Researchers at Argonne National Laboratory, working with a collaborator at North Carolina State University, have now introduced a machine learning framework designed to break this bottleneck. The framework, called Phy-ChemNODE, is described in a study published on 15 August 2025 in Frontiers in Thermal Engineering. It combines a non-linear autoencoder with a neural ordinary differential equation, or NeuralODE, and trains both networks together in an end-to-end manner. The autoencoder compresses the full thermochemical state of the reacting system into a low-dimensional latent space, and the NeuralODE learns the temporal dynamics of the system directly in that compressed representation.

The approach builds on an earlier framework known as ChemNODE, in which a neural network replaces the expensive physics-based computation of chemical source terms. Rather than training the network offline and then coupling it to a numerical solver, ChemNODE uses an a-posteriori learning paradigm: the predicted source terms are passed to an ODE solver for time integration, and the network weights are optimized to minimize the loss between the predicted and ground truth thermochemical states. Because NeuralODEs learn continuous-time dynamics that can be integrated with existing ODE solvers, the predicted state remains adherent to the ground truth trajectory even over long time horizons, avoiding the error accumulation that plagues conventional offline-trained surrogates.

For large chemical mechanisms, however, evolving the full thermochemical state with a NeuralODE remains costly. Phy-ChemNODE addresses this by coupling a non-linear autoencoder to the NeuralODE so that the dynamics are learned in a reduced-order latent space. The training loss combines three mean absolute error terms: a reconstruction loss measuring how faithfully the encoder-decoder mapping reproduces the ground truth, a data loss measuring the accuracy of the full encoder-NODE-decoder prediction, and a latent loss comparing the NODE prediction in latent space with the encoder mapping of the ground truth. The reconstruction and latent terms ensure that the encoder and decoder mappings remain bijective, so every point in latent space corresponds to a unique physical state.

The most distinctive feature of the new framework is the direct incorporation of elemental mass conservation constraints into the loss function during training. Traditional black-box machine learning models of chemical kinetics do not inherently respect conservation laws, which can degrade simulation accuracy and hinder integration with multidimensional CFD solvers. Phy-ChemNODE adds penalty terms that compare the autoencoder-NODE predicted mass fractions of each element, such as carbon, hydrogen, and oxygen, with the ground truth values, weighted by atomic masses, the number of atoms of each element in each species, and molecular weights. This ensures that total mass and the mass of each elemental species are conserved in an a-posteriori manner throughout deployment.

For the proof-of-concept demonstration, the team considered an autoigniting methane-oxygen homogeneous reactor at a constant pressure of 20 atmospheres, corresponding to practical high-pressure isobaric operating conditions for stationary gas turbine engines. The detailed chemical mechanism consists of 32 species and 266 chemical reactions. Ground truth data was generated with the Cantera software toolkit, which solves the coupled ODE system with full detailed chemistry. The thermodynamic and composition space comprised 9 equispaced initial temperatures between 1600 and 2000 K and 11 equivalence ratios, yielding 99 initial conditions, each integrated to chemical equilibrium and downsampled to 200 time points. The data was randomly split by initial condition into training, validation, and test sets.

The network architecture was finalized through multiple ablation studies. Both the encoder and decoder have 5 hidden layers with 64 neurons each and Exponential Linear Unit activations, while the NODE has 4 hidden layers of 64 neurons to model the chemical source terms in latent space. Time integration during the forward pass uses a fourth-order explicit Runge-Kutta solver, and gradients are computed with backward adjoint automatic differentiation. The model was trained for 10,000 epochs using the ADAM optimizer with exponential learning rate decay, implemented in the Julia programming language with the Flux.jl library on two AMD EPYC 7713 64-core processors over a walltime of 96 hours.

Hyperparameter studies revealed that incorporating the elemental mass conservation soft constraints into the loss function results in lower validation loss for the same number of epochs, enabling more efficient training and faster convergence. In a-posteriori autoregressive inference tests, the trained model showed excellent agreement with ground truth for temperature and species mass fractions, including major species such as carbon monoxide and water and minor radicals such as hydroxyl. Scaled test set mean absolute errors for both major and minor species confirmed that Phy-ChemNODE captures the temporal dynamics with high fidelity across the training and test conditions.

The speedup results are striking. Running inference on an Intel i7-1165G7 workstation with 16 cores, Phy-ChemNODE yielded speedups of 6x with an implicit backward differentiation formula solver and 860x with an explicit RK45 solver relative to the full chemical mechanism in overall simulation walltime. The authors attribute this acceleration partly to a significant reduction in temporal stiffness: the evolution of the latent variables proved much smoother than that of the thermochemical scalars in physical space, mirroring earlier observations of stiffness reduction in autoencoder-NODE models of advection-dominated dynamical systems. The model trained with mass conservation constraints also conserved carbon, hydrogen, and oxygen elemental mass fractions far better during deployment than an unconstrained counterpart.

The researchers outline several avenues for extending the framework, including scaling the training workflow to wider ranges of thermodynamic conditions with multiple pressures, larger kinetic mechanisms, and multicomponent fuels; integrating Phy-ChemNODE with CFD solvers for accelerated multidimensional reacting flow simulations on high-performance computing platforms; demonstrating the approach for constant-volume combustion; incorporating uncertainty quantification; and exploring training methodologies to enhance out-of-distribution generalizability. If those extensions succeed, physics-constrained neural ODE surrogates could bring realistic fuel chemistry within reach of routine engine and turbine simulations, transforming how energy conversion systems are designed and optimized.

Subject of Research: machine learning combustion kinetics

Article Title: Phy-ChemNODE: an end-to-end physics-constrained autoencoder-NeuralODE framework for learning stiff chemical kinetics of hydrocarbon fuels

Article References: Kumar, T., Kumar, A., & Pal, P. (2025). Phy-ChemNODE: an end-to-end physics-constrained autoencoder-NeuralODE framework for learning stiff chemical kinetics of hydrocarbon fuels. Frontiers in Thermal Engineering, 5, Article 1594443. https://doi.org/10.3389/fther.2025.1594443

Image Credits: AI Generated

DOI: 10.3389/fther.2025.1594443

Keywords: chemical kinetics, combustion, machine learning, NeuralODE, autoencoder, physics-informed neural network, CFD, methane, Phy-ChemNODE, end-to-end, physics-constrained, autoencoder-NeuralODE