Physics and Uncertainty in the Community Research Earth Digital Intelligence Twin

Explore the development and implementation of physics-constrained AI models for Earth system prediction in this 49-minute conference presentation from IPAM's Mathematics and Machine Learning for Earth System Simulation Workshop. Learn about the Community Research Earth Digital Intelligence Twin (CREDIT), an open foundational research platform developed by the National Center for Atmospheric Research for creating AI-based Earth system prediction models. Discover how global physics constraints on mass and energy applied during training and inference enhance prediction accuracy, enable stable emulator rollouts, and correct biases in training data such as drizzle bias. Examine decomposed latent perturbations that can be applied to pretrained models to generate physically realistic and calibrated ensemble models with significantly reduced computational requirements compared to diffusion methods or training ensembles from scratch. Understand how decomposing transforms across different scales reveals which scales most influence ensemble spread and how interpolating in latent space between ensemble members creates additional members for analyzing regime transition sensitivities. Gain insights into emerging research on loss functions and spline representations of vertical profiles that advance the field of AI-driven Earth system modeling.