Hybrid Inference on Heterogeneous Architectures for ANN Chemistry Surrogate Modeling in Reactive CFD
Please login to view abstract download link
Reactive CFD simulations can become prohibitively expensive even for small or moderately sized chemistry mechanisms due to the stiffness of the chemistry caused by the large range of timescales involved in the chemical reactions. Artificial neural network (ANN) based chemistry surrogate modeling offers promising avenues for reducing the computational cost. However, the efficient integration of these models into existing CFD codes and their use in HPC applications remains underexplored. This works systematically studies the computational cost reduction potential obtained with ANN-based chemistry surrogate models on heterogeneous compute architectures. A variation of thermodiffusively unstable H2 flames solved with detailed chemistry integration (large computational costs) serve as baseline configurations.
