Abstract
Thermal analysis is used to predict operating temperatures in electronic packages. Thermal analysis is deterministically conducted using finite element, control volume, finite difference, or analytical methods. Input parameters include package geometry, material properties, and cooling conditions. Accuracy of the predicted temperatures depends on the accuracy of the material properties and the imposed heat transfer coefficient. The run-time for a thermal model depends on the complexity of the package geometry. For a complex computational fluid dynamics (CFD) model, the run-time may extend over several hours.In this paper, we utilize a predictive model using artificial neural networks (ANNs). The forward and backward propagation steps of the neural network model are programmed using MATLAB, with a cost function that minimizes the error between the inferred and the input data generated using a finite element model. The input data are randomized and then normalized so that the ANN weights can be reasonably determined.The input data are generated using analytical and ANSYS thermal models of a chip-on-substrate geometry applicable to electronic packaging. The chip size, the TIM bond thickness, the heat spreader thermal conductivity and thickness, and the heat transfer coefficient imposed on top of the heat spreader are varied in the thermal model. The imposed heat transfer coefficient combines the effect of the cooling fluid and the extended surface area of a heat sink or cold plate attached to the lid.Various input sets are analyzed to co-optimize the size of the training set, the number of hidden layers in the ANN model, and the inference error. It is demonstrated that ANN methods can be effectively used to predict package operating temperatures.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have