SCEC2024 Poster #177: Interpolation Methods for CyberShake Seismic Hazard Analysis Data Products

Group A, Poster #177, Ground Motions (GM)

Interpolation Methods for CyberShake Seismic Hazard Analysis Data Products

Amelia Kratzer, Kevin R. Milner, & Scott Callaghan

Poster Image:

Poster Presentation

2024 SCEC Annual Meeting, Poster #177, SCEC Contribution #13689 VIEW PDF

CyberShake is an HPC platform that performs physics-based probabilistic seismic hazard analysis using wave propagation simulations. CyberShake results enable improved earthquake risk assessment and mitigation strategies, but they are very computationally intensive to perform. To avoid running costly simulations for every location of possible interest, we investigated methods for interpolating CyberShake results. We tested bilinear interpolation and machine learning methods to generate two key CyberShake data products: hazard curves and intensity measures (IMs). We chose to interpolate sites for which we have existing CyberShake simulated results to compare our interpolations with and evaluat...e their accuracy. For bilinear interpolation, we discovered that IM interpolations were more accurate than direct hazard curve interpolations. Additionally, including scale factors based on velocity structures, specifically Vs30, Z1.0, and Z2.0, in the IM interpolations helped to improve our accuracy. We found that our results are closer to simulated CyberShake results at 10 seconds than at 2 seconds, likely due to basin effects which are difficult to capture using the bilinear interpolation approach. For our machine learning approach, we created two separate models: a hazard curve interpolation setup composed of 51 smaller networks, each responsible for interpolating the probability at a different shaking value, and an IM interpolation setup consisting of a single network with 23 inputs - the input sites’ distances, IM values, and velocity metrics. As with bilinear interpolation, we found that the ML interpolation of IMs was more accurate than that of hazard curves. This is likely because the hazard curves are produced from the underlying IMs, so it is more effective to interpolate the IMs directly. While both bilinear and ML interpolation of IMs are viable methods to accurately interpolate CyberShake simulated results, our results suggest that bilinear interpolation produces the most accurate outcomes.
SHOW MORE