Strongly Coupled Data Assimilation for Paleoclimate Reconstruction Using Deep Learning‐Based Models

Abstract Coupled data assimilation (CDA) could be beneficial for paleoclimate reconstruction, but intermediate‐complexity climate models are often used. The deep learning (DL)‐based surrogate model that can realistically simulate the climate system provides alternative to CDA, so that CDA of multi‐timescale proxy data using DL‐based models is investigated for paleoclimate reconstruction. Results reveal that at the annual time scale, coupling an oceanic component to the atmospheric model is beneficial for the surface air temperature (SAT) forecasts, but coupling an atmospheric component to the oceanic model slightly degrades the sea surface temperature (SST) forecasts, since coupling the atmospheric component introduces high frequent noises to the oceanic model. Compared to the coupled model component, cross‐component update by assimilating cross‐component proxy data has larger positive impacts on SAT and SST reconstructions. Moreover, simultaneously assimilating atmospheric and oceanic proxy data with a coupled model yields the most accurate SAT and SST reconstructions.