Variational Target Encoding for Integrating Climate Models

Abstract:

Earth System Models (ESMs) are fundamental tools for simulating major physical and biogeochemical processes to forecast climate variability. Different modeling centers develop different ESMs, which leads to inter-model variability and uncertainty about the future climate. Current Bayesian approaches to integrate multi-model ESM projections into a single forecast can suffer from high computational costs and limited flexibility due to strong parametric and linearity constraints. Instead, we propose an empirical Bayesian model called a Variational Target Encoder (VTE) that integrates multi-model ESM projections into a single spatiotemporal forecast based on observations of the target climate process. The VTE parameterizes its posterior distribution and conditional likelihood with two neural networks that respectively encode ESMs into a latent distribution and decode the latent distribution into the target forecast distribution. We show that the VTE is computationally efficient, highly flexible, automatically regularized, and has asymptotically correct uncertainty quantification. We conducted two experiments on NA-CORDEX data, using the daily precipitation and temperature fields from four ESMs as input and the corresponding fields from another ESM as the target climate. We fit separate VTEs (precipitation and temperature) on historical model runs (1950-2005) and tested their predictions on future projections (2006-2100) under RCP 8.5. The VTEs accurately recovered their respective future target fields and generated posterior credible intervals for each field with the prespecified coverage.