A hybrid physics and machine learning framework for trajectory prediction and uncertainty quantification of undrained behaviour in gassy clay

The undrained response of gassy clay is strongly path-dependent and uncertain, making continuous trajectory prediction difficult when only limited experimental data are available. This study develops a lightweight physics-guided machine learning framework for path-level prediction and uncertainty quantification under undrained shearing. The framework combines MCC-based physical-reference trajectories, which provide stress-history-dependent mechanical information, with experimental gassy-clay trajectories, which provide target-domain response information. Both data sources are converted into a common rate-based representation, and three independent Random Forest regressors are trained using domain-weighted samples to predict the strain-normalised rates of p ′ , q , and Δ u w . Complete response paths are reconstructed through closed-loop sequential rollout. Grouped validation identifies β / α = 30 as the most balanced domain-weighting ratio. Monte Carlo perturbations of OCR, initial pore water pressure, and gas content are then propagated through the frozen predictor to construct pointwise empirical P5–P95 uncertainty envelopes. The supplementary saturated overconsolidated trajectory reproduces the overall q hardening response with a full-path RMSE of 12.687 kPa and retains the principal direction and curvature of the p ′ – q path. An independent strength-level validation series comprising seven observations gives an RMSE of 0.0219 and an MAE of 0.0166 for the normalised undrained shear strength. The Monte Carlo analysis further quantifies the trajectory-level variability of p ′ , q , and Δ u w under the combined perturbations of OCR, initial pore water pressure, and gas content. The proposed framework provides an interpretable response-level complement to conventional constitutive modelling for gassy clay under data-limited conditions. • Propose a hybrid physics and machine learning framework for gassy clay prediction. • Enable path-level prediction of effective and deviatoric stress and pore pressure. • Quantify uncertainty in undrained responses linked to gas content using Monte Carlo.