Correcting Climate Model Sea Surface Temperature Simulations with Generative Adversarial Networks: Climatology, Interannual Variability, and Extremes

Climate models are vital for understanding and projecting global climate. However, these models frequently suffer from biases that limit their accuracy in historical simulations and the trustworthiness of future projections. Addressing these challenges requires overcoming internal variability, hindering direct alignment between model simulations and observations and thwarting conventional supervised learning methods. Here, we employ an unsupervised Cycle-consistent Generative Adversarial Network (CycleGAN), to correct daily Sea Surface Temperature (SST) simulations from Community Earth System Model 2 (CESM2). Our results reveal that CycleGAN not only corrects climatological biases but also improves the simulation of major dynamic modes including El Niño-Southern Oscillation (ENSO) and Indian Ocean Dipole mode, as well as SST extremes. Notably, it substantially mitigates climatological SST bias, decreasing the Root Mean Square Error (RMSE) by 58%. Furthermore, it markedly refines the representation of the annual cycle in the tropical Pacific, reducing the RMSE by 31% and boosting the pattern correlation coefficient (PCC) by 34%. Intriguingly, CycleGAN effectively addresses the well-known excessive westward bias in ENSO SST anomalies, a common issue in climate models. Additionally, it augments the simulation of SST extremes, raising the PCC from 0.56 to 0.88 and lowering the RMSE from 0.5 to 0.32. This enhancement is attributed to better representations of interannual variability and variabilities at intraseasonal and weather scales. This study offers a novel approach to correct global SST simulations, and underscores its effectiveness across different time scales and primary dynamical modes.