Optimization of Power Module Cooling Plate: An Application of Deep Learning for Thermal Management Devices 2024-01-2583

To meet the ever-increasing demands of the engineering industry, novel approaches to design optimization are essential, especially in fast-paced production environments. Conventional CAD and simulation tools often struggle to keep up with the complexity and speed required for designing critical components. In this context, leveraging Deep Learning technologies presents a promising solution by integrating knowledge from simulations and designs to drastically accelerate product development. With the drive for Electrification, conventional power electronics and systems are becoming more energy dense and hence requires compact and efficient thermal management solutions. Higher energy density is attributed to high power electrical components fitted in packs with shrinking characteristic dimensions and hence needs more efficient and compact thermal management solutions. Conventional engineering design approaches have limitations to push the boundaries of efficiency and power density of air or liquid cooled solutions. This paper explores the application of recent algorithms based on Geometric Deep Learning to construct a predictive model capable of accurately assessing the thermal performance of cooling plates for power modules within a matter of seconds. Leveraging this predictive model, non-parametric design optimization allowed us to rapidly explore an extensive and innovative design space, overcoming the limitations of traditional parametric optimizations. This study reveals significant thermal performance gains achieved through the optimization process, while considering the manufacturing constraints of the cooling module. The flexibility of this approach is demonstrated by comparing the baseline design to those optimized for cost and/or thermal performance. The results showcase cooling modules that significantly outperform the standard product, bringing more than 30% improvement in pressure drop, together with reduced costs, the overall weight being reduced by more than 10%. The Deep Learning based optimization workflow converged towards innovative designs which could not have been achieved with conventional CAD and CAE tools.