Data-Driven Process Analytics for Early Defect Detection in High-Volume Semiconductor Fabrication

Abstract

High-volume semiconductor fabrication generates large amounts of sensor, equipment, and inspection data, but defects are often confirmed only after several wafers have already been processed. This study presents a data-driven process analytics framework for early defect detection using process signals, equipment-state records, event information, and early quality indicators. Historical wafer-processing data were cleaned, standardized, and converted into statistical and time-dependent features. Principal component analysis, one-class support vector machine, random forest, and autoencoder models were compared across production batches. The results showed that the autoencoder achieved the highest defect detection rate, strongest recall, and earliest warning, while conventional statistical monitoring produced lower performance. Chamber-pressure variation, plasma-power instability, endpoint-time deviation, gas-flow imbalance, and chamber age were identified as the most important defect indicators. The proposed framework can support faster engineering investigation, reduce wafer exposure to unstable process conditions, and improve process control in high-volume semiconductor manufacturing without replacing established quality inspection procedures.