Using multi-regression machine learning to assess uncertainties of surface wave phase velocities derived from ambient noise

With the increasing use of ambient seismic noise to investigate the crust and uppermost mantle structures, accurately estimating the uncertainties of surface wave phase velocities derived from noise data is more crucial than before. Additionally, the rapid growth of Machine Learning (ML) has sparked interest in its potential contributions to geophysical research. In this study, we calculate Noise Correlation Functions (NCFs) with different stacking durations from 2.5 years of continuous ambient noise data recorded by 84 stations in Iran and neighboring countries. Then, phase velocity dispersion curves are measured at a period range of 10–50 s. We assess phase velocity uncertainties as a function of four variables of NCFs by using a multi-variable regression model, including Signal-to-Noise Ratio (SNR), inter-station distance, stacking duration, and period of NCFs. The empirical relationship between phase velocity uncertainty and these variables is utilized to understand the variations in uncertainty. Three machine-learning algorithms are adopted to test for performance, including Artificial Neural Networks (ANNs), Support Vector Regressions (SVRs), and Gradient Boosting Regression (GBR). SVR shows the highest performance with the highest correlation coefficient ( R 2 ) and a lower Mean Absolute Error (MAE). Additionally, sensitivity analysis of these three models shows that the inter-station distance plays a crucial role in predicting the uncertainty of phase velocity, followed by stacking duration, SNR, and period. The model of SVR can be utilized to estimate uncertainties of phase velocities in other applications of ambient noise tomography in the future.