Deep Learning for Classification of the Chemical Composition of Particle Defects on Semiconductor Wafers

Manual classification of particle defects on semiconductor wafers is labor-intensive, which leads to slow solutions and longer learning curves on product failures while being prone to human error. This work explores the promise of deep learning for the classification of the chemical composition of these defects to reduce analysis time and inconsistencies due to human error, which in turn can result in systematic root cause analysis for sources of semiconductor defects. We investigate a deep convolutional neural network (CNN) for defect classification based on a combination of scanning electron microscopy (SEM) images and energy-dispersive x-ray (EDX) spectroscopy data. SEM images of sections of semiconductor wafers that contain particle defects are fed into a CNN in which the defects' EDX spectroscopy data is merged directly with the CNN's fully connected layer. The proposed CNN classifies the chemical composition of semiconductor wafer particle defects with an industrially pragmatic accuracy. We also demonstrate that merging spectral data with the CNN's fully connected layer significantly improves classification performance over CNNs that only take either SEM image data or EDX spectral data as an input. The impact of training data collection and augmentation on CNN performance is explored and the promise of transfer learning for improving training speed and testing accuracy is investigated.