Research Reports
Report Details

Abstract

Reliable knowledge of subsurface soil properties is needed to predict earthquake motions and assess seismic hazards. However, these properties are often uncertain because site investigations provide only limited measurements. Conventional seismic inversion methods can estimate soil properties from recorded ground motions, but they are computationally demanding and may depend strongly on the assumed starting model. This project developed a physics-informed neural network framework that combines seismic observations with the physical equations governing wave propagation. The framework was designed to recover unknown input motions, estimate shear-wave velocity, and account for energy loss through material damping. It was evaluated using four- and six-layer soil profiles derived from published site data, followed by an application using field recordings from the Treasure Island Downhole Array. The framework recovered base motions from single surface recordings with errors below 6% and reproduced the main features of the reference shear-wave velocity profiles using only two recordings. The results also demonstrated the feasibility of estimating velocity and damping simultaneously, although damping and thin soil layers were more difficult to resolve. The field application captured the overall variation of shear-wave velocity with depth and the principal material contrasts. For transportation infrastructure, poorly characterized low-stiffness zones and other subsurface anomalies can contribute to roadway deformation and subsidence, making improved subsurface characterization important for proactive asset management. These findings show that physics-informed machine learning can support seismic deconvolution and subsurface characterization when measurement data are limited. Further development could provide engineers with an additional tool for site-response analysis, earthquake-hazard assessment, and geotechnical design.

How to Cite

Pu, Yuze, and Kami Mohammadi. Physics-Informed Machine Learning for Seismic Site Characterization from Limited Recordings, CTIPS-26-009. North Dakota State University - Upper Great Plains Transportation Institute, Fargo: Center for Transformative Infrastructure Preservation and Sustainability, 2026.