Abstract: Conventional methods for back-analysis of slope soil parameters relied on subjective experience, while purely data-driven approaches tended to neglect underlying physical mechanisms. To address these limitations, a back-analysis method based on physics-informed neural networks (PINNs) that integrated elastoplastic deformation analysis theory was proposed. The governing equations and boundary/initial conditions of slope elastoplastic analysis were embedded as constraints into the neural network training process. A back-analysis framework jointly driven by physical laws and monitoring data was established. The soil parameters to be back-analyzed were treated as trainable hyperparameters in the network. By simultaneously optimizing these and the network's own parameters, the model's predictions were constrained to satisfy both the physical principles and the observed displacements. Validation via two-dimensional (2D) numerical cases demonstrated that that the proposed method could efficiently back-analyze elastic modulus, cohesion, and internal friction angle using only monitored displacement data, eliminating the need for extensive pre-computed numerical simulations to generate training samples. Its computational efficiency was shown to be significantly higher than that of conventional data-driven methods. Furthermore, the trained PINN model itself served as a high-fidelity surrogate model capable of directly predicting displacement fields, thereby achieving an integration of parameter back-analysis and forward analysis.

Key words: slope, back-analysis, physics-informed neural network (PINN), elastoplastic analysis, data-driven and physics-informed