Finite-Element-Informed Neural Surrogates for Elastic Site Characterization from Sparse Data: A Feasibility and Sensitivity Study

Data-driven site characterization in geotechnical engineering increasingly relies on high-dimensional waveform data and computationally intensive inverse modeling. Full waveform inversion and finite element model updating typically rely on gradient-based or Bayesian optimization, requiring many serial forward simulations, making large-scale applications computationally expensive. In this pilot study, we propose an alternative methodology based on physics-informed neural surrogate models—FE–PINN (continuous time) and FE–NODE (discrete time)—that embed fully assembled spatial finite-element mass–damping–stiffness matrices directly into physics loss residuals. This framework infers subsurface material parameters from sparse geodata without requiring repeated forward solver execution during iteration. Using a two-dimensional synthetic site characterization example, we demonstrate that the proposed surrogates can recover soil parameters with good accuracy, exhibit robustness to measurement noise, and reveal sensitivity to parameter initialization. Across a 240-case sensitivity sweep varying data sparsity, noise (0–50%), parameter initialization (±30%), and material heterogeneity, the proposed surrogates successfully recover soil parameters from as few as 2 measured surface degrees of freedom. Specifically, FE–PINN demonstrates strong noise tolerance, maintaining relative parameter errors below 5% for up to 25% measurement noise and 30% initialization error. In contrast, FE–NODE achieves fast convergence in clean, homogeneous cases but fails to reliably converge in heterogeneous media, where parameter errors reach 90–113% due to discrete finite-difference error propagation. The results highlight the potential and operational limits of finite-element-based neural surrogates as an auto-differentiable framework for geotechnical inverse analysis.