Correcting errors in seasonal Arctic sea ice prediction of Earth system models with machine learning

Abstract. While Earth system models are essential for seasonal Arctic sea ice prediction, they often exhibit significant errors that are challenging to correct. In this study, we integrate a multilayer perceptron (MLP) machine learning (ML) model into the Norwegian Climate Prediction Model (NorCPM) to improve seasonal sea ice predictions. We compare the online and offline error correction approaches. In the online approach, ML corrects errors in the model's instantaneous state during the model simulation, while in the offline approach, ML post-processes and calibrates predictions after the model simulation. Our results show that the ML models effectively learn and correct dynamical model errors in both approaches, leading to improved predictions of Arctic sea ice during the test period (i.e., 2003–2021). Both approaches yield the most significant improvements in the marginal ice zone, where error reductions in sea ice concentration exceed 20 %. These improvements vary seasonally, with the most substantial enhancements occurring in the Atlantic, Siberian, and Pacific regions from September to January. The offline error correction approach consistently outperforms the online error correction approach. This is primarily because the online approach targets only instantaneous model errors on the 15th of each month, while errors can grow during the subsequent 1-month model integration due to interactions among the model components, damping the error correction in monthly averages. Notably, in September, the online approach reduces the error of the pan-Arctic sea ice extent by 50 %, while the offline approach achieves a 75 % error reduction.