from 170 sites was used, with initial void ratio, natural water content, liquid limit, and plasticity index as inputs. KGR-PINN achieved the best test performance (R2=0.9890, RMSE = 0.0691), outperforming conventional machine-learning models and empirical correlations. Ablation results confirmed the complementary contributions of the three knowledge-guided components to predictive accuracy, robustness, and physical consistency. Knowledge-Guided Robust SHAP (KGR-SHAP) analysis identified initial void ratio as the dominant predictor, followed by natural water content and liquid limit, and revealed their interaction effects. In a cross-database evaluation on 214 marine-clay records from the TC304 database, not used for model training, KGR-PINN achieved R2 = 0.9570 and RMSE = 0.0990. The proposed KGR-PINN provides a physically consistent approach for preliminary compressibility estimation of marine clay within the soil-property ranges represented by the available data.