A hybrid optimal estimation and machine learning approach to predict atmospheric composition

Abstract

We present a HYbrid REtrieval Framework (HYREF) that predicts subcolumn carbon monoxide (CO) concentrations from Cross-track Infrared Sounder (CrIS) observations, trained to replicate the TRopospheric Ozone and its Precursors from Earth System Sounding (TROPESS) retrievals based on optimal estimation (OE). Unlike the OE algorithm, which produces retrievals for only a small fraction of available CrIS observations due to computationally expensive but physically accurate radiative transfer, the addition of machine learning (ML) techniques enables full coverage by providing high-resolution predictions for every valid CrIS sample. Importantly, in addition to CO concentrations, TROPESS-HYREF also predicts key retrieval diagnostics, namely column averaging kernels, degrees of freedom, and retrieval errors, that are essential for meaningful comparison with other observations, models, and ingestion into data assimilation. The framework is designed to emulate and extend the OE retrieval, rather than replace it, by providing full spatial coverage and enhanced resolution consistent with the underlying physical solution. The new framework achieves excellent performance with correlation coefficients r>0.99 and a bias <0.1% when benchmarked against an independent test set, and reproduces fine-scale spatial patterns in CO fields observed during a major wildfire over North America. A scale analysis reveals substantial variability in CO concentrations below the nominal 0.80° resolution of the TROPESS OE retrieval, which TROPESS-HYREF successfully resolves. Inference is computationally efficient, with daily global predictions completed in minutes on a single compute node. By filling observational gaps while maintaining consistency with the OE retrieval, this fusion of OE-derived physical information and ML-driven efficiency provides a practical pathway to high-resolution atmospheric CO monitoring with robust diagnostics.