Abstract: To address the challenges of data-scarce sample size, sparse spatial distribution, and inadequate physical consistency in predicting marine clay parameters, we developed a Gaussian process regression (GPR) model that integrates a physics-guided augmentation(PGA) algorithm with a weighted kernel strategy. In the proposed PGA-GPR model, physical knowledge, including the effective stress principle, the upper bound of shear strength, and the overconsolidation ratio, is incorporated together with a multi-kernel weighting mechanism. This design improves the model’s ability to capture nonlinear behavior and ensures physical consistency. Using data on Norwegian marine clays from the TC304b database, we validated the model’s ability to predict marine clay parameters as a function of depth. The results show that, under sparse-sample conditions, PGA-GPR increases determination coefficient R² by 17%–53% relative to conventional machine learning models. It also provides higher prediction accuracy and more stable performance. Moreover, compared with the stratified random field model previously proposed by the authors, PGA-GPR more effectively captures depth-dependent variations in the overconsolidation state of marine clays, demonstrating both predictive effectiveness and physical consistency. Additionally, PGA-GPR provides reliable prediction intervals, with at least 84% of the measured values for each soil parameter falling within the 95% confidence interval. These results suggest that the model offers a promising methodology for geotechnical engineering modeling in data-scarce scenarios.

Key words: marine clay, data-scarce sample, physics constrain, data augmentation, Gaussian process regression model