These slides were presented at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

The cooling system of an automobile requires heavy consideration in the design phase, particularly for electric vehicles which have more heat sources than a conventional internal combustion engine. It is in the designer'™s best interests to create a cooling system that maintains operating temperatures within optimal limits while consuming less power for better mileage. Furthermore, there are hard limits for EV battery pack temperature that must be maintained to avoid thermal runaway. With the advent of many thermal 1D-CFD commercial simulation software, it has been made possible to simulate these cooling systems with a high degree of accuracy, allowing for several design iterations to be tested and analysed without requiring costly and time-consuming physical testing. These software packages, however, still fall short of creating optimal design iterations of the cooling circuits. This can mainly be attributed to limited optimization algorithms available in the software. Rerunning simulations for every possible design is also computationally expensive and time-consuming. Hence, it is preferred to have an optimization process that can provide a number of potential designs that meet the required targets, in a short period of time. In this work, optimization of an EV cooling system to achieve the reduction of the maximum e-motor temperature by 5°C is presented. A 1-D model of the existing cooling system is developed in GT-Suite to generate the Design of Experiments (DOE) matrix for the selected optimization parameters, such as the coolant flow rate, heat exchanger dimensions and so on. Subsequently, an ML-based approach for design optimization is applied using SIMULIA ISIGHT. Using the DOE matrix with parameters and outputs, a Radial Basis Function (RBF) neural network approximation of the cooling circuit is created. RBF approximations are characterized by reasonably fast training and compact networks, and they are useful in approximating a wide range of nonlinear spaces. The tool also has features to analyse the sensitivity of parameters, allowing the selection of only the most significant parameters for design optimization. The non-dominated sorting genetic algorithm (NSGA2) is used to identify optimal solutions. This multi-objective algorithm can identify multiple local minima for the given objectives and constraints, which results in three optimal design iterations capable of achieving the 5°C temperature reduction with different parameter values, for example, the lowest coolant flow rate or the smallest heat exchanger size, etc. The overall ML-based optimization process is about 30 times faster than the conventional optimization process and can significantly speed up the design phase. The created designs are retested in the original 1D-CFD software, showing high accuracy and low error of 1% in the motor temperature, thus validating the optimization results.