Revision as of 10:13, 1 July 2024

Abstract

Multi-physics scientific codes, like those used in weather prediction and spacecraft simulation, often involve the coupling of many subdomain models. The resulting models are generally expensive to run. These costs, along with the model’s complexity, make downstream tasks like model calibration and uncertainty quantification especially difficult. In this work, we simplify structure in multi-physics models by learning an undirected graphical model corresponding to the system state variables, where the edges in the learned graph represent conditional dependence between variables. Depending on the application of interest, the resulting graph may (1) reveal the probabilistic structure of the joint distribution; (2) identify the most important coupling variables (those that must be shared between subdomain models), and those which may be safely neglected; (3) identify candidate variables for first-pass verification tasks; or (4) decouple parts of a model, while maintaining accuracy in the model predictions. We illustrate these possibilities through two multi-physics numerical models, the Multiple Prediction Across Scales-Atmosphere (MPAS-A) code base and a fire detection satellite model

Full Paper