Time-lapse seismic inversion with for CO2 saturation and storage - A new hybrid machine learning approach

Using convolutional neural networks, a new deep learning approach is constructed assuming continuous temporal changes in the subsurface model. SeDeepLabv3Net and Efficient deep convolutional neural networks forecast CO2 saturation using full waveform shot collects. Site-specific geological data, fluid dynamics, rock physics, and seismic modeling are used to create realistic synthetic datasets for CO2 storage sites. Following numerical research, synthetic testing showed promising results, prompting us to use SeDeepLabv3Net’s learnt weights to highly nonlinear and field data. This inversion approach was used to invert all time-lapse data simultaneously, and the resultant time-lapse and difference models were compared to seismic data. We found that pre-training convolutional neural networks on synthetic datasets and applying the weights to field data predicts CO2 saturation. This method effectively addresses field training data shortages, enabling long-term GCS monitoring.