Forecasting Induced Seismicity in Oklahoma using Machine Learning Methods

Abstract:

Many of the earthquakes in Oklahoma have been associated with wastewater injection. In our study, we use machine learning methods – Random Forest (RF) to forecast the seismicity rate in Oklahoma. We divide the study area into uniform grids and count injection wells, modeled pore pressure and poroelastic stress points, and earthquakes in each grid. The parameters related to injections are used as features to forecast earthquake rate in a RF model. We split the data into training (2010--2016) and test (2017--2020) dataset. Our model can forecast the rapid seismicity decrease in recent years. The model also shows that pore pressure and poroelastic stress are the most important features in the forecasting. The findings are consistent with the known mechanisms of induced seismicity. Our study suggests that machine learning techniques can be applied to forecasting earthquakes and understanding the physics behind induced seismicity.