Application of unsupervised machine learning approach for characteristics of microseismic events at a CO2 injection site

Abstract:

Mechanistic understanding of seismic events is fundamentally important in many geoscience applications due to its significance to public safety, efficiency of energy recovery from the subsurface, and mitigation strategy for natural and induced earthquake hazards. Here, we use a subset of the microseismic data measured at the Illinois Basin Decatur Project (IBDP) site where CO2 has been injected over three years to delineate the fault zone architecture by evaluating the characteristics of seismic response. One of the biggest challenges we face during subsurface energy activities with fluid injection is the presence of hidden/unknown faults which can be reactivated by stress and pore pressure changes caused by fluid injection. As in the recent re-analysis of microseismic data at the IBDP, a majority of microseismic activities have been located within the Precambrian basement bedrock underlying the CO2 injection reservoir formation. Since CO2 injection within storage target formations is maintained below fracture pressure, induced activity generally develops at natural pre-existing regions of mechanical weakness. Hence, the low magnitude microseismic activities at the IBDP site may reveal the hidden/unknown faults that are more susceptible to changes in stress conditions. In this work we explore an unsupervised machine learning (ML) approach that combine the nonnegative matrix factorization (NMF) as dimension reduction and the Hidden Markov Model to construct a probabilistic architecture. Waveform data has been transformed into the spectrogram that was used as the input data for the ML process. The HMM identifies small changes in spectral content of the data whose spatio-temporal patterns can serve as the fingerprints of waveform characteristics. These fingerprint patterns are clustered to identify similar characteristics of the waveform data which can reveal time dependent patterns of the microseismic activities associated with hidden/unknow fault/fracture structures. We will present how the spatio-temporal patterns of these clusters are related to pore pressure and stress perturbation caused by CO2 injection activities and a new methodology to reveal the hidden/unknow fault architectures. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.