Full Waveform Inversion via Uncertainty-Aware Deep Learning for CO2 Plume Detection

This study presents a probabilistic deep learning framework developed for full waveform inversion (FWI) to detect and characterise subsurface velocity anomalies induced by stored CO2. Three convolutional neural network architectures—LinkNet, UNet, and a feature pyramid network (FPN)—were trained and evaluated on the Kimberlina CO2 synthetic dataset from the OpenFWI repository, which simulates CO2 injection scenarios within stratified reservoirs. The FPN architecture with an SE-ResNet50 backbone achieved the best performance (R2=0.9386, MSE=3.1676×10−4, MAE=4.5249×10−3). Uncertainty quantification was performed using variational inference with the Flipout estimator, which enables the separation of aleatoric and epistemic uncertainty. The predicted velocity models accurately reconstruct the background stratigraphy and the low-velocity anomaly associated with the CO2 plume. The aleatoric uncertainty remains at around 101 m/s and is spatially localised to delineate the geometry of the low-velocity zone. In contrast, the epistemic uncertainty is several orders of magnitude smaller, at around 10−3 m/s, and increases specifically over the lowest-velocity region of the anomaly. This structured distribution indicates where the velocity reconstruction is least reliable. Calibration substantially reduced the overall miscalibration area, decreasing the magnitude of the aleatoric component while slightly increasing that of the epistemic uncertainty over the CO2 plume anomaly, although the spatial structure of both components remained unchanged. This uncertainty decomposition and calibration help to distinguish between data noise and model limitations. It provides direct information for designing acquisitions and assessing risks, which is an area that is not widely covered in the current literature.