Efficient inference and learning of a generative model for ENSO predictions from large multi-model datasets

Abstract

Abstract Historical simulations of global sea-surface temperature (SST) from the fifth phase of the Coupled Model Intercomparison Project (CMIP5) are analyzed. A state-of-the-art deep learning approach is applied to provide a unified access to the diversity of simulations in the large multi-model dataset in order to go beyond the current technological paradigm of ensemble averaging. Based on the concept of a variational auto-encoder (VAE), a generative model of global SST is proposed in combination with an inference model that aims to solve the problem of determining a joint distribution over the data generating factors. With a focus on the El Niño Southern Oscillation (ENSO), the performance of the VAE-based approach in simulating various central features of observed ENSO dynamics is demonstrated. A combination of the VAE with a forecasting model is proposed to make predictions about the distribution of global SST and the corresponding future path of the Niño index from the learned latent factors. The proposed ENSO emulator is compared with historical observations and proves particularly skillful at reproducing various aspects of observed ENSO asymmetry between the two phases of warm El Niño and cold La Niña. A relationship between ENSO asymmetry and ENSO predictability is identified in the ENSO emulator, which improves the prediction of the simulated Niño index in a number of CMIP5 models.