Abstract:Recently, artificial neural networks (ANNs) have been frequently embedded in computational fluid dynamics (CFD) solvers as surrogate tools for solving chemical reaction kinetics, thereby accelerating the computation of chemical reaction source terms. Compared with other chemical acceleration methods, such as mechanism reduction and tabulation, ANNs have the potential to simultaneously save computational cost and preserve mechanisms' high fidelity. However, conventional data-driven ANNs highly rely on the training data for solving combustion reaction kinetics, and non-physical prediction errors can hardly be avoided. This paper proposes a physics-informed neural network for solving combustion reaction kinetics (CRK-PINN), which regularizes the ANN through physical principles, including the laws of mass action, Arrhenius, enthalpy conservation, element conservation, and mole fraction conservation. In the absence of training data, CRK-PINN can independently solve chemical reaction kinetics, reconstruct the combustion process, and infer intermediate species and temperature. Under the supervision of training data and physical principles, CRK-PINN significantly suppresses the non-physical errors of data-driven ANNs and presents lower data dependence. This is demonstrated by the 0-D autoignition prediction and direct numerical simulation (DNS) of laminar and turbulent flames. Compared with the direct integration method, CRK-PINN results in an acceleration of 6.0 ∼ 14.6 times for solving a 10-species 21-step H2 -Air chemical mechanism. This further leads to a 2.3 ∼ 4.9 times speedup for simulations of reacting flows with satisfactory accuracy based on the OpenFOAM toolbox. Novelty and significance statement We propose a novel Combustion Reaction Kinetics Physics-Informed Neural Network (CRK-PINN) for solving detailed combustion reaction kinetics with features of multiscale and high stiffness. Soft constraints of physical losses have been developed in PINNs to solve ordinary differential equations (ODEs). Data pre-processing methods such as logarithm and normalization are commonly used to alleviate the multiscale issue. However, these techniques have not been effectively integrated to properly address detailed combustion reaction kinetics. Our CRK-PINN incorporates logarithmic-normalized combustion reaction kinetics ODEs and several conservation equations as physical constraints into a neural network, highlighting both technical challenges and our novelty. By incorporating comprehensive physical constraints, CRK-PINN can independently solves detailed combustion reaction kinetics without relying on training data. Compared with data-driven ANNs, CRK-PINN offers a surrogate model with improved physical completeness, error reduction, and low data dependence. Compared with the direct integration method, CRK-PINN provides satisfactory accuracy and evident computational acceleration.

Stiffness-Reduced Neural ODE Models for Data-Driven Reduced-Order Modeling of Combustion Chemical Kinetics

Kinetics Parameter Optimization via Neural Ordinary Differential Equations

Kinetics parameter optimization of hydrocarbon fuels via neural ordinary differential equations

Artificial Neural Network Based Chemical Mechanisms for Computationally Efficient Modeling of Hydrogen/carbon Monoxide/kerosene Combustion

Neural Ordinary Differential Equations for Model Order Reduction of Stiff Systems

A Posteriori Evaluation of a Physics-Constrained Neural Ordinary Differential Equations Approach Coupled with CFD Solver for Modeling Stiff Chemical Kinetics

An Efficient Data-Driven Approach for Modeling and Optimization of Reactivity-Controlled Compression Ignition Engine Fueled with Polyoxymethylene Dimethyl Ethers and Hydrogen

Dynamic Stiffness Removal for Direct Numerical Simulations

Embedding Physics into Neural ODEs to learn Kinetics from Integral Reactors

Comparative Analysis of Detailed and Reduced Kinetic Models for CH4 + H2 Combustion

A Deep Learning-Based ODE Solver for Chemical Kinetics.

Learning kinetics from non-ideal reactors by implicitly solved finite volumes and global reaction neural networks

Reduced-order Model for Fluid Flows via Neural Ordinary Differential Equations

Data-driven reduced-order models via regularized operator inference for a single-injector combustion process

Data-Driven Tabulation for Chemistry Integration Using Recurrent Neural Networks

CRK-PINN: A physics-informed neural network for solving combustion reaction kinetics ordinary differential equations

Learning stiff chemical kinetics using extended deep neural operators

Dynamic Adaptive Chemistry with Mechanisms Tabulation and in Situ Adaptive Tabulation (ISAT) for Computationally Efficient Modeling of Turbulent Combustion

Projection-based Reduced Order Modeling of Multi-Species Mixing and Combustion

An Improved Approach Towards More Robust Deep Learning Models for Chemical Kinetics

Neural Differential Equations for Inverse Modeling in Model Combustors