Published: Aug 12, 2021 by Daning Huang
From Aug. 10-12 the Lawrence Livermore National Laboratory (LLNL) hosted the Machine Learning for Industry (ML4I) Forum, and Dr. Huang gave a talk in the Physics-Constrained Machine Learning session, titled “Physics-Infused Differential-Algebraic Reduced-Order Models for Multi-Disciplinary Systems”
Abstract: The multi-disciplinary design, analysis and optimization is becoming indispensable for modern high-performance engineering systems. However, this class of problem usually comes with an extremely high computational cost, even for only the coupled analysis. Reduced-order models (ROMs) and surrogates are typical approaches to reducing the computational cost to a tractable level. However, the existing ROMs and surrogates suffer from the curse of dimensionality that roots from the need to parameterize and sample the system configuration parameters. This work presents our recent progress in physics-fused differential-algebraic ROM, where first-principle-based analytical models are augmented with a data-driven physically-interpretable component. A demonstration is presented for hypersonic aerothermal load prediction, a key ingredient for the aerothermoelastic analysis and optimization of the structures of hypersonic vehicles. The new ROM technique can achieve an accuracy close to that of CFD solvers when predicting the flow solutions over a wide range of complex surface deformations with a limited number of high-fidelity solutions. These results underline its potential to be used as a new generation of ROM for a broader engineering applications for which classical analytical models are available.
The slides are available here. The implementation is mostly done by our lab member Carlos.
