AI-Based Performance Prediction and Its Application on the Design and Simulation of Cooling Plates for Battery Electric Vehicles

With the increasing focus on the electrification of personal mobility, shortened development cycles with high cost pressure have to be managed quicker than ever in the Automotive Industry at minimal development cost. As a global automotive supplier, Mubea develops novel manufacturing technologies for new products in the electric powertrain. One example in the battery case of battery electric vehicles is the thermal management system of the traction batteries - a key factor in the battery case. As a new production technology for this product, the Mubea Rollbonding Process of aluminum offers several advantages, such as a high design freedom of the channel structure. For the product development – even though Mubea consequently makes use of automated simulation workflows – the turn-around time of a single CFD performance evaluation is still high. Therefore, our goal is to use all the historic simulation results from past projects to build a predictive model that allows the prediction of simulation results in real-time. However, given the high freedom in the design space allowed by the Rollbonding process, standard Machine Learning approaches, based on parameters, are not suitable. Hence, there is the need to directly process 3D geometries as such. Using historical engineering data, the unique deep learning approach of Neural Concept is able to predict unseen designs in seconds rather than hours, directly from the raw CAD file. This innovative approach allows Mubea to iterate faster and shorten the response time on customer enquiries. In conjunction with other design disciplines and manufacturing data, we look forward to have not only an AI-based design evaluation but also a tolerance-aware design optimization. In this paper we present an innovative strategy to utilize historic simulation results, and the corresponding 3D geometries, to predict the performance of new designs instantaneously. After explaining the underlying approach, first results are discussed. It can be shown that with as little as 100 training samples, this approach is able to deliver predictions with sufficient accuracy and over 90 % of lead-time reduction. Finally, we explain how Neural Concept and Mubea are collaborating to embed this approach in the Mubea design and simulation environment.