Machine Learning Emulator Achieves 10-15 Day Ocean Forecasts Using Correlation-Aware Loss

A machine learning emulator adapted from the GraphCast architecture demonstrates skillful medium-range ocean forecasting capabilities. The system produces accurate predictions for 10-15 day lead times using a 24-hour time step and single initial condition, trained on NOAA's UFS-Replay dataset without autoregressive training methods. Researchers implemented Mahalanobis distance as a correlation-aware loss function that explicitly accounts for correlations between target variable tendencies, showing improved forecast skill compared to traditional Mean Squared Error approaches. Spatial correlation analysis reveals this loss function acts as a statistical-dynamical regularizer for slow ocean processes. The ocean-only emulator operates with prescribed atmospheric conditions, building on previous successes in atmospheric state forecasting through machine learning techniques. This research represents a significant advancement in applying artificial intelligence to global ocean dynamics prediction.