A machine learning modelling outline for short-term seismic hazard assessment - RASIM 2022

By its nature, seismology has been a data intensive field since the first digital computers were introduced in the 1950s. The first empirical models were developed in the same period and are still widely used today, for instance, the Gutenberg-Richter frequency-magnitude relationship (Gutenberg and Richter, 1956). Machine learning (ML) methods make it possible to redefine these empirical models using large datasets and examining considerably more parameters, where their significance to the problem being addressed is methodically and repeatably evaluated using robust statistical methods. In this paper we develop a simple logistic model for the short-term prediction of seismic events above a threshold magnitude from a mine’s event catalogue comprising more than a million events with 14 basic parameters from which a pool of over 830 derived parameters was developed, covering a series of 11 precursory time windows. The logistic model was fitted stepwise using SAS with α = 0.025 for both entry and acceptance. 25 variables from the pool met the acceptance criterion. The findings were surprising in that parameters a rock engineer or seismologist might have expected to be the most significant in the logistic model were not; instead, it was an increase in standard deviation of the immediately-prior seismic magnitudes that carried the most weight, followed by the average maximum source displacement and median apparent volume change. These findings make intuitive sense because the increasing scatter of the magnitude shows an increasing number of large events being recorded that could be indicative of instability in the system that could lead to large events. The model is rudimentary and enormous scope exists for more exploration in terms of data pre-processing, model types, and response variable definitions.