Precision monitoring of the subsurface carbon-dioxide plume ensures long-term, sustainable geological carbon storage. Carrigan et al. (2013) and Yang et al. (2014) showed that electrical resistivity tomography (ERT) can accurately map the evolution of the CO2 saturation during geological carbon storage. To better monitor the CO2 plume migration in a storage reservoir, we develop an unsupervised spatiotemporal clustering to process the CO2 saturation maps derived from the ERT measurements acquired over 80 days by Carrigan et al. (2013). Using dynamic time wrapping (DTW) K-means clustering, four distinct clusters were identified in the CO2-storage reservoir. The four clusters exhibit a Davies-Bouldin (DB) index of 0.71, a Calinski-Harabasz (CH) index of 262791, and a DTW-silhouette score of 0.58. Unlike traditional clustering methods, the DTW K-means incorporates a temporal distance metric. Traditional clustering methods, such as Euclidean K-means, agglomerative and meanshift clustering, exhibit a lower performance with DB index of 0.83, 0.95, and 1.01, respectively, and CH index of 157866, 131593, and 69438, respectively. Subsequent statistical analysis indicates that contrast stretching and fast-Fourier transform are strong geophysical signatures of the spatiotemporal evolution of CO2 plume. We also identified a strong correlation between injection flow rate and the spatial evolution of regions with high CO2 content. Finally, the previously computed spatiotemporal clusters were further clustered to discover distinct temporal sequences emerging with respect to the overall CO2 plume distribution in the subsurface. Six distinct temporal clusters of CO2 plume evolution were detected over a period of 2 months. A tensor-based feature extraction was critical for processing the ERT data acquired over 80 days to capture both the temporal and spatial components relevant to the evolution of CO2 plume in the storage reservoir.

Mechanical discontinuity embedded in a material plays an essential role in determining the bulk mechanical, physical, and chemical properties. This paper is a proof-of-concept development and deployment of a reinforcement learning framework to control both the direction and rate of the growth of fatigue crack. The reinforcement learning framework is coupled with an OpenAI-Gym-based environment that implements the mechanistic equations governing the fatigue crack growth. Learning agent does not explicitly know about the underlying physics; nonetheless, the learning agent can infer the control strategy by continuously interacting the numerical environment. The Markov decision process, which includes state, action and reward, is carefully designed to obtain a good control policy. The deep deterministic policy gradient algorithm is implemented for learning the continuous actions required to control the fatigue crack growth. An adaptive reward function involving reward shaping improves the training. The reward is mostly positive to encourage the learning agent to keep accumulating the reward rather than terminate early to avoid receiving high accumulated penalties. An additional high reward is given to the learning agent when the crack tip reaches close enough to the goal point within specific training iterations to encourage the agent to reach the goal points as quickly as possible rather than slowly approaching the goal point to accumulate the positive reward. The reinforcement learning framework can successfully control the fatigue crack propagation in a material despite the complexity of the propagation pathway determined by multiple goal points.

Fluid injection into subsurface causes rock deformations, which give rise to mechanical waves in the surrounding rock. This article focuses on the infrasound signals (2-80 Hz) recorded by hydrophones during a meso-scale (~10 meter) hydraulic fracturing experiment at depth of 1.5 kilometer. We present a full-waveform-based data-driven workflow to map the spatiotemporal evolution of the infrasound sources produced during hydraulic fracturing. The infrasound source locations are compared against the simultaneously created microseismic source locations. Orientation of the infrasound source point cloud strongly agrees with natural fracture orientation, as inferred from the discrete fracture-network modelling. Finally, we arrive at a conceptual model of fluid-injection driven infrasound generation in subsurface and posit that the reopening of natural fractures is the main mechanism of the infrasound generation. A joint analysis of signals from microseismicity and infrasound sources can improve subsurface fracture imaging.