FPGA implementation of an optimized neural network for CFD acceleration

In this work, an evaluation of FPGAs as the central computation platform in domain-specific AI-accelerated CFD simulations is performed. This evaluation is performed in three categories: power efficiency, speed, and accuracy. The specific domain in the study is the FDA nozzle benchmark, which is simulated using SimpleFoam, a laminar solver that is a component of the OpenFOAM CFD toolbox. The proposed AI model is a low-parameter feed-forward neural network with three fully connected layers, trained using steady-state solutions distinguished by various Reynolds numbers, all of which are computed by the OpenFOAM framework. The proposed model can then generate the steady-state CFD simulation result given the initial few iterations generated by the solver. Moreover, this paper introduces a hardware implementation for inference of the simulation results using an SoC chip with minimal hardware resource utilization. The suggested hardware design is developed from scratch for Zynq-7000 System-on-Chip, using only VHDL, and requiring no dependencies on third-party commercial AI frameworks or costly FPGA boards designed for AI-related applications. The proposed workflow in the test case study achieves a 98% reduction in simulation time while maintaining relatively high accuracy and a 95.6% reduction in energy consumption compared with the regular CFD workflow.