DeepFlame: A deep learning empowered open-source platform for reacting flow simulations

Recent developments in deep learning have brought many inspirations for the scientific computing community and it is perceived as a promising method accelerating computationally demanding reacting flow simulations. In this work, we introduce DeepFlame, an open-source C++ platform with capabilities of utilising machine algorithms pre-trained models to solve reactive flows. We combine individual strengths computational fluid dynamics library OpenFOAM, framework Torch, chemical kinetics program Cantera. The complexity cross-library function data interfacing (the core DeepFlame) minimised achieve simple clear workflow code maintenance, extension upgrading. As demonstration, apply our recent work on predicting (Zhang et al. Combust. Flame vol. 245 pp. 112319, 2022) highlight potential simulation. A thorough validation conducted via broad range canonical cases assess its accuracy efficiency. results demonstrate that convection-diffusion-reaction implemented DeepFlame are robust accurate both steady-state transient processes. addition, number methods aiming further improve efficiency, e.g. dynamic load balancing adaptive mesh refinement, explored. Their performances also evaluated reported. With speed-up two orders magnitude achieved hydrogen ignition case when performed medium-end graphics processing unit (GPU). Further gain efficiency expected hydrocarbon other complex fuels. similar level acceleration obtained AI-specific chip – (DCU), highlighting leveraging next-generation architecture hardware. Program Title: CPC Library link files: https://doi.org/10.17632/3pg9xmypp3.1 Developer's repository link: https://github.com/deepmodeling/deepflame-dev Licensing provisions: GPLv3 Programming language: Nature problem: Solving chemically flows direct (quasi-direct) simulation usually troubled by following problems: 1. widely-used (CFD) toolbox, OpenFOAM features poor ODE solvers chemistry oversimplified transport models, yielding non-negligible errors results; 2. source term evaluation most expensive accounts more than 80% total time. Solution method: An bringing together Cantera PyTorch libraries built study. present implementation, CVODE solvers, detailed all adopted assist flow. Note here introduced combination heterogeneous accelerate solving procedure evaluation.