Early Detection of Non-Conformance in Monitoring of CO2 Storage Reservoirs Using Auto-Encoder Neural Networks

Planning and managing operations of subsurface reservoir assets in terms of conformance is crucial for responsible CO2 storage. One important component of conformance management is the monitoring of the reservoir dynamics in response to the implemented operational strategies, particularly for early detection of deviations from intended behaviour (i.e., non-conformance). In recent work, we have introduced a model-based quantitative workflow to objectively assess the usefulness of monitoring for conformance verification in CO2 storage and shown how to use state-of-the-art supervised learning techniques to achieve a more practical workflow. In the present work, we investigate the use of a semi-supervised anomaly detection approach based on auto-encoder neural networks as an alternative to circumvent limitations of the supervised classification approaches explored so far. The results of our case study of a real storage aquifer show that auto-encoders trained on simulated time-lapse seismic data from (only) conformance scenarios can be used to accurately detect scenarios where the migration of CO2 deviates from the desired range of behaviours. These promising results confirm that the proposed approach can be used to derive efficient conformance classification workflows without an explicit finite dataset representing non-conformance to be defined in advance.