A Strain-Driven PINN Incorporating Constitutive Models for Unsaturated Soils
The hydraulic and mechanical properties of unsaturated soils are relatively complex. To address the highly nonlinear hydro-mechanical behavior of unsaturated soils, this study proposes a Physics-Informed Neural Network (PINN) model constrained by the constitutive model of unsaturated soils. Driven by axial strain, the model uses a fully connected neural network to predict five variables, including axial stress, suction, effective saturation, void ratio and confining pressure. The constitutive relationships of soil-water characteristic curves, and boundary conditions are all embedded in the loss function to ensure that the predictions comply with physical laws. The model was validated by using unsaturated triaxial test data under various overconsolidation ratios and suction conditions. Sensitivity analysis indicated that λ = 0.1 yielded the lowest combined error among the candidate weights examined, indicating a high agreement between the predicted results and experimental data. The established model can reproduce the deviator stress and volumetric changes in unsaturated expansive soil under different stress histories, and the model demonstrates promising prediction performance across the investigated strain stages under the available triaxial test conditions. Comparisons with the MLP baseline and a conventional constitutive model indicate that the proposed PINN can reproduce the main trends in the investigated hydro-mechanical responses.
Authors
- 劉承伯 (ORCID: https://orcid.org/0009-0005-2542-0479)
- Shengshen Wu (ORCID: https://orcid.org/0000-0001-8458-299X)
- Yuxin Dong
- Xinru Gong
Institutions
- Indiana University Indianapolis (US)
- China Agricultural University (CN)
Publication Details
- Journal
- Applied Sciences
- Published
- 2026-11-13
- DOI
- https://doi.org/10.3390/app162211264
- Primary Topic
- Soil and Unsaturated Flow
- Type
- article
- Field-Weighted Citation Impact
- 0.00