Estimating CO₂ saturation and porosity using the double difference approach based invertible neural network

Accurate monitoring of injected CO₂ volumes and CO₂ migration is critical for the success of carbon sequestration projects. In this work, we first use unsupervised deep learning to denoise seismic data. Then, we propose the use of Invertible Neural Networks (INNs) to estimate porosity and CO₂ saturation and their respective uncertainty maps from seismic data. Our network is based on the double difference approach. The use of Invertible Neural Networks (INNs) over other network architectures is motivated by the fact that INNs can produce comparable posterior pdfs of model parameters to those obtained using Markov Chain Monte Carlo methods at significantly less computational time. We use two seismic vintages and well-logs from the Cranfield reservoir in order to validate our approach.