Application of Multivariate Gaussian Correction in Predicting the Performance of Geocell-Reinforced Sandy Soils

Journal of Ocean Engineering and Marine Energy · Springer Nature · 2026-10-12 · DOI: 10.1007/s40098-026-01638-8

Аннотация

Reliable prediction of Duncan-Chang constitutive parameters remains challenging when laboratory data are limited, while direct application of data-driven predictions to a specific engineering site may introduce additional uncertainty. This study develops an integrated prediction–updating–constitutive framework in which Duncan-Chang parameters are first predicted from basic soil properties using machine learning, subsequently updated using site-specific information through a Johnson-transformation-based multivariate Gaussian conditional framework, and finally propagated through an incremental constitutive model for geocell-reinforced sand. The multivariate Gaussian and Johnson-transformation procedures follow established probabilistic concepts, while the contribution of the present study lies in their integration with data-driven parameter prediction and reinforced-soil response analysis. Among the candidate machine-learning models, GRNN-Bayes provided the most balanced held-out prediction performance, with test-set RMSE and MAE of approximately 1.58° and 1.17° for the friction angle, while the MAPE and MaxError for k were 10.70% and 30.33. Validation of the incremental constitutive formulation against published PET- and LLDPE-reinforced sand tests yielded a mean R 2 of approximately 0.982, with mean absolute deviations of approximately 3.71% for peak deviatoric stress and 6.04% for failure strain. In the site-specific Wei River application, the 19 paired model cases showed mean reinforcement-induced peak-stress increases of 380.35, 420.85, and 442.87 kPa at confining pressures of 200, 300, and 400 kPa, corresponding to relative increases of approximately 41.1%, 31.1%, and 25.0%, respectively. The corresponding failure-strain changes were comparatively small and pressure-dependent. These results demonstrate the potential of the proposed workflow for combining data-driven parameter estimation, site-specific probabilistic updating, and reinforced-soil constitutive analysis within the represented calibration domain.

Читать у издателя →

Похожие статьи