Accurately predicting Thermal and Flow parameters in Battery Pack Thermal Management System is very important to ensure maximum system efficiency, long range and vehicle safety. Traditionally, conventional Computational Fluid Dynamics (CFD) methods are used to get detailed insights into these parameters. But these methods remain computationally very expensive, which limits their application in real-time predictions. This study presents an alternate structured simulation framework that combines CFD fundamentals with Physics-Informed Neural Networks (PINNs) to enable fast and efficient prediction of thermal and flow behavior. PINNs incorporates the governing partial differential equations (PDEs) - mass, momentum and energy – directly into the neural network's loss function, enabling efficient learning with limited dependency on training datasets.
In this study, we present a PINN framework trained on limited datasets derived from general CFD simulations of flow and Conjugate Heat Transfer (CHT) cases for a battery pack. The trained PINN model is then validated against a separate set of test datasets obtained from independent studies, across key performance parameters including pressure drop, velocity, temperature distribution, and hotspot location. Comparative analysis indicates that the proposed PINN framework predicts key parameters within 5% deviation from test dataset values. The PINN model also significantly reduces computational time from 5-8 hours per case for traditional detailed CFD to 2-3 seconds. Additionally, the PINN framework requires only a closed-volume CAD geometry and does not rely on 3D-mesh. Limited data can be used to speed up the training process, thereby eliminating the time associated with pre-processing, repeated model setup for different cases, and subsequent corrections. Findings from this study demonstrate PINNs as a potential surrogate modelling tool for real-time thermal monitoring and rapid design iteration in battery cooling systems. This reduces dependency on computationally expensive simulations, thereby reducing development cycles and faster time for customers. This work shows a scalable physics-informed neural-network framework that can be extended beyond battery systems to other components at system and vehicle levels.