Prediction of Circuit Board Warp Using Convolutional Neural Network on Small Amount of Datasets

Abstract

A convolutional neural network, which reproduce a function by data, was used to predict the amount of warp distortion of a four-layer circuit board in a reflow soldering process. The data used for training are material properties such as Young's modulus, board thickness, and residual copper content as input data, while the warpage strain data to be predicted is the amount of warp of the circuit board obtained from the measurement. Since several distortion data to be predicted was insufficient to be used for training, a newly proposed data augmentation method used to increase a total amount of data. Proposed data augmentation is evaluated through the result of the predictions and discussed.