Sim-to-Real Deep Learning for CO$_2$ Plume Inversion

Simulation-trained deep learning for scientific inverse problems can exploit physical variables that are unavailable at real-world deployment. We study this simulation-to-real mismatch through satellite-based CO$_2$ plume inversion, comparing convolutional neural networks trained with no wind information, ERA5 reanalysis winds, and higher-fidelity SMARTCARB simulation winds. On held-out synthetic power plants, explicit wind substantially improves emission estimates, with SMARTCARB winds yielding the lowest error. Controlled perturbations show that the networks learn physically meaningful wind--emission relationships, while ERA5 wind errors propagate systematically into predicted emissions. We then transfer the models trained on synthetic data to real OCO-3 satellite observations without using real data for training. Despite being less faithful to the simulated transport, the ERA5-trained model performs best on real observations; unlike the SMARTCARB-trained model, it receives the same meteorological product during training and deployment. More broadly, for simulation-trained deep learning, the most accurate physical input in simulation may not be the one that transfers best to real-world deployment.