An Efficient AI-Based Classification of Semiconductor Wafer Defects using an Optimized CNN Model

Wafer maps used to display defect patterns in the integrated circuits industry include crucial information that quality engineers may utilize to identify the cause of a defect and increase yield. In this paper, we put forth a framework for accurately and quickly categorizing semiconductor wafer faults utilizing particularly CNN-based models. This paper seeks to provide a scalable, adaptive, and user-friendly implementation of convolutional neural networks for applications classifying semiconductor defects. In categorizing the defects found on semiconductor wafers, the suggested CNN model obtained an accuracy of 90.50% & 92.28% and losses of 0.39 & 0.29 while performing the training and validation, respectively, along with the misclassification rate of 0.0772. The suggested model also learns rapidly on the validation set at a rate of 1e-03 per second. The proposed custom CNN model architecture incorporates only two convolution layers, resulting in a greatly reduced number of parameter weights and biases. Specifically, the number of parameters is only 44000, which makes the model more compact, cost-effective, and robust against random noise. Moreover, this model can function well under low power and processing limits.