Tutorials

This directory provides step-by-step tutorials for using DFODE-kit to develop and deploy neural network models for accelerating combustion kinetics simulations. The tutorials guide you through the complete pipeline: from data sampling and augmentation to model training and testing. We provide two example cases to demonstrate the workflow for different flame configurations.

Pipeline Overview

DFODE-kit's workflow consists of the following stages:

1. Data Sampling: Extract thermochemical states from canonical flame simulations using low-dimensional manifold sampling. This ensures coverage of high-dimensional composition spaces efficiently.

2. Data Augmentation and Labeling: Enrich datasets with physics-constrained perturbations to approximate turbulent conditions, then generate supervised labels using Cantera's CVODE solver.

3. Model Training: Train a neural network (e.g., MLP with GELU activations) to predict state changes, incorporating constraints for mass/energy conservation.

4. Model Testing: Evaluate the model via a priori (single-step predictions) and a posteriori (full CFD simulations) validations.

5. Model Deployment: Integrate the trained model into CFD solvers like DeepFlame for accelerated chemistry integration.

Each tutorial includes Jupyter notebooks and scripts to walk you through these stages.

Available Tutorials

1. One-Dimensional Freely Propagating Flame (oneD_freely_propagating_flame/)

This tutorial demonstrates the pipeline for a 1D laminar premixed hydrogen/air flame under premixed conditions, validating model accuracy in reproducing flame propagation behavior. This example samples data from a single 1D flame simulation and performs a posteriori validation on the same simulation case.

• 1_sample_train/: Covers sampling, augmentation, labeling, and training.

  • dfode_kit_init.ipynb: Initialize simulation parameters (e.g., equivalence ratio, temperature, pressure) via config_dict.
  • Allrun: Execute the DeepFlame simulation to generate canonical flame data.
  • dfode_kit_tutorial.ipynb: Comprehensive Jupyter notebook providing a step-by-step guide through the entire pipeline, including code execution, explanations of each stage, and instructions on both python interface and command line utilities.

• 2_model_test/: Test the trained model.

  • priori/: A priori testing via single-step predictions on labeled datasets to evaluate model accuracy against ground-truth ODE solutions from Cantera CVODE.
  • posteriori/: A posteriori validation by integrating the trained model into full CFD simulations, comparing flame propagation and structure against direct CVODE integration.

2. Two-Dimensional HIT Flame (twoD_HIT_flame/)

This tutorial evaluates the DNN model in a 2D propagating premixed hydrogen/air flame within homogeneous isotropic turbulence (HIT), assessing turbulence-chemistry interactions beyond laminar regimes. The setup follows configurations for capturing flame front wrinkling and turbulent burning velocities. This example samples data from multiple 1D flame simulations and performs a posteriori validation on a separate 2D HIT case.

• 1_sample_train/: Similar pipeline as above, adapted for 2D HIT.

  • Includes initialization, simulation, sampling, augmentation, labeling, and training for HIT conditions.
  • dfode_kit_tutorial.ipynb: Comprehensive guide.

• 2_model_test/: Model testing in turbulent scenarios.

  • priori/: Single-step predictions.
  • posteriori/: Full turbulent flame simulations.

Getting Started

To get started with the tutorials, refer to dfode_kit_tutorial.ipynb in each 1_sample_train/ subdirectory for step-by-step guidance through the sampling, augmentation, labeling, and training steps.

Note that running the simulations requires DeepFlame to be installed. Refer to the DeepFlame GitHub repository and documentation for installation instructions.