U-Net-Accelerated Indirect Boundary Element Method for Efficient Seismic Basin Amplification Simulation

Abstract

The indirect boundary element method (IBEM) based on layered half-space Green’s functions (HSGFs) is a powerful tool for parametric studies of basin amplification effects. Once the HSGF is computed for a given basin model, simulating different source configurations can be accomplished in minutes. However, even a one-time calculation of the HSGF for layered media could remain a significant bottleneck, as closed-form solutions are usually unavailable and numerical evaluation is required. This study presents a deep-learning (DL)-assisted IBEM scheme that accelerates HSGF computation through supervised interpolation. We organize the space–frequency domain HSGF as fourth-order tensors indexed by wavefield component, source position, receiver position, and frequency, with source and receiver indices structured according to the basin’s discretization geometry, thereby recasting a significant fraction of the computation as a structured interpolation problem. The dense HSGF sampling intrinsic to IBEM provides abundant self-consistent data, eliminating the need for external training. A case-specific U-Net architecture is trained on a downsampled subset of fully computed HSGFs and then applied to predict the remaining entries. This problem-specific strategy maximizes utilization of IBEM-generated data while circumventing the generalization challenges that limit many learning-based forward solvers. Validation tests on ellipsoidal basin models show that the hybrid scheme achieves over 99% accuracy compared with full-computation References while reducing the overall computation time by approximately 35%–55% depending on the mask rate, corresponding to a saving of about one to two days in the present examples. This methodology could be extended naturally to other basin geometries and boundary integral formulations, and holds promise for large-scale applications in seismic hazard assessment and ground motion simulation.