Extracting sub-seasonal to seasonal predictability offered by low-frequency models of variability, like El Niño Southern Oscillation (ENSO), remains a challenge for Southern California precipitation. Here, we use a multi-input multi-output autoencoder-decoder (MIMO-AE) machine learning network to extract the nonlinear covariability patterns of Tropical Pacific Ocean sea surface temperatures and Southern California precipitation. The predictive skill of the network at predicting Southern California precipitation is further assessed using Long Short-Term Memory models. MIMO-AE will be trained on multi-model ensembles of Earth System Models of various configurations including conventional low resolution models, high-resolution models, as well as super-parameterized models that have cloud resolving models embedded within each model grid box. Using transfer learning approaches, we will fine-tune the network to observational data. The benefit of including variables that inform the network of intrinsic atmospheric variability, like the geopotential height at 500mb over the Northern Pacific Ocean will also be assessed. Preliminary studies with a MIMO-AE network trained on a single model dataset and observational data indicate enhanced predictability of Southern California precipitation as compared to other ENSO indices, like the Niño3.4 index and the ENSO Longitude Index.