A physics-informed neural network framework for consolidation parameter prediction using controlled clay-sand mixtures

• "First PINN for Cv prediction: R²=0.9963, outperforms empirical by ΔR²=+1.31 • "Clay Activity-based constraints: no classification needed, applicable for all types of clay minerals" • "Beats best ML model: Δ35.7% RMSE, Δ44.3% MAPE over baseline (p=0.024)" • "Multi-physics loss: integrates Kozeny-Carman, activity, Cc-mv consistency" • "SHAP validates physics: plastic limit & activity as primary drivers which is in line with soil mechanics theory" This paper introduces a novel Physics-Informed Neural Network (PINN) model for predicting the coefficient of consolidation (Cv) in high plasticity clays. The model was trained from experimental data obtained from controlled clay-sand mixtures. The input parameters include clay content, Atterberg limits, initial void ratio, compaction energy, applied pressure and consolidation characteristics like compression index (Cc) and volumetric compressibility (mv). Additional parameters like plasticity index, porosity, activity-clay interaction and compaction efficiency were derived from feature engineering. The proposed PINN model combines data-driven loss and physics-based loss into a total loss function. The physics loss includes three constraints derived from modified Kozeny-Carman equations, activity-based mineralogical relations, and compression-volume consistency. Hyperparameter optimization identified the optimal configuration: 800 epochs, learning rate 0.001, architecture [128, 64, 32], and physics loss weights distributed as 0.7, 0.25, and 0.05. Five-fold cross-validation demonstrated robust performance (R 2 =0.9903±0.0026), significantly outperforming baseline neural networks (R 2 =0.9682±0.0126, p=0.0116) with 73.9% reduction in Root Mean Square Error (RMSE = 6.37 × 10 -11 m²/s) and 5.71% improvement in Mean Absolute Percentage Error (MAPE = 4.48%). External validation showed the PINN (R²=0.9968) substantially outperformed empirical correlations (best R 2 =0.1636) and conventional machine learning models (best R 2 =0.9878). SHapley Additive exPlanations (SHAP) interpretability analysis validated physically meaningful decision-making, with plastic limit and activity emerging as primary drivers. This framework provides a transferable, physics-consistent solution applicable across diverse clay types for foundation design and site characterization.