Automated Pattern Recognition in Semiconductor Test Data: A Machine Learning Approach to Analyzing Cumulative Frequency Plots

Manual analysis of test data plots in semiconductor manufacturing creates a significant bottleneck in production optimization, with product engineers dedicating considerable hours to reviewing patterns and anomalies in cumulative frequencies plots. This study presents an automated approach that uses computer vision and machine learning to streamline this process. We developed a multistage pipeline that detects and extracts plots from various document formats using Faster R-CNN, generates feature embeddings via Vision Transformer (ViT-B/32), and employs Affinity Propagation clustering alongside Isolation Forest for anomaly detection. An integrated LightGBM model provides comprehensive statistical assessments, including descriptive statistics, distribution analysis, and intelligent evaluation of plot abnormalities. Field validation demonstrated a 95% accuracy in plot detection, a 45% reduction in analysis time (from 8 to 4.4 hours) and a 35% increase in anomaly detection rates compared to manual methods. In particular, the system identified subtle deviations in the pattern that manual inspection might overlook. The successful deployment of this system underscores the viability of AI-driven automation in critical semiconductor manufacturing processes, contributing to faster yield optimization and reduced time to market.