Machine learning of induced seismicity to inform regulatory decision-making

UK Government regulations require oil and gas operators involved in hydraulic fracturing (“fracking”) to adopt a traffic light monitoring system (BEIS, 2013). This involves (1) assessing the locations of faults before fracking; (2) monitoring seismic activity in real time; and (3) immediately suspending operations if an earthquake exceeds an arbitrary magnitude (M) 0.5. However, insights from other areas of fracking and injection indicate this regulatory approach could pose several problems: active faults are often identified only in the aftermath of earthquakes (often M > 3), and recent studies have shown significant time lags (many months) may exist between injection and seismicity. Further, the cumulative effects of clusters of wells are undetermined. Here, these problems will be addressed using machine learning. Bayesian artificial intelligence (e.g. Hincks et al., 2018) will be used to develop, test and calibrate induced seismicity models in advanced fracking (e.g. Alberta; see Schultz et al., 2018) and wastewater disposal (e.g. Oklahoma) regions. This will enable us to understand the relative importance of operational and geologic parameters, improve forecasting of induced seismicity for densely clustered wells, and investigate time dependencies between fluid injection and seismicity.

The project will involve the development and application of a Bayesian network to understand the causes of induced seismicity using freely available ‘big data’ (i.e. injection and seismicity records) and existing hydrogeological/fracture models (e.g. fault networks, permeability, pore pressure). These data will be used to learn system behavior, and identify spatial and temporal patterns in seismicity. In addition to operational and seismicity records, the model will also integrate geological datasets, e.g. stratigraphic data, and further tests will determine the extent to which forecast skill is improved. The student will also have the opportunity to handle seismic data to compute seismic constraints on the strength and magnitude of the stress field (e.g. focal mechanisms, shear wave splitting). As well as identifying key drivers and timescales, this machine learning approach will enable inherent uncertainties in the relationships to be evaluated. The project will use these constraints to formulate evidence-based recommendations for regulatory controls that take into account the complexity and time-dependent behavior involved in operational shale gas energy exploitation.

The SPITFIRE DTP programme provides comprehensive personal and professional development training alongside extensive opportunities for students to expand their multi-disciplinary outlook through interactions with a wide network of academic, research and industrial/policy partners. The student will be registered at the University of Southampton and hosted in the School of Ocean and Earth Science. Specific training will include: (i) training on computer modeling and data analysis, using Uninet and R; (ii) training in the use of ArcGIS and Petrel for 3D geological modeling; (iii) training in the use of relevant codes to handle earthquake waveform data.