DFODE: Deep learning package for solving Flame chemical kinetics with high-dimensional stiff Ordinary Differential Equations

DFODE is an open-source deep learning package designed to accelerate computationally intensive reacting flow simulations by replacing conventional numerical integration of chemical source terms governed by high-dimensional stiff ordinary differential equations (ODEs).

Overview

DFODE provides:

• Efficient sampling module for extracting high-quality base data from low-dimensional manifolds in canonical flames
• Robust data augmentation strategy to expand training data for high-dimensional turbulent flames
• Neural network implementation with optimized data preprocessing and hyperparameter tuning
• Seamless interfaces for deploying trained models within the open-source CFD solver DeepFlame
• Physical constraints derived from conservation laws to ensure reliability in CFD applications

Features

• Low-dimensional manifold sampling from canonical flame configurations

  • 0D homogeneous reactors
  • 1D laminar premixed flames
  • 2D counterflow diffusion flames
  • 1D detonation tubes

• Data augmentation with physics-based constraints

  • Random perturbation of thermochemical states
  • Heat release and molar element ratio filtering
  • Mass conservation enforcement

• Deep neural network model

  • Multi-layer perceptron architecture
  • Box-Cox transformation for data preprocessing
  • Physics-informed loss functions
  • Support for both CPU and GPU training/inference

• Physical-aware correction

  • Elemental conservation enforcement
  • Energy conservation constraints
  • Heat release prediction error control

• CFD solver integration

  • Seamless interface with DeepFlame
  • Support for both CPU and GPU inference
  • Flexible deployment options

Installation

# Clone the repository
git clone https://github.com/DeepFlame-ML/DFODE.git

# Install dependencies
# Installing this package in editable mode is recommended
# in case users would like to experiment with different
# sampling schemes or make adjustments
pip install -e /path/to/your/DFODE/package

Usage

• For an example of using this package to sample data from low-dimensional flame simulations, please ensure that DeepFlame has been properly installed and refer to /your/path/to/DFODE/sampling_cases/oneDFlame.orig/case_setup.ipynb for instructions