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

Abstract

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 fP, HV, VS30, and depth Z2.5 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.

Reducing the uncertainties in the NGA-West2 ground motion models by incorporating the frequency and amplitude of the fundamental peak of the horizontal-to-vertical spectral ratio of surface ground motions

“…A third possible solution is using machine learning in predicting site parameters and local site effect. For instance, Ji et al (2021) and Zhu et al (2022) used parameters obtained from HVSR to predict earthquake site response for stations in Japan, and Akhani and Pezeshk (2022) reduced the uncertainties in GMMs by optimizing the regression coefficients. Also, Tamhidi et al (2022) proposed a new approach in predicting ground motion IMs at a target location (uninstrumented sites), which would allow estimating the site-specific HVSR-based proxies.…”

“…For brevity, g i is used hereafter as the period-dependent coefficient. Previous studies (Hassani and Atkinson, 2016c; Zhu et al, 2022) proposed a model that peaks at the site fundamental frequency to capture the resonance effect. Although the model shown in Equation 8 is not restrained to capture resonance effects, it can be seen that for most of the periods, the maximum prediction of the proposed model (Equation 8) occurs when the site resonance frequency is close to the frequency (or period) of interest.…”

Machine learning‐based peak ground acceleration models for structural risk assessment using spatial data analysis

“…While data-driven models have shown significant promise in various domains, they do indeed come with limitations, especially when compared to physics-based models. 54,55 Physics-based models, rooted in fundamental principles and governing equations, have the advantage of being able to interpolate and predict ground motion intensity measures without relying heavily on historical data from the specific location. These models leverage a deep understanding of the underlying physical processes, allowing them to provide accurate predictions even in scenarios where there might be limited or no historical data available.…”

“…By capturing a richer set of spatial information, such as proximity to fault lines or geological characteristics, the models can gain a more comprehensive understanding of the factors influencing the PGA and achieve more accurate predictions. While data‐driven models have shown significant promise in various domains, they do indeed come with limitations, especially when compared to physics‐based models 54,55 . Physics‐based models, rooted in fundamental principles and governing equations, have the advantage of being able to interpolate and predict ground motion intensity measures without relying heavily on historical data from the specific location.…”

Unsupervised probabilistic machine learning applied to seismicity declustering: a new approach to represent earthquake catalogues with fewer assumptions

“…Earthquake catalogues are key datasets widely used by the scientific community for understanding the statistical behaviour of earthquakes, their spatio-temporal evolution and their triggering factors. They can also highlight the 3D geometry of seismically active structures, contribute to the quantification of seismic hazard and improve earthquake forecasting (Zhu et al, 2023). In addition, new generations of high-definition seismic catalogues are being built with more powerful detection procedures.…”