Acceleration of combustion computation fluid dynamics simulations through machine learning

Jo, Jun Hyoung

Permalink

https://hdl.handle.net/2142/132783 Copy

Description

Title

Acceleration of combustion computation fluid dynamics simulations through machine learning

Author(s)

Jo, Jun Hyoung

Issue Date

2025-12-03

Director of Research (if dissertation) or Advisor (if thesis)

Lee, Tonghun

Department of Study

Mechanical Sci & Engineering

Discipline

Mechanical Engineering

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Deep Operator Networks
• Scientific machine learning
• Hydrogen combustion
• Combustion CFD
• Reacting flow CFD

Abstract

This work demonstrated the acceleration of combustion computational fluid dynamics simulations through Deep Operator Networks (DeepONets), achieving up to an 18 times speedup in the chemical source term evaluation. Another numerical experiment showed an overall reduction of approximately 30% in computation time. This was achieved by directly replacing the ODE solver for chemical reactions with cost-effective machine learning inference. Moreover, the proposed framework offered a practical and scalable method for combustion CFD simulations. The trained DeepONet models are integrated into an open-source CFD framework, OpenFOAM, using LibTorch, replacing the conventional chemistry and heat-release subroutines with inference results from DeepONets. Validation through quasi-zero-dimensional and one-dimensional hydrogen combustion cases shows that the proposed framework accurately predicts chemical species and temperature evolution computed by the baseline numerical solver, while delivering significant computational accelerations. Moreover, to enable efficient training of the machine learning models, a physics-informed approach is investigated using the UCSD hydrogen chemistry. Furthermore, a tailored DeepONet architecture is developed to accommodate various thermochemical variables as well as increase scalability for larger fuel mecahnisms. The model incorporates physics-informed loss functions derived from the governing chemical kinetics equations, allowing the models to learn the chemical kinetics with higher accuracy from fewer datasets compared to purely data-driven approaches. Furthermore, a sequential optimization procedure and specialized normalization techniques are introduced to handle the stiffness and wide dynamic range inherent in combustion chemistry. Overall, this study demonstrated that operator learning can provide an efficient, accurate, and CFD compatible surrogate for detailed chemical kinetics. Real CFD demonstrations were included to compare the accuracy and acceleration of the machine learning integrated CFD framework.

Graduation Semester

2025-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/132783

Copyright and License Information

Copyright 2025 Jun Jo

Owning Collections