CFD-Informed Neural Networks for Centrifugal Fan Design: Combining Simulation and Deep Learning to Predict the Performance of Parametrically Varied Fan Blades

Abstract

This project presents the development of a surrogate model for centrifugal fan performance, focusing on the Same Sky CBM-97S series DC fan. The objective is to develop a validated CFD model of the centrifugal fan and to use the model to train neural networks that can predict the performance of varying blade geometries. To achieve this a steady state CFD model was developed in STAR-CCM+ using k − ω turbulence model and a moving reference frame to simulate the rotating impeller. After validation against experimental data, a simplified fan geometry was parametrized, and a dataset consisting of 200 simulations with varying blade count and blade angle of attack was generated. Two different supervised Neural Networks were then trained: the first predicts the pressure and velocity fields across a 2D plane section, while the other estimates performance parameters such as outlet mass flow and pressure based on the geometric inputs. Both models demonstrated high accuracy. The flow field model achieved R2 values above 0.94. The fan performance model showed mass flow and pressure predictions with errors below 6%. The project shows that the neural networks trained on the CFD simulation data are able to accurately predict a two dimensional flow field and performance variables. Future work may involve extending to 3D field predictions using a physics-informed neural network, incorporating additional parameters such as blade length, and optimizing network architecture for enhanced performance.