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

Abstract

Electronics systems are becoming more and more compact and require sophisticated optimization techniques to achieve optimal performance. The growing complexity of the design space makes it very difficult for engineers to identify the key design variables that need to be optimized. One state of the art approach is to apply parametric optimization procedure to the thermomechanical setup of an electronics system. This includes an automatic workflow calling different software tools in batch and starts with generating a geometry based on the parameter values of a single design out of a possible sampling scheme. For each design the geometry is consumed by the following CFD and FEA analysis. Using this process, you can optimize your design by using scalar responses of the simulation e.g. for minimal heat sink mass or thermal stress while constraining the temperature within a given limit. Conducting a constrained Multi-Objective Optimization leads to hundreds or even thousands of design evaluations and thus to a high numerical effort. In order to keep the numerical effort reasonable, 0D (scalar) ROMs are used that approximate the responses within the design space of the design variables. With sufficient quality, they can be used for (pre-) optimization where thousands of designs and different optimization runs can be realized quickly. This allows us in several steps to convert the Multi-Objective Optimization task into a Single- Objective Optimization by summarizing positive correlated objectives and formulating preferences based on the results. One key element in this approach is the quality of the ROMs. Traditionally regression models like Polynomial Methods, Moving Least Squares or Kriging are used to set them up. Nowadays ROMs can be generated using machine learning techniques like Deep Feed Forward Networks or even mixed approaches that combine Neural Network approximation with traditional models like Kriging. The described optimization methodology is applied to a optimization of a circuit board design. In total 12 Parameters are defined in the parametric model including number of fins and fin thicknesses of both heatsinks, fan flow rates and fan locations. The analysis starts with a Multi-Objective Optimization to reduce the thermal resistance of both heat sinks, pressure loss, thermal stress and deformation on board. At the same time constraining the heat sink mass and temperatures of the heat sources within given limits. Positive correlated objectives will be summarized into one singe objective and another objective is converted to constraint. This approach with be conducted using traditional ROMs and machine learning ROMs. Finally, we will discuss these ROMs and approaches for an optimal way to find the best designs.