Use of Double Machine Learning for Microseismic Data Analysis: What Factors Influence Microseismic Events?

Abstract

Abstract This study proposes a new workflow for analyzing microseismic data using causal inference techniques. This approach allows us to identify the causal relationships between a new microseismic event and its spatiotemporally proximal, prior microseismic events, while taking into account confounding variables that influence both the cause-and-effect variables. We applied this workflow to microseismic data acquired from hydraulic fracturing operations on 2 horizontal wells in the Marcellus Shale. Our results revealed several new insights into the microseismic source mechanisms, including: 1) The magnitude of a new microseismic event does not depend on the number or spatial and temporal concentrations of the spatiotemporally proximal, prior events; 2) When the maximum magnitude of prior event in a region increases, a new event occurs much earlier in that region; 3) A region with a large number of microseismic events will produce a new microseismic event much earlier in time than a region with fewer events. These causal relationships suggest that accurately selecting confounders is crucial for obtaining accurate causal estimates. Failure to properly select confounders can result in significant overestimation or underestimation of the causal estimates, as high as +/- 100%. Our results also confirm that causation and correlation are two distinct concepts. A causal analysis with true confounders reveals the true causal relationship that cannot be quantified using correlation/association methods. This is demonstrated using Double Machine Learning (DML) to compute the average treatment effect for both a true confounder variable and a random confounder variable.