A Workflow to Integrate Numerical Simulation, Machine Learning Regression and Bayesian Inversion for Induced Seismicity Study: Principles and a Case Study.

The main objective of this work is to develop a new workflow integrating numerical simulations of fluid flow and dynamic rupture propagation, machine learning regression techniques and Bayesian inversions of subsurface model parameters. We present the theory behind each step as well as practical application of the proposed methodology on the May 2012, M_w4.8 Timpson, TX, earthquake. Numerical simulations show that the triggering of the earthquake is related to the wastewater disposal with the dominant role of poroelastic stress changes. Dynamic rupture simulations allow us to reproduce the size of the earthquake. Using the results from a set of simulations, we form a training dataset and compare the performance of different regression algorithms. Random Forest, Bagging and K-Neighbors regression algorithms are the most promising and we use them in the inversion procedure to replace numerical simulations. We test multiple inversion scenarios and cross-validate them with the results of numerical simulations to constrain stress state and fault frictional parameters by matching the observations (the moment magnitude of the real event). We also discuss the limitations of the current methodology and propose further extensions in the future. [ABSTRACT FROM AUTHOR]