24AI Using Multi-input Multi-output Autoencoders (MIMO-AE) to enhance predictability of Southeastern US Precipitation from Tropical Pacific Sea Surface Temperatures

Predictability of extratropical sub-seasonal to seasonal climate variability remains challenging. While tropical climate variability phenomenon like El Nino Southern Oscillation (ENSO) exhibit relationships with remote extratropical regions, the correlations are generally weak, and they offer poor predictive skills. Here, we apply a Multi-Input Multi-Output Autoencoder (MIMO-AE) network to assess the predictability offered by the Tropical Pacific when accounting for the non-linear relationship between Tropical Pacific and Southeastern US, following a previous work that used a similar network to find enhanced predictability over Southwest US (Passarella and Mahajan, 2023). The MIMO-AE network learns the encodings that capture the non-linear relationship between TP-SST and SE-US-PRECIP as it learns to reconstruct both variables simultaneously while passing through a low dimensional (one node here) bottleneck layer. We train the network on a historical simulation with US Department of Energy’s Energy Exascale Earth System Model (E3SMv1). We also use a transfer learning approach to apply the network to observational data. Further, we use a long short term memory (LSTM) model to predict the values at the one-node bottleneck layer, which is termed as the MIMO-AE index. Preliminary results suggest that the MIMO-AE index exhibits a high correlation (>0.9) to SE-US-PRECIP, while also exhibiting high predictive skill owing to the contribution from the TP-SST, likely due to the persistence of SST anomalies, to the index. We find that MIMO-AE network offers significant correlation skills over a lead time of up to four months for SE-US-PRECIP, enhancing predictability over the region as compared to that offered by ENSO indices like Nino3.4 index. We plan to present results from a MIMO-AE network pre-trained on a larger set of simulation ensembles from the E3SM and CMIP6 archive.