DeepM&Mnet for hypersonics: Predicting the coupled flow and finite-rate chemistry behind a normal shock using neural-network approximation of operators

Zhiping Mao, Lu Lu, Olaf Marxen, Tamer A. Zaki, George Em KarniadakisJournal of Computational Physics 2021 · OpenReview

TL;DR

This paper introduces DeepM&Mnet, a novel framework that leverages DeepONets to efficiently predict complex hypersonic flows with non-equilibrium chemistry. By approximating the underlying physics as learnable operators, DeepM&Mnet achieves significant speedups over traditional CFD methods and enables robust data assimilation from sparse sensor measurements.

What Problem Does This Paper Address?

Simulating hypersonic flows, particularly those involving shock waves and rapid chemical reactions (non-equilibrium chemistry), is computationally demanding. These flows exhibit coupled multiphysics and multiscale dynamics, making them challenging for conventional Computational Fluid Dynamics (CFD) solvers. The high temperatures behind a shock trigger dissociation reactions, altering gas composition and significantly impacting fluid properties like viscosity and heat conduction. Accurately capturing these phenomena requires solving complex, coupled equations, leading to excessive computational costs. Furthermore, integrating sparse experimental data into these simulations for accurate predictions, especially in data-poor regimes, remains a significant challenge.

Methods

The core innovation is the use of DeepONets (Deep Operator Networks). Unlike standard neural networks that approximate functions, DeepONets are designed to approximate operators. This means they can learn the mapping from an input function (e.g., a set of species densities) to an output function (e.g., velocity and temperature fields). The authors train individual DeepONets to predict specific fields based on others, effectively learning the coupled physics.

They then build upon this by proposing DeepM&Mnet, a composite framework that uses pre-trained DeepONets as building blocks. Two architectures are presented:

  1. Parallel DeepM&Mnet: Assumes data is available for all variables and uses a neural network to coordinate predictions from multiple DeepONets.
  2. Series DeepM&Mnet: Designed for scenarios where only sparse data (e.g., velocity and temperature) is available, using a sequential approach with DeepONets.

Both architectures incorporate mechanisms for data assimilation and can be enhanced with regularization and physical constraints like mass conservation.

Datasets

The authors generated training and testing datasets using a finite difference approach based on the governing equations for hypersonic flow with non-equilibrium chemistry. This involved simulating a benchmark case of a normal shock wave at Mach numbers between 8 and 10. The data consists of seven field variables: densities of five chemical species (N₂, O₂, N, O, NO), velocity (U), and temperature (T).

Key Insights and Ideas

  • Operators, Not Just Functions: The key insight is that complex physical systems can be effectively modeled by learning the operators that govern their behavior, rather than just approximating individual field values. DeepONets excel at this.
  • "Plug-and-Play" Physics: Pre-trained DeepONets act as modular components that encode specific physical phenomena. This allows for a flexible and efficient construction of complex multiphysics models.
  • Data Assimilation in Data-Poor Regimes: DeepM&Mnet offers a powerful way to integrate scarce sensor data into simulations, significantly improving prediction accuracy and enabling extrapolation beyond the training data range.
  • Speedup is Dramatic: The trained DeepONets are orders of magnitude faster than traditional CFD solvers, making them practical for real-time applications or large-scale simulations.
  • Physics-Informed Regularization: The pre-trained DeepONets inherently incorporate the underlying physics, acting as a form of regularization that stabilizes training and improves accuracy, even without explicit regularization terms in some cases.

Key Results

  • Speedup: Trained DeepONets achieve a speedup of approximately 100,000X compared to the CFD solver used for training data generation.
  • Accuracy: DeepONets provide accurate predictions for hypersonic flows with non-equilibrium chemistry within the training Mach number range.
  • Extrapolation: While initial extrapolation outside the training range is not perfect, combining a few sparse measurements with pre-trained DeepONets significantly improves accuracy.
  • DeepM&Mnet Performance: Both parallel and series DeepM&Mnet architectures demonstrate good accuracy in predicting all seven field variables, with the series architecture being particularly effective when only velocity and temperature data are available.
  • Global Mass Conservation: Incorporating the global mass conservation constraint demonstrably improves prediction accuracy and stabilizes the training process.

Limitations and Future Work

  • Extrapolation Challenges: While DeepM&Mnet improves extrapolation with sparse data, it's noted that initial predictions without data can be unsatisfactory. Further work is needed to robustly extrapolate "on the fly."
  • Convergence Guarantee: Similar to other inverse problems with nonlinear physics, convergence is not guaranteed for the NN training within DeepM&Mnet. Regularization terms are sometimes necessary to ensure stable training.
  • Generalizability: The authors suggest that the DeepM&Mnet framework is general and can be applied to other complex multiphysics and multiscale problems beyond hypersonics. Future work could explore its application in diverse domains.
  • Data Requirements: While DeepM&Mnet excels at data assimilation, the performance of extrapolation is still influenced by the number of available data points.

Summary updated

Abstract

Highlights • Efficient DeepONets are developed for the non-equilibrium chemistry at high Mach numbers. • We obtain good prediction when the Mach number is out of the input. • We develop a new inference tool, DeepMMnet for complex multiphysics and multiscale models. • DeepMMnets assimilate sparse data using pre-trained DeepONets in a “plug-and-play” mode. • The codes are compact and easy to adopt in diverse domains as well as maintain or enhance further. Abstract In high-speed flow past a normal shock, the fluid temperature rises rapidly triggering downstream chemical dissociation reactions. The chemical changes lead to appreciable changes in fluid properties, and these coupled multiphysics and the resulting multiscale dynamics are challenging to resolve numerically. Using conventional computational fluid dynamics (CFD) requires excessive computing cost. Here, we propose a totally new efficient approach, assuming that some sparse measurements of the state variables are available that can be seamlessly integrated in the simulation algorithm. We employ a special neural network for approximating nonlinear operators, the DeepONet [23], which is used to predict separately each individual field, given inputs from the rest of the fields of the coupled multiphysics system. We demonstrate the effectiveness of DeepONet for a benchmark hypersonic flow involving seven field variables. Specifically we predict five species in the non-equilibrium chemistry downstream of a normal shock at high Mach numbers as well as the velocity and temperature fields. We show that upon training, DeepONets can be over five orders of magnitude faster than the CFD solver employed to generate the training data and yield good accuracy for unseen Mach numbers within the range of training. Outside this range, DeepONet can still predict accurately and fast if a few sparse measurements are available. We then propose a composite supervised neural network, DeepM&Mnet, that uses multiple pre-trained DeepONets as building blocks and scattered measurements to infer the set of all seven fields in the entire domain of interest. Two DeepM&Mnet architectures are tested, and we demonstrate the accuracy and capacity for efficient data assimilation. DeepM&Mnet is simple and general: it can be employed to construct complex multiphysics and multiscale models and assimilate sparse measurements using pre-trained DeepONets in a “plug-and-play” mode.
BibTeX
@article{mao2021deepm,
  author = {Zhiping Mao and Lu Lu and Olaf Marxen and Tamer A. Zaki and George Em Karniadakis},
  title = {{DeepM\&Mnet for hypersonics: Predicting the coupled flow and finite-rate chemistry behind a normal shock using neural-network approximation of operators}},
  journal = {Journal of Computational Physics},
  year = {2021},
  url = {https://openreview.net/forum?id=BrxWmMSEG7q},
}
Figure