Surrogate modeling of parameterized multi-dimensional premixed combustion with physics-informed neural networks for rapid exploration of design space

Abstract

Parametric optimization is a critical component in designing and prototyping combustion systems. However, existing parametric optimization methods often suffer from either insufficient prediction accuracy or computationally prohibitive data generation costs. This study develops and validates an efficient and robust field-resolving surrogate modeling framework for combustion systems based on physics-informed neural networks (PINNs). This study comprises three progressive steps. (1) We propose a PINNs framework for premixed combustion and validate its accuracy with direct numerical simulation (DNS) for one-dimensional and two-dimensional canonical premixed combustion problems. Results show a robust predictive capability of PINNs with R2 generally above 0.99. (2) We continue to develop a field-resolving surrogate modeling framework for parameterized two-dimensional methane-air jet combustion system with five input parameters: inlet equivalence ratio, temperature, velocity, diameter, and coflow temperature. The PINNS surrogate model after unsupervised training is used to predict the combustion fields for 32 input parameter combinations, and the results are compared against corresponding numerical simulations. The surrogate model accurately predicts the temperature fields with R2 over 0.95, except in case of flashback, where the accuracy is slightly lower. Good consistency between PINNs and numerical simulations is also demonstrated for four combustion performance metrics: flame length, maximum temperature, outlet temperature, and outlet CO2 mass fraction. (3) Based on the validated field-resolving PINNs surrogate model, sensitivity analysis of the parameterized combustion system is performed, and its implications for engineering applications are discussed. Most notably, distinct from common data-driven and machine learning approaches, the proposed PINNs models are entirely physics-driven and do not use any training datasets, saving tremendous amount of time in the design of experiments (DOE) and case-by-case simulations. Overall, the results of this study demonstrate PINNs can be an efficient and robust surrogate modeling approach for parameterized combustion system visualization, design, optimization, and active control.