Amortized probabilistic retrieval of atmospheric CO2 from OCO-2 spectra using deep learning with Laplace approximations and normalizing flows
Published in Preprint (under review), 2026
Recommended citation: Calle-Saldarriaga, A.; Jimenez, F.; Grosskreuz, J.; Wang, J.; Hobbs, J.; Katzfuss, M. (2026). "Amortized probabilistic retrieval of atmospheric CO2 from OCO-2 spectra using deep learning with Laplace approximations and normalizing flows." Preprint, under review. https://arxiv.org/abs/2606.17413
Abstract: Retrieving atmospheric CO2 from OCO-2 spectra requires inverting an expensive full-physics forward model, and the operational Gaussian optimal-estimation solution can be poorly calibrated when the forward model is misspecified. We propose an amortized neural inference framework in which normalizing flows and Laplace approximations deliver calibrated, non-Gaussian posterior uncertainty orders of magnitude faster than the operational solver, while remaining robust to aerosol-induced forward-model error. Joint work with NASA JPL’s Uncertainty Quantification for Remote Sensing group.