A physics-informed machine learning approach for predicting dynamic behavior of reacting flows with application to hydrogen jet flames

Abstract

Traditional data-driven modeling approaches suffer from large error accumulation over time, divergence from expected physical behavior, and poor generalizability for out-of-distribution samples. To address this, we present Physics-informed hybrid Multiscale and Partitioned Network (PiMAPNet), a physics-informed machine learning (ML) strategy for generating multi-scale and multi-physics predictions by integrating low-resolution physics-based models with neural networks. Motivated by prior work on hybrid methods that combine coarse-grain simulations with ML, PiMAPNet employs state-space decomposition on the hydrodynamic (velocity and pressure) and thermochemical (temperature and species mass fractions) quantities for improved predictions of multiphysical processes. In addition, the ML model utilizes a mixture-of-experts (MoE) architecture that partitions the thermochemical state-space and employs a separate ML model to represent a specific region within this partition. We demonstrate this ML framework on a reacting hydrogen/air jet flame configuration. Results demonstrate that both the purely data-driven ML model and a traditional PIML approach could not represent the entire state-space, which resulted in unphysical behavior in long-term predictions. In contrast, the MoE-based PiMAPNet achieves higher accuracy and demonstrates improved robustness over extended time windows and out-of-distribution scenarios. Through our analysis, we show that PiMAPNet offers faster inference speed than a numerical simulation with comparable accuracies in multiple physical quantities.

Novelty and Significance Statement

This study introduces a novel physics-informed machine learning framework that enhances the predictive accuracy for chemically reacting flows by integrating low-resolution physics-based models with neural networks. The novelty of the framework lies in its specialized treatments for hydrodynamic and thermochemical variables. Additionally, the thermochemical state-space is partitioned to effectively capture the evolution of different regions within the state-space. The significance of our work is its ability to deliver highly accurate and robust predictions over extended time periods and for out-of-distribution scenarios. Furthermore, the separate treatment of different physical processes enables this framework to be extendable to other multi-physics systems, such as plasma physics or multiphase flows, making it a valuable tool for researchers across various domains in computational physics.