How well can we predict earthquake site response so far? Machine learning vs physics-based modeling

In site-specific site-response assessments, observation-based approaches requiring a target–reference recording pair or regional network cannot be implemented at many sites of interest. Thus, various estimation techniques have to used. How effective are these in predicting site responses (average over earthquakes)? To address this question, we conduct systematic comparison using large data set which consists detailed metadata and Fourier outcrop linear based on observations 1725 K-NET KiK-net sites. We first develop machine learning (i.e. random forest ( RF)) amplification models training (1580 sites). Then test compare their predictive powers 145 independent testing with that the one-dimensional (1D) ground response analysis (GRA). The standard deviation residuals between predictions, is, between-site (site-to-site inter-site) variability, is used as benchmark. Results show model few predictor variables, surface roughness, peak frequency f P, HV , V S 30 depth Z 2.5 achieves better performance than physics-based modeling (GRA) 1D velocity profiles. This implies can more existing information GRA inflicted by high level parametric uncertainties. finding warrants further exploration effect characterization, especially transferability across different regions.