Assessing the impact of novel polymers and thermal management in a power electronics module using machine learning approaches

We highlight the predictive utility of machine learning (ML) techniques in estimating thermal performance benefits in power electronics modules, resulting from the use of high thermal conductivity polymers and thermal management techniques. The thermal performance of a commercial 1.2kV/444A SiC half-bridge module is quantified via high fidelity numerical simulations. Parametric simulations are conducted by considering the thermal conductivity of i) encapsulant (anisotropic), ii) baseplate and iii) heat sink as variable parameters, along with the coolant temperature and convection heat transfer coefficient. These simulations generate a data set of more than 2500 data points, which is used to train and evaluate the performance of machine learning algorithms to estimate the maximum junction temperature (Tj) of the package. Parameters are varied to represent a broad spectrum of possibilities ranging from high thermal conductivity polymer-based heat sinks to copper heat sinks; and air to two-phase liquid cooling technologies. The performance of three different statistical machine learning models is evaluated: polynomial regression, random forest, and support vector machines in predicting Tj. While polynomial regression does not predict Tj with a reasonable accuracy, random forest and support vector machines demonstrate excellent prediction accuracies w ith overall R
²
of 99.6 and 99.98%, respectively. To estimate the relative contribution of the underlying thermal parameters, we use SHAP (Shapley Additive exPlanations) dependence plots in combination with random forest algorithm to identify parameters which strongly influence Tj. We observe that the thermal conductivity of heat sink material and heat transfer coefficient have the maximum impact on Tj reduction, whereas the thermal conductivity of the polymeric encapsulant has the least influence on Tj. The presently used approach of simulations-based training of ML algorithms can be adapted for the thermal design and parameter optimization in other electronics packages.