Integrating Numerical Simulations, Machine Learning and Bayesian Inversion in Investigations of Hydraulic Fracturing and Wastewater Disposal Operations, and Associated Seismic Activity
Abstract
The “shale revolution” which started in the United States in the mid-2000s significantly increases in the hydrocarbon production from unconventional reservoirs. As a result, wastewater disposal and hydraulic fracturing has become commonly conducted operations. Thus, understanding of fault and fracture mechanics in the presence of fluid flow is of a great importance to operators, regulators and scientists. In this study, we focus on applications of numerical simulations, machine learning and stochastic inversion in investigations of these processes and associated seismic activity. We first develop a forward modeling method by integrating fluid flow in a poroelastic medium and dynamic rupture on faults to estimate time of earthquake triggering and its magnitude. We explore the parameter space and find that formation permeability and elastic properties, along with fault-well distance have a big impact on accumulation of pressure and stress perturbations and eventually earthquake triggering. We apply the methodology to a real case study of the 2012 Mw 4.8 Timpson (TX) induced earthquake. We reproduce not only the size of the mainshock but also main features of the aftershock sequence, building a direct physical link between wastewater injection and the earthquake. Combining the forward modeling method with machine learning regression and Bayesian inversion, we develop a methodology to better constrain fault frictional parameters and background stress states, which translates to improvement in simulation results. We further develop our methodology to simulate complex hydraulic fracture propagation and its interaction with natural fractures or bedding planes by introducing a dual permeability model for flow simulations, mixed-mode failure criterion and cohesive elements along the fracture path. We investigate fracture propagation, activation, and interaction behaviors, and analyze the relationship between producing reservoir volume and microseismic cloud extent. We find that in many cases the Stimulated Reservoir Volume (SRV) does not correspond well with the Drained Reservoir Volume (DRV). We apply the methodology to the HFTS-1 (Hydraulic Fracturing Test Site – 1; Midland Basin) experiment case study and discuss implications of our results on field data analyses.
Citation
Szafranski, Dawid (2022). Integrating Numerical Simulations, Machine Learning and Bayesian Inversion in Investigations of Hydraulic Fracturing and Wastewater Disposal Operations, and Associated Seismic Activity. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /197097.