An AI-Accelerated CFD Application on a Benchmark Device: FDA Nozzle

Abstract

In this study, we suggest a procedure for speeding CFD (computational fluid dynamics) analysis up by combining a conventional opensource CFD solver with a traditional AI module. The studied case is the FDA benchmark nozzle with various Reynolds numbers. The considered CFD simulations belong to a group of steady-state simulations and utilize the laminar flow solver SimpleFoam in the OpenFOAM toolbox. The proposed module is implemented as a Feed-Forward Neural Network (FFNN) supervised learning procedure. Our method distributes the data by creating a combined AI model for each quantity of the simulated phenomenon for various Reynolds numbers. The model can then be combined after the initial iteration phase to decrease the execution time or to lower memory requirements. We analyze the performance of the proposed method depending on the estimation accuracy of the data of interest, velocity, and pressure. For test data, we achieve time-to-solution discounts of nearly a factor of 10. Comparing simulation results based on the FFNN test results and 3D visualization shows the average accuracy for all the parameters over 99% for the velocity and the pressure.