Accelerating Design of ML and AI Experiments in Scientific Simulation

Across industry and scientific fields, Artificial Intelligence (AI) and Machine Learning (ML) have ushered in a sea change in how engineers and scientists interact with their designs and tools. This is felt especially in the field of computational science, where simulations sit on the same silicon as these “thinking” algorithms. In the field of computational fluid dynamics, current Reynolds Averaged Navier Stokes (RANS) closure models typically cannot properly model all regimes of a given flow state. However, it has now been shown that by incorporating high fidelity simulation data into the closure model through machine learning, these closure models can be significantly improved. Furthermore, data-driven models have shown promise of some extrapolatory power, meaning that improvement can potentially be achieved even outside the narrow configuration of their training dataset. That being said, significantly more work is required to understand the interplay between the data-driven model, the simulation, and expectations of robustness and accuracy as the model's application is expanded into increasingly diverse scenarios. This problem is characteristic of many challenges faced by those looking to adopt data-driven techniques into their solvers. As such, in this Phase II effort, we are proposing to significantly advance our in situ AI toolkit demonstrated in Phase I. We plan to enable powerful new capabilities for Catalyst to serve AI and ML technologies more efficiently. Further, we will conduct exciting research into adjoint-driven model sensitivity analysis, which will provide powerful insights into how a data-driven model impacts a simulation state, and how it can be improved given quantities of interest directly relevant to the simulation.