Online Learning of Sub-Grid Stress Models for Large Eddy Simulation of Wall Bounded Turbulent Flows
DescriptionData-driven approaches for the development of sub-grid stress (SGS) closure models for large eddy simulation (LES) have been gaining popularity as they offer encouraging results for improved predictive capacity over traditional models. Due to the time and scale-resolving nature of LES, these models must be trained on instantaneous high-fidelity turbulent data. As we enter the age of exascale simulations, generating new databases of training data for traditional offline learning creates an I/O and storage bottleneck. This limitation is resolved by performing online (in situ) learning, wherein the ML model is trained concurrently with the flow simulation producing the training data. Moreover, online learning allows fine tuning of existing models to new problem classes on the fly. This talk will cover the software infrastructure developed to perform online learning at scale. We leveraged an open-source tool called SmartSim to deploy a database to store training data in-memory and enable two-way asynchronous communication between the simulation and model training applications running concurrently on an HPC cluster. In addition, we demonstrate online training of a SGS model on a number of flow problems and compare the model’s predictive capacity against other classical and offline trained SGS models.
TimeMonday, June 2615:30 - 16:00 CEST
Computer Science, Machine Learning, and Applied Mathematics