Supervised and Unsupervised Machine Learning Applications for Induced Seismic Data Analysis of Fracture Initiation and Propagation

Abstract:

Quantifying in-situ subsurface conditions and predicting fracture development are critical to reducing risks of induced seismicity and improving modern energy activities in the subsurface. In this work, we developed a novel integration of both supervised and unsupervised machine learning methods to identify mechanical failure processes observed under controlled laboratory experiments and estimate reservoir response of fault (re)activation from induced seismic data observed at a field demonstration site of CO2 injection. For laboratory experiments, we integrated acoustic emission (AE) data collected during three-point bending testing of 3D printed and natural samples, digital image correlation (DIC) and micro-CT data. In addition, synthetic wave data generated from numerical simulations under different failure mechanisms is augmented to constrain classification of the relevant waves related to rock deformation. The methods tested with laboratory data have been applied for microseismic data obtained at Illinois Basin Decatur Project (IBDP) to characterize the different patterns observed in the Precambrian basement rock formation and reservoir formation where CO2 was injected into the reservoir formation. An unsupervised machine learning has been developed as a fingerprint-based clustering approach with nonnegative matrix factorization and hidden Markov model. The set of clusters resulting from the fingerprint-based approach will be also used to train a supervised approach such as convolutional neural networks to detect the new seismic events that can be used to improve the (hidden) fault identification. This research will improve characterizing seismic waveforms by machine learning approaches and the detection of low-magnitude seismic events leading to the discovery of hidden fault/fracture systems. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.