Machine learning algorithms for design optimization of Ceramic Substrates-DBC for Power Applications using local approaches

Direct Bonded Copper (DBC) substrates are key components in power electronic devices, due to their ability to combine thermal, mechanical, and electrical properties in a single system. They are made of a ceramic layer coupled with copper layers on one or both sides and can be subjected to thermal cycling and significant mechanical stress during operations. The difference in thermal expansion coefficients for copper and ceramic layers causes tension and compression states at interface surfaces, which can lead to delamination, with critical consequences for the reliability of the component. For this reason, the development of innovative solutions, such as the introduction of optimized patterns or dimples to improve durability, has been essential. Strain Energy Density (SED) has been used to assess how the combination of different geometric parameters affects the strain energy close to the notch tip at the interface of ceramic and copper. Machine Learning (ML) algorithms allow to investigate several geometric configurations sampled by means of Latin Hypercubic Sampling (LHS) method, which consists in extracting design variables from predefined domains, allowing the analysis of a reduced number of configurations compared to what would be done with a Full Factorial approach. Each variable’s effect on objective function is evaluated through Spearman correlation method, and the design points are defined as the minimum of the response surface obtained by numerical simulations.