Reconstruction of Arctic sea ice thickness (1992–2010) based on a hybrid machine learning and data assimilation approach

Abstract. Arctic sea ice thickness (SIT) remains one of the most crucial yet challenging parameters to estimate. Satellite data generally present temporal and spatial discontinuities, which constrain studies focusing on long-term evolution. Since 2011, the combined satellite product CryoSat-2 (CS2) and Soil Moisture and Ocean Salinity (SMOS), CS2SMOS, enables more accurate SIT retrievals that significantly decrease modelled SIT errors during assimilation. Can we extrapolate the benefits of data assimilation to past periods lacking accurate SIT observations? In this study, we train a machine learning (ML) algorithm to learn the systematic SIT errors between two simulations of the model TOPAZ4 over 2011–2022, one with CS2SMOS assimilation and another without any assimilation, to predict the SIT error and extrapolate the SIT prior to 2011. The ML algorithm relies on SIT coming from the two versions of TOPAZ4, various oceanographic variables, and atmospheric forcing from ERA5. Over the test period of 2011–2013, the ML method outperforms TOPAZ4 without CS2SMOS assimilation when compared to TOPAZ4 assimilating CS2SMOS. The root-mean-square error (RMSE) in Arctic-averaged SIT decreases from 0.42 to 0.28 m and the bias from −0.18 to 0.01 m. Also, despite the lack of observations available for assimilation in summer, our method still demonstrates a crucial improvement in SIT. Relative to independent mooring data in the central Arctic between 2001 and 2010, mean SIT bias reduces from −1.74 to −0.85 m when using the ML algorithm. In the Beaufort Gyre, our method approaches the performance of a basic correction algorithm. Ultimately, the ML-adjusted SIT reconstruction reveals an Arctic mean SIT of 1.61 m in 1992 compared to 1.08 m in 2022. This corresponds to a decline in total sea ice volume from 19 690 to 12 700 km3, with an associated trend of −3153 km3 per decade. These changes are accompanied by a distinct shift in SIT distribution. Our innovative approach proves its ability to correct a significant part of the primary biases of the model by combining data assimilation with machine learning. Although this new reconstructed SIT dataset has not yet been assimilated into TOPAZ4, future work could enable the correction to be further propagated to other sea ice and ocean variables.