Comprehensive deep learning for combustion chemistry integration: Multi-fuel generalization and a posteriori validation in reacting flow

The application of deep neural networks (DNNs) holds considerable promise as a substitute for the direct integration of combustion chemistry in reacting flow simulations. However, challenges persist in ensuring high precision and generalization across various fuels and flow conditions, particularly in a posteriori time-evolving flame simulations. This study performs comprehensive deep learning with multi-fuel generalization and computational fluid dynamics (CFD) validations. The process begins with generating thermochemical base states from low-dimensional canonical laminar flames to facilitate generalization and minimize the complexity of data generation. An effective perturbation and data augmentation strategy is then employed to broaden the coverage of the composition space for multi-dimensional flame configurations. Without the need for extensive tuning, three DNNs were consistently trained for three representative fuels: hydrogen, ethylene, and Jet-A. These DNN models were subsequently integrated into our recently developed open-source CFD package, DeepFlame (https://github.com/deepmodeling/deepflame-dev), for a posteriori reacting flow simulations and thoroughly validated against laminar flames and two representative turbulent premixed flames. The DNNs strongly agreed with the direct integration results across various combustion characteristics, including laminar and turbulent flame speeds, dynamic flame structures influenced by turbulence-chemistry interactions, and conditional scalar profiles. These findings underscore the exceptional accuracy and generalization capability of the employed deep learning approach. Moreover, by leveraging graphics processing units for model inference, the integration of DNN into CFD simulations resulted in significant speed-ups, achieving factors of approximately 72 for ethylene/air flames and 102 for Jet-A/air flames. The integrated DNN-CFD solver and test cases (https://www.aissquare.com) are openly shared, providing valuable tools to advance DNN development for chemical kinetics.