Machine learning techniques for estimating seismic site amplification in the Santiago basin, Chile

Seismic site amplification and seismic hazard maps are crucial inputs for decision making and risk evaluation in places where seismicity imposes a significant risk to human life and infrastructure. In this work, we propose a novel machine learning (ML) based methodology to integrate qualitative and quantitative data to map the degree of seismic amplification in an area of Chile, one of the most seismically active countries on Earth. Our method uses measurements of surface shear wave velocities (V s30 ) and predominant frequencies (f 0 ) combined with gravity anomaly maps to update the geographic extension of seismic amplification classes. Additionally, we trained the predictive models to interpolate and extrapolate V s30 and f 0 to the unsampled sites. Applying this method to the Santiago basin resulted in (i) a refined seismic amplification map, and (ii) maps of V s30 and f 0 estimated with improved accuracy. The best predictions, obtained by ML techniques and validated through cross-validation, are possibly due to the inclusion of spatial covariates for algorithm training, enhancing the ability of the model to capture the spatial correlations of geological, geophysical and geotechnical data. The estimation of predominant frequencies (f 0 ) is improved considerably by including gravity as a covariant. The accuracy of the f 0 predictions apparently depends more on the choice of covariates than on the algorithm used, while the V s30 predictions are more sensitive to the chosen algorithm. These results illustrate the great potential of machine learning predictive algorithms in digital soil mapping, which surpass traditional geostatistical techniques. The major contribution of this work is to introduce a novel methodology, based on artificial intelligence models, to extend local measurements of site-specific dynamic properties. This information can be used to quantitatively estimate seismic hazard over a regional scale. • We propose a methodology for integrating data to map dynamic site characterization using machine learning. • Our method uses measurements of Vs30 and f0 combined with gravity anomaly and geology. • We trained predictive models to interpolate Vs30 and f0 in the Santiago de Chile basin. • The best predictions were obtained thanks the inclusion of spatial covariates for algorithm training. • The estimation of f0 is considerably improved by including gravity as a covariant.