Parametric Estimation of the Local Stress Field: A Hybrid Combinatorial and Machine Learning Based Technique

Accurate determination of state of stress at a field site is key in order to evaluate the geomechanical impacts of any industrial operation that involves injection of fluids into the deep subsurface. Direct borehole measurements of the orientation of the stress field are typically expensive, spatially limited, restricted to shallow depths, and potentially compromised by the drilling itself. Though indirect measurements of the stress field orientation with seismic interferometry or through analysis of earthquake focal mechanisms may address some limitations of borehole-based approaches, they are more fundamental in nature, time consuming, and require advanced processing techniques and expert analysis that may not be readily available to small, independent operators. This technological gap represents a significant barrier for cost-effective and accurate prediction of state of stress with several negative environmental and economical impacts. To address the aforementioned critical challenge, the overarching objective of this project is to use a suite of combinatorial analysis, machine learning, and automation to develop a novel, efficient and user-friendly software platform that quickly returns and/or predicts a parameterized estimate of the stress field based solely on a user-provided data. Capitalizing on our preliminary results, phase I will develop a beta version of our software product by building a geomechanical model representative of a generalized storage complex with properties similar to those in regions of active induced seismicity. This will be followed by a global sensitivity analysis of thousands of combinatorial synthesis earthquakes induced by fluids injected into the subsurface. Phase II will develop a rich database and model of various synthetic earthquake catalogs that arise based on pre-existing and evolving stress conditions and associated statistical measures of the seismic signatures. These analyses will be enhanced by utilizing deep learning algorithms such as Fully Convolutional Neural Networks to combine all key parameters and take advantage of the cross-parameter interdependencies to reduce dimensionality and provide useful trends in state of stress. As more and more new data is gathered, the deep learning code will automatically (dynamically) update their prediction algorithms using optimizations and will become more accurate over time. Key unique features of our software product for parameterization of the stress field include 1) harnessing spatiotemporal earthquake signatures, 2) accurate stress field parameterization by inverting seismic signatures, 3) a non-invasive and spatially expansive solution, 4) an automatized smart software that can be installed on PCs/Apps/clouds, and can enrich (expand) its databank and accuracy over time, and 5) cost-effectiveness. Together, these attributes constitute a significant first-step in accurate estimation of the state of stress, potentially leading to a new line of commercialization for monitoring/mitigating induced seismicity for multiple applications including carbon storage, waste water disposal from hydrocarbon extraction, hydraulic fracturing, enhanced oil recovery, and enhanced geothermal systems. C-Crete’s experience in projects with similar complexity along with our collective state-of-the-art facilities and equipment as well as our key relationships with key industry players will catalyze our commercialization efforts.