Multi-Objective Evolutionary Optimization of Multi-node Network for Thermal Modelling of Electronic Package

In 1996, the concept of Compact Thermal Model (CTM) was proposed by the European research project referred to as DELPHI. Its objective was to from a set of data, generated by numerical simulations, to create the simplest multi-node thermal model that allows preserving an acceptable accuracy whatever the operating conditions of the inputs. The established model is a black-box object combined to a network of resistors that links a single temperature-sensitive node to major surfaces of heat extraction. This surrogate model is built with the aim of approximating the thermal behavior of an electronic component submitted to a large range of boundary conditions. Over the last two decades, new Reduced Order Model (ROM) methods were studied but at board modeling level, nodal analysis model remains the most practical solution to minimize numerical model size and computation times. However, the agreement of DELPHI's CTM standardized method suffers many limitations such as the choice of appropriate optimization techniques or the definition of training multi-objective criteria. The present work discusses the use of Differential Evolution (DE) algorithms to formulate a robust chromosomes-genes fitting procedure where a relevant multi-node network can be extracted. So, the performances of the Classic-DE algorithm were analyzed on several test cases of an electronic component family, referred to as Quad Flat No-lead package (QFN). Whatever the studied package size, a deduced six-node matrix proves its ability for training data to yield high-accuracy resistance-network models and to perform well for training-independent validation scenarios of boundary conditions. The prediction of the component most sensitive temperature using a very simple black-box model form never exceeds on average 1%.