Physics-informed neural surrogates for next-generation aerothermochemical modeling

Zanardi, Ivan

Permalink

https://hdl.handle.net/2142/132477 Copy

Description

Title

Physics-informed neural surrogates for next-generation aerothermochemical modeling

Author(s)

Zanardi, Ivan

Issue Date

2025-11-03

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

Panesi, Marco

Doctoral Committee Chair(s)

Panesi, Marco

Committee Member(s)

• Bodony, Daniel J.
• Alam, Syed B.
• Duraisamy, Karthik

Department of Study

Aerospace Engineering

Discipline

Aerospace Engineering

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Thermochemical Nonequilibrium
• Plasma
• Hypersonic
• Surrogate Modeling
• Machine Learning

Abstract

Thermochemical nonequilibrium is a fundamental physical process in a wide range of reactive flow environments, spanning planetary entry and astrophysical flows to plasma-assisted combustion and flow control. State-to-state (StS) collisional–radiative (CR) models provide the highest level of physical fidelity by explicitly resolving the species population dynamics of individual quantum energy levels of each species. However, the dimensionality of these models increases exponentially with the number of resolved states, and their inherent numerical stiffness renders them prohibitively expensive for large-scale simulations. To address these challenges, this thesis proposes a modular, three-stage, physics-constrained, data-driven framework that preserves the fidelity of detailed state-specific CR kinetics while achieving the computational efficiency required for practical aerospace applications. The first stage reduces the dimensionality of the problem via an encoding–decoding procedure built on a Petrov–Galerkin reduced-order model (ROM). The ROM obliquely projects the full-order dynamics onto a low-dimensional subspace using adjoint sensitivities, producing accurate reduced representations without the empirical or physical assumptions common in conventional thermochemistry ROMs. The second stage addresses numerical stiffness with two novel neural-operator-based surrogate models that replace the expensive implicit integration of kinetic systems. By embedding physical laws and governing equations directly into their architectures, these machine-learning-based surrogates achieve greater reliability, accuracy, and generalization than many existing approaches, in which physics constraints are often absent or only weakly enforced. The final stage couples the surrogates with CFD solvers via operator splitting, enabling efficient, high-fidelity simulations of realistic plasma flow configurations. Across benchmark problems ranging from zero-dimensional reactors to multidimensional plasma flows, the proposed ROM achieves up to seven orders of magnitude reduction in floating-point operations compared to the full-order StS model, while keeping errors below 1% for macroscopic quantities and under 10% for microscopic state populations. The neural-operator-based surrogate models further accelerate the solution of the (reduced) system of equations, replacing conventional implicit integration and providing an additional one to three orders of magnitude speedup, while maintaining accuracy within a few percent of the reference solution even in strongly nonequilibrium and transient regimes. Remarkably, these surrogates also maintain acceptable accuracy in extrapolative scenarios beyond the training domain, a capability that stems directly from the deep integration of physical constraints into their design. These results highlight the potential of physics-constrained data-driven modeling to make high-fidelity nonequilibrium simulations tractable for next-generation aerospace applications.

Graduation Semester

2025-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/132477

Copyright and License Information

Copyright 2025 Ivan Zanardi

Owning Collections