WombatWisdom & WombatWiser: MHD Simulations with in-situ ML Training

WOMBAT is a computational MHD fluid dynamic code primarily utilized to study AGN jets. WombatWisdom is a forthcoming reworked polyglot agent-based version that uses a directed graph to explicitly handle program execution. The new design will better leverage HPC systems by decomposing the simulation into many subprocesses, and is being developed in partnership with HPE as an application of their new Dragon distributed runtime. This architecture is broadly applicable, and is in development in the EXCLAIM weather simulation rework.

Key to the redesign is the use of shared memory architectures, which allows the simulation’s output to be exposed to other processes as it is being produced. WombatWiser will encompass a set of in-situ machine learning (ML) codes that will have direct, concurrent read access to the WombatWisdom outputs. There are a number of applications we intend to develop, including structure and shock classifiers, quality monitoring with the ability to terminate poor runs early, and neural network training to generate artificial observations or additional, artificial, simulations. Here we present the directed graph structure and initial test results of WombatWisdom and the first results of ML shock identification from WombatWiser, trained on outputs from the original WOMBAT.