The NSF National Center for Atmospheric Research has developed the Community Research Earth Digital Intelligence Twin (CREDIT) an open, foundational research platform for developing AI Earth System Prediction models. Recent work with the platform has focused on adding model-architecture-agnostic physics constraints and on learned scale-focused latent perturbations to generate calibrated ensembles. Global physics constraints on mass and energy applied during both training and inference lead to both more accurate predictions and stable rollouts of the emulator. They can also improve upon the original training data by correcting properties such as drizzle bias. Decomposed latent perturbations can be applied to a pretrained model and tuned to produced physically realistic and calibrated models with far less compute than is required by diffusion methods or ensembles trained from scratch. Decomposing the transform across scales reveals what scales most influence ensemble spread. Interpolating in the latent space between two ensemble members produces additional ensemble members that can be further rolled out to understand the sensitivites of regime transitions. Further emerging work on loss functions and spline representations of vertical profiles will also be discussed.
Copyright. All Rights Reserved.