Description
The use of Computational Fluid Dynamics (CFD) is highly valuable for investigating airflow in electric vehicles to achieve efficient thermal management by numerically solving the relevant governing equations in real-world scenarios. However, depending on the complexity of the problem, CFD simulations can lead to significantly high computational costs [1]. To address this challenge, Machine Learning (ML) is increasingly being integrated with CFD, aiming to reduce the need for computationally demanding simulations when studying multiple scenarios.To optimize the thermal management system and improve energy efficiency in electric vehicles (EVs), this study employs a data-driven surrogate modeling approach. High-fidelity cabin airflow simulations are first generated using the open-source CFD software OpenFOAM [2]. These simulation results serve as the training data for a neural network developed in PyTorch [3], enabling rapid and accurate prediction of cabin airflow patterns. The resulting model supports the efficient operation of the HVAC system, ultimately contributing to improved thermal comfort, reduced energy consumption, and extended driving range. The neural network is based on the U-Net architecture, which uses skip connections to enhance both accuracy and efficiency. U-Net is well-suited for fluid flow prediction due to its ability to capture detailed flow features with low computational cost [4]. A comparison is made between the CFD model and the ML-based surrogate model, with boundary conditions for velocity and temperature used as input parameters. An automated script generated 132 simulation cases for analysis.
The surrogate model dramatically reduces computational costs, achieving near real-time inference speeds (2.1 seconds), while maintaining a prediction accuracy exceeding 95% in most regions of the cabin. Despite minor limitations in resolving complex turbulent structures with high velocity gradients, the model demonstrates high accuracy in capturing key thermal comfort metrics, with less than 4% deviation for temperature predictions. These findings emphasize the surrogate model's potential for integration into EV thermal management systems, supporting both design optimization and real-time control strategies.
| Period | 4 Jul 2025 |
|---|---|
| Event title | 20th OpenFOAM Workshop |
| Event type | Conference |
| Location | Vienna, AustriaShow on map |
| Degree of Recognition | International |
Research Field
- Vehicle System Simulation