Keywords

Physics-Based Surrogate Models, Data-Driven Subgrid-Scale Closure Models, Physics-Informed Neural Networks, Turbulence-Chemistry Interactions, Machine Learning Framework for Thermo-Fluids

Abstract

The design of next-generation combustion systems relies on computer simulations that capture the multidimensional and multiscale physics of turbulent reacting flows. These simulations are computationally intensive, requiring trade-offs to produce results within allowable time and cost constraints. These trade-offs often consist of general assumptions and modeling simplifications that approximate complex physical phenomena. Consequently, these models can limit both the accuracy and the regime of applicability to real systems since these approximations are often heuristic and inconsistent with the governing physics. Moreover, in today’s rapid design-to-prototype development phase, making timely and informed design decisions is important. This requires fast and reliable computer simulations, to ensure that development can be accelerated while still safeguarding the physical accuracy of the results.

This study demonstrates how Artificial Intelligence can help overcome some of the challenges associated with the computational cost of reacting flows simulations. This work shows how these data-driven approaches can learn the complex behavior of flames to dramatically speed up simulations, thereby saving valuable time and computational resources. By accelerating simulations without sacrificing predictive fidelity, these surrogate combustion models can help overcome traditional time and resource bottlenecks that hinder analysis and design evaluations early in the project life-cycle.

However, reducing computational cost is only part of the challenge when simulating reacting flow systems. It is equally important to improve the assumptions and underlying physical models that often oversimplify turbulent combustion phenomena, particularly turbulence-chemistry interactions.

To enable the development of reliable and robust machine learning surrogates, this work interrogates a high-fidelity Direct Numerical Simulation (DNS) dataset of a turbulent premixed methane jet. This ultra-high-resolution "digital microscope" serves as a ground-truth to examine the fundamental parameters that describe and represent turbulence-chemistry interactions.

Building upon these insights, a Random Forest Machine Learning model was developed, trained, and tested using the DNS dataset to predict the subgrid-scale (SGS) scalar dissipation rate. The model’s development was guided by a physics-informed feature set. Results show that the machine learning-based model outperformed traditional algebraic LES closures for the SGS scalar dissipation rate. This demonstrates that machine learning models are highly effective for replacing and improving upon existing heuristic-based models in capturing the SGS scalar dissipation rate behavior.

Completion Date

2025

Semester

Fall

Committee Chair

Vasu, Subith

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Department of Mechanical and Aerospace Engineering

Format

PDF

Identifier

DP0029714

Document Type

Thesis

Campus Location

Orlando (Main) Campus

Download

Share

COinS