Multi model ensemble analysis with Neural Network Gaussian Processes

We introduced a NNGP model to take an ensemble of climate model output (CMIP6) and predict the corresponding reanalysis field (ERA5). We applied our model to both temperature and precipitation. Out of sample experiments showed that the NNGP significantly improves forecast accuracy and image reconstruction fidelity (pictured) over previous approaches based on model averaging. Current work seeks to further improve this models robustness to distribution shift (climate change) and better quantify the prediction uncertainty.