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

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