Deep Learning Modelling for Composite Properties of PCB Conductive Layers

This paper presents the development of a novel modelling approach, based on the use of deep learning (DL), to predict the orthotropic composite properties of copper-patterned conductive layers of printed circuit boards (PCBs). This data is needed to assess the bulk PCB properties with existing methods for laminar composites. Image datasets of copper patterned artwork, required with this approach, are gathered and the composite (homogenised) orthotropic elastic modulus of the respective conductive layouts is evaluated through an automated, macro-script executed, finite element analyses. The modulus values are assigned as labels to each image of a copper layout in the dataset. A regression convolutional neural network is developed and optimised using a training dataset and validated using the test dataset.The results show that the DL model can predict the orthotopic values of the elastic modulus of highly non-structured copper patterns accurately, with the absolute errors of the predicted vs. true (FEA evaluated) property value being less than 3% of the composite propriety range for 99% of the patterns in the validation dataset. The advantages of the proposed machine learning solution over existing techniques are that it can be digitalised and made available to the end-user as an easy-to-use and computationally fast toolset. The modelling approach can enable design engineers effectively explore PCB design alternatives, with awareness of their thermo-mechanical properties and the effect they have on the assembly performance and components' reliability.