Deep learning based combustion chemistry acceleration method for widely applicable NH3/H2 turbulent combustion simulations

Abstract

Simulating reacting flows with detailed chemistry is often prohibitively expensive due to the complexity of reaction mechanisms and the numerical stiffness arising from disparate chemical time scales. While recent advancements in neural networks offer potential for efficiently capturing the dynamics of stiff chemistry, its application to dual-fuels with drastic differences in reactivity such as ammonia (${\mathrm{NH}}_3$) and hydrogen ($H_2$) remains challenging. In this study, we present a neural network model with variable time steps aimed at enhancing the efficiency of combustion chemistry simulations focusing on the complex dual-fuel ${\mathrm{NH}}_3/H_2$ under premixed combustion. We improved the ”sampling-training” workflow based on previous HFRD method to overcome the challenge of generalizing neural network models to fuel blends under premixed combustion. This workflow involves three improvements: defining the base manifold using unity Lewis number laminar flames, introducing continuously controllable randomization, and employing a training process with mass conservation and heat release rate similarity constraints. Our approach is validated against simulations of planar turbulent premixed flames and temporally-evolving jet flames across various conditions. The model demonstrates high accuracy and consistency, achieving a chemical calculation acceleration of 7 times and an overall simulation acceleration of 5 times using a model with 4 hidden layers and 800 neurons on the same CPU device. When a GPU is adopted, the chemical calculation acceleration increases to 30 times, and the overall simulation acceleration reaches 10 times.

Novelty and Significance Statement

Utilizing detailed chemistry in reacting flow simulations drastically increases computational cost due to numerical stiffness and disparate time scales. A promising approach is to replace the time-consuming ODE solvers with compact neural networks. Despite the rapid development of the neural network approach for accelerating combustion kinetics calculations, the application of this concept to fuel blends with varying mixing ratios and reactivities remains insufficient, particularly in turbulent premixed flames. In this study, we improved the neural network framework that could predict the kinetics of fuel blends of low-reactivity fuel ${\mathrm{NH}}_3$ and high-reactivity fuel $H_2$, highlighting the applications to binary fuel with large reactivity differences. Specifically, the unity Lewis number laminar flames are leveraged as an economical thermochemical base manifold, given their close resemblance to turbulent flame profiles. A continuously controllable randomization method is introduced to balance model capacity and computational efficiency by adjusting key parameters. Additionally, a loss function with mass conservation and heat release rate similarity constraints ensures stable long-term predictions. The well-trained neural network model was coupled with CFD codes to simulate two challenging cases across a wide range of turbulent intensities and fuel compositions. The results show visually identical scalar fields and highly accurate statistical outcomes, even under intense turbulence and after O($10^4$) neural network model calls, with a O(10) acceleration in computation.