Dimensionless physics-informed neural networks for remote-excitation SH-wave propagation in truncated domains

Physics-informed neural networks (PINNs) are explored for the first time as a mesh-free solver for two-dimensional shear–horizontal (SH) wave propagation in truncated domains excited by remote seismic sources. Building on the Domain Reduction Method (DRM), we derive a jump (substructuring) condition that transfers free-field displacements and tractions onto an interior interface and combine it with Lysmer–Kuhlemeyer (LK) absorbing and transmitting boundary conditions. A dimensionless scaling, extended here from one to two dimensions, removes heuristic loss-weighting and markedly improves training robustness. Three truncation configurations of increasing sophistication are benchmarked against finite-element (FEM) solutions for (i) a synthetic toy model, (ii) a field-scale homogeneous basin, and (iii) a heterogeneous basin with depth-varying shear-wave velocity. Across all cases, the dimensionless PINN reconstructs displacement fields with normalized errors below 9%, while the most effective configuration — separating the jump interface from the absorbing boundary by half a characteristic wavelength — reduces FEM truncation error to $<$ 4% and total PINN error to $<$ 8%. The proposed framework, therefore, offers an accurate and flexible alternative for large-scale site response and soil–structure interaction analyses subjected to remote excitation.