This conference paper was submitted for presentation at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

Automotive fans for engine bay cooling of internal combustion powered engines (ICE) or battery-powered motorized engines are a critical component of vehicle performance. Aside from their cooling capabilities that is assessed by the throughput of mass flow that they achieve, their acoustic behavior is also of paramount concern. The reason for this is that passenger vehicle noise perception is a measure of vehicle comfort. Having said that, fan acoustic behavior is even more critical for battery-powered vehicles where the fan'™s acoustics is no longer masked by the internal combustion powered engines'™ loud noise derived from the piston motion and associated combustion process. Dassault Systèmes'™ lattice Boltzmann method (LBM) software is able to perform direct noise computations and, thereby, assist in developing low noise fans using parametric design-of-experiment (DOE) studies that identify the fan geometry with best performance and acoustic characteristics. But, the question arises, can machine learning be coupled to these physics-based LBM acoustics simulations to accelerate the DOE study for low noise? The envisioned process is to train simple feed-forward neural networks (also known as multilayer perceptron networks) with high-fidelity computational fluid dynamics LBM acoustic run data, where the inputs to the neural network are fan geometry, while the outputs are fan surface pressure, volumetric velocity field and sound pressure level (SPL) at a designated microphone location. Subsequently, once the machine learning models are trained, unseen fan geometries are fed into the multilayer perceptron models to predict the fan surface pressure, volumetric velocity field and sound pressure levels (SPLs) at a designated microphone location at a fraction of the computational cost. This approach is in line with the growing trend of combining traditional computer-aided engineering (CAE) software with machine learning focused processes, allowing simulation and design engineers to explore more design options in a DOE without incurring excessive computational expenses, fostering innovation and efficiency in fan design processes.