Deep Learning-Based Correction of Decadal Predictions of the PDO and TAG Indices

Conventional multi-model ensemble averages (MME) from hindcasts, such as those from the Decadal Climate Prediction Project component A (DCPP-A), can suffer from diluted predictive skill due to the averaging of models with varying performance. This dilution limits the skillful prediction of key Decadal Climate Variability (DCV) modes, such as the Pacific Decadal Oscillation (PDO) and the Tropical Atlantic SST Gradient (TAG). We propose a Long Short-Term Memory (LSTM) network to correct the ensemble averages of each participating model and hence the MME. The LSTM is trained on individual model ensemble averages and the MME to learn non-linear relationships, producing a more skillful time series for the PDO and TAG indices, relative to the conventional MME, as quantified by multiple performance metrics. These enhancements include superior temporal accuracy, reduced error, and a closer statistical correspondence in phase, timing, and magnitude with the observations. Our study demonstrates that deep learning can extract a more skillful signal from climate ensembles, offering an approach to more reliable decadal predictions.