Entropy-Constrained Machine Learning with Residual Data Augmentation for Modeling Chemical Kinetics

Jul 16, 2026

A physics-constrained machine learning framework accelerates direct numerical simulation of turbulent reacting flows by replacing detailed chemical source-term evaluation with a surrogate predicting reaction rates from a reduced thermochemical state. Thermodynamic consistency is enforced by incorporating the second law of thermodynamics through a non-negative entropy-generation constraint, restricting thermochemical state evolution to physically admissible directions and improving time-integration stability. Validated on DNS of a 2D planar lean premixed methane–air flame in turbulence, the method matches detailed-chemistry fidelity while reducing computational cost by over an order of magnitude and enables parametric inlet exploration using residual-based synthetic data augmentation.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

Introduction

Model Formulation and Setup

Hybrid Neural Network-Kernel Regression Framework

arXiv :2607 .09582v1 [physics .flu-dyn] 10 Jul 2026 ENTROPY-CONSTRAINED MACHINE LEARNING WITH RESIDUAL DATA AUGMENTATION FOR MODELING CHEMICAL KINETICS Okezzi Ukorigho, Opeoluwa Owoyele∗ Department of Mechanical and Industrial Engineering Louisiana State University, Baton Rouge, LA 70810, USA ABSTRACT We present a physics-constrained machine learning framework for accelerating the direct numerical simulation (DNS) of turbulent reacting flows. The model replaces the direct evaluation of detailed chemical source terms with a surrogate that predicts reaction rates from a reduced thermochemical state. To improve physical consistency, the second law of thermodynamics is incorporated asa training constraint by enforcing non-negative entropy generation, which restricts the evolution of the thermochemical state to physically admissible directions and improves stability during time integration. The approach is demonstrated on DNS of a two-dimensional planar lean premixed methane-air flame interacting with a turbulent flow field. The model reproduces detailed-chemistry results with high fidelity while achieving more than an order-of-magnitude reduction in computational cost. Furthermore, a residual-based synthetic data augmentation strategy enables parametric exploration by constructing new training data from the original dataset, allowing accurate simulation at new inlet conditions without additional detailed-chemistry CFD runs. These results demonstrate that thermodynamically constrained machine learning can provide reliable and computationally efficient surrogates for detailed chemistry in high-fidelity combustion simulations. Keywords turbulent reacting flows · chemical kinetics · direct numerical simulation · entropy-constrained machine learning · second-law constraints · data augmentation 1 Introduction Direct Numerical Simulation (DNS) of reacting flows remains an important tool for improving our understanding of turbulence–chemistry interactions, but its application to practical fuels and geometries remains limited by the high cost of resolving detailed reacting-flow physics [1–3] . As a result, several approaches have been developed to accelerate chemistry calculations. These include skeletal and reduced mechanism generation techniques [4–6], as well as manifold-based approaches such as flamelet-generated manifolds [7, 8], intrinsic low-dimensional manifolds [9, 10], and data-driven manifold reduction methods that project the thermochemical state space onto lower-dimensional manifolds [11–14] . Over the past decade, machine learning (ML) has gained traction as a tool for accelerating chemical source-term evaluations. Early work by Christo et al. [15] demonstrated neural-network-based prediction of chemical source terms, followed by Blasco et al. [16], who modeled the temporal evolution of chemically reacting systems. Subsequent studies have expanded these ideas, including micro-mixing–based data generation strategies for ML training [17], Arrhenius-and law-of-mass-action-inspired neural network architectures [18], neural ordinary differential equations for kinetics integration [19, 20], and more recently, operator-learning approaches for accelerating combustion chemistry [21, 22] . Concurrently, a growing body of work has sought to incorporate physical constraints into ML-based source-term prediction, enforcing conservation of mass or chemical elements either as soft constraints through loss-function regularization [23–26] or as hard constraints through post-prediction correction layers [27, 28] . These efforts are valuable ∗ Corresponding author: [oowoyele@lsu.edu](oowoyele@lsu.edu) Entropy-Constrained Machine Learning with Residual Data Augmentation because purely data-driven models often do not automatically encode such laws, despite their fundamental importance and their foundation in centuries of accumulated scientific knowledge. This work is positioned within the same broader effort, which seek

FAQs

What is the main goal of the proposed machine learning approach?

To accelerate direct numerical simulation of turbulent reacting flows by learning a surrogate for chemical source terms that predicts reaction rates from a reduced thermochemical state.

How does the method enforce physical consistency during training?

It incorporates the second law of thermodynamics by enforcing non-negative entropy generation, which constrains thermochemical state evolution to physically admissible directions and improves stability.

How is residual-based synthetic data augmentation used?

It constructs new training data from the original dataset, enabling accurate simulations at new inlet conditions without additional detailed-chemistry CFD runs.

Jul 16, 2026

Upload by Valentina