Machine Learning–Based Predictive Maintenance for Semiconductor Fabrication Equipment: Failure Modes, Data Challenges, and Yield Impact

Abstract

This study quantitatively evaluated the effectiveness of machine-learning-based predictive maintenance for improving equipment reliability, failure prediction, maintenance efficiency, and production performance in semiconductor fabrication equipment. A longitudinal and quasi-experimental design was applied to an integrated dataset comprising 118 semiconductor manufacturing and supporting equipment units. Following data-quality screening, 2,342,816 equipment-cycle records, 17,923,648 sensor observations, 1,286,540 wafer-processing runs, 9,742 maintenance events, and 2,184 verified failures were retained, with overall data completeness reaching 98.7%. Data challenges were substantial: failure-related observations represented only 1.9% of condition observations, and 143,904 records (5.8%) were excluded because of missing measurements, duplication, timestamp inconsistency, uncertain failure labels, or nonproduction status. Equipment-condition analysis showed that vibration intensity had the strongest relationship with failure occurrence (r = .61; β = .34), followed by alarm frequency (r = .58), vacuum deterioration (r = .55), and temperature deviation (r = .52), with all major relationships significant at p < .001. Failure-mode profiles differed by equipment category: progressive vibration, thermal, pressure, and vacuum degradation in vacuum and pump systems provided the longest median warning lead time (52.7 hours), whereas more abrupt failures in ion implantation provided the shortest (26.8 hours). Conventional machine-learning and deep-learning models were compared using classification, discrimination, and remaining-useful-life metrics. The CNN-LSTM hybrid achieved the strongest independent-test performance, with 97.8% accuracy, 96.4% precision, 96.1% recall, an F1-score of .963, ROC-AUC of .994, and PR-AUC of .979. Remaining-useful-life prediction achieved an MAE of 7.4 hours, RMSE of 11.2 hours, and R² of .94, with a median warning lead time of 41.6 hours. AI-supported predictive maintenance reduced unplanned downtime by 38.5%, failure frequency by 33.3%, emergency repairs by 41.0%, and MTTR by 30.9%, while MTBF increased by 50.0%. Equipment availability increased from 93.1% to 96.5%, and overall equipment effectiveness increased from 75.9% to 81.6%. Degraded equipment conditions increased wafer defect density by 75.0%, reduced throughput by 12.6%, and increased cycle time by 25.2%, whereas predictive maintenance substantially moderated these adverse effects. Yield impact was largest for plasma etching, where degradation increased defect density by 68.4% and reduced first-pass yield by 4.6 percentage points. Temporal and cross-equipment validation confirmed comparatively stable model performance. Generalization to previously unseen equipment remained a data challenge, with accuracy declining from 97.8% to 96.2%.