A denoising diffusion probabilistic modeling approach for predicting CO2 plume evolution from seismic shot gathers

Geological carbon sequestration (GCS) is being pursued globally as a geoengineering measure for reducing carbon dioxide (CO2) emission. Periodic monitoring of the injected CO2 plume is required to minimize the risk of leakage from host formations. Time lapse seismic monitoring provides one of the few non-invasive imaging techniques for routine GCS monitoring. The inversion process required to convert seismic data to CO2 plume estimates is highly nonlinear and challenging because of uncertainty in geological and petrophysical properties. Deep learning based inversion methods are being actively explored to expedite inversion. However, most of the existing deep learning based seismic inversion studies are deterministic, offering none or very limited prediction uncertainty estimation. Here we present a probabilistic machine learning method for seismic inversion of CO2 saturation. Specifically, the denoising diffusion probabilistic model (DDPM), which is a type of probabilistic deep generative model, is trained to learn the inverse mapping between the seismic gather shots and CO2 plume shapes, with quantification of the prediction uncertainty. Efficacy of the DDPM approach is demonstrated over a realistic seismic and CO2 plume dataset pertaining to the Frio-II (Texas, USA) field experiments. Results show DDPM successfully captures the evolution of CO2 plumes. The superior generative abilities of DDPM allow for accurate predictions of CO2 saturation and estimating prediction uncertainties by Monte Carlo sampling. The combination of physics-informed synthetics, and probabilistic DDPM tools shall constitute a cost-effective tool for monitoring, verification, and accounting of CO2 volumes.