DFODE-Kit: Deep learning package for solving flame chemical kinetics with high-dimensional stiff ordinary differential equations
Description
Recent advances in deep learning for solving flame chemical kinetics offer promising solutions to the long-standing trade-off between accuracy and computational efficiency in combustion simulations. This work introduces DFODE-kit, an open-source Python package designed to replace the conventional, computationally intensive integration of chemical source terms governed by high-dimensional, stiff ordinary differential equations (ODEs), thereby substantially accelerating chemistry evaluation in combustion simulations. The package provides: i) an efficient sampling module that extracts high-quality thermochemical states from low-dimensional manifolds in canonical flames; ii) an effective data augmentation module that enriches the dataset to approximate the high-dimensional composition space encountered in turbulent flames; and (iii) an optimized neural network training module with multiscale preprocessing and physics-informed constraints to enhance model fidelity and stability. The trained models are seamlessly integrated into our previously released CFD solver DeepFlame, and can also be adapted for use with other widely used platforms such as OpenFOAM via custom interface modifications. Illustrative examples for a posteriori validations demonstrate that DFODE-kit models achieve excellent predictive accuracy. Furthermore, in isolated chemistry evaluations, the DNN models attain up to O(10^2) acceleration compared with CVODE, while end-to-end CFD runs typically see multi-fold speed-ups. The package, dataset, and example scripts are released to support reproducible benchmarking and community adoption.