This use case focuses on scaling machine learning solutions across semiconductor production sites and applications through MLOps, transfer learning, interpretable AI, heterogeneous data handling, and reusable model pipelines. The goal is to move from isolated ML prototypes toward productionized, monitorable, and scalable ML systems for defect recognition, virtual inspection, anomaly detection, equipment monitoring, and predictive maintenance. The use case also emphasizes FAIR data, harmonized storage, deployment infrastructure, and model reuse across equipment types. Its expected impact includes faster ML development, reduced data needs, better defect detection, longer component lifetime, reduced maintenance, lower scrap, and improved line yield.
| # |
Technology Name |
Description |
| 1 |
Machine Learning and Deep Learning |
Productionized machine learning models at scale are developed for semiconductor manufacturing to support early fault detection and quality control. The work includes deep learning forecasting models and AI-powered ML pipelines for OEE prediction and continuous improvement, explainable unsupervised anomaly detection for root-cause analysis, weakly supervised anomaly detection using user feedback over time, and defect and defect-pattern classification of silicon wafer images. The main goal is to save material and personnel costs, support concentration on high-value tasks, and improve quality. |
| KPI # |
Related Technology |
KPI Target |
| 1 |
ML/DL |
Estimated 25% reduction of required data for new ML solutions using DBACS. |
| 2 |
ML/DL |
Availability of an MLOps structure, with an estimated 75% reduction of time for ML model development. |
| 3 |
ML/DL |
Estimated 55% reduction of time for monitoring a given industrial process. |
| 4 |
ML/DL |
Availability of early warnings for production excursions and anomalies. |
| 5 |
ML/DL |
Achieve the targeted 3% improvement in average accuracy for the base materials 1 and 2 model. |
| 6 |
ML/DL |
Save 500 kEUR per year in material and personnel costs. |
| KPI # |
Baseline |
| 1 |
Only first approaches exist for machine learning applications in semiconductor manufacturing. |
| 2 |
Within the Arrowhead-Tools project, a first ML method in the area of ion implantation showed good results for improving quality and costs. |
| 3 |
No deep approach for ML and DL is available in semiconductor production. |
| KPI # |
Validation Method |
| 1 |
Reconstruction, prediction error, and misclassification rate are evaluated using metrics such as RMSE, MAE, and MCR for sub-use cases UC15-subB, UC15-subC, UC15-subD, and UC15-subE. Predicted and measured values are compared on a held-out test set, with targets such as RMSE below 10% of the signal range or R² above 0.9. |
| 2 |
Anomaly detection performance is evaluated using F1-score and MCC for UC15-subA. Detected anomalies are compared with labeled or simulated events in the same test sequence, with target values of F1-score above 0.85 and MCC above 0.8. |
| 3 |
Global Feature Importance is evaluated for UC15-subA by checking whether the model highlights physically meaningful and context-relevant features. Ranked feature importance is compared with engineering knowledge or baseline explanation models, with a target of more than 80% alignment. |
| 4 |
Local Explanation Fidelity is evaluated for UC15-subA by measuring consistency between local explanations and model predictive behavior. The validation compares influential features identified by the XAI method with those effectively driving the model output, targeting more than 70% overlap or agreement with model-based sensitivity maps. |
| 5 |
Inference latency and efficiency are evaluated for UC15-subB by measuring average inference time and memory usage on target hardware. The target is latency below 50 ms per inference and memory usage below 500 kB, enabling integration into on-device control and monitoring systems. |
- Industrial process phases made more effective and precise through integrated and tailored AI algorithms, methods, and tools.
- 20% improvement in prediction precision.
- 25% improvement in classification precision.
- 20% reduction of required data for scaling new ML models to new machines.
- 70% reduction of time for developing and integrating a ML model in the digitalized industrial domain.
- Reduced manufacturing costs in the long term.
- Improved overall productivity of the manufacturing factory by approximately 20%.
The use case contributes to more effective and precise industrial process phases by productionizing ML and DL models for early fault detection, anomaly detection, defect recognition, OEE prediction, and quality control in semiconductor manufacturing. The targeted 25% reduction of required data for new ML solutions directly supports scaling ML models to new machines and applications with less data. The availability of an MLOps structure and the estimated 75% reduction of ML model development time directly support the AIMS5.0 objective of reducing the time needed to develop and integrate ML models in digitalized industrial domains. Defect and defect-pattern classification, explainable anomaly detection, and weakly supervised anomaly detection support prediction and classification improvements, while early warnings for production excursions and anomalies help reduce process monitoring effort. The expected yearly saving of 500 kEUR in material and personnel costs contributes to long-term manufacturing cost reduction and improved productivity.
| AI Toolbox Tool |
Use within the Use Case |
| Time-Series Anomaly Detection Tool |
Used to detect production excursions and abnormal temporal patterns in semiconductor manufacturing data. It supported early warnings, predictive monitoring, and anomaly detection for equipment and process behavior. |
| Depth-based Isolation Forest Feature Importance |
Used to support explainable unsupervised anomaly detection by identifying which features contributed to anomalous behavior. It helped make anomaly-detection results more actionable for root-cause analysis. |
| Extended Isolation Forest Feature Importance |
Used to provide interpretable feature-importance analysis for isolation-forest-based anomaly detection. It supported explainability in weakly supervised and unsupervised monitoring scenarios. |
| Tabular Prediction Tool |
Used to train and evaluate ML models on structured semiconductor production data, including prediction, classification, and monitoring tasks across the UC15 sub-use cases. |
| Tabular Foundation Models |
Used to support rapid development and scaling of tabular ML models for new semiconductor datasets, contributing to reduced data requirements and faster model development. |
| Image Classification Tool |
Used to classify defects and defect patterns in silicon wafer images. It supported visual inspection, defect recognition, and quality-control automation. |
| Object detection tool |
Used to detect defect regions or relevant visual patterns in wafer images before classification or further quality analysis. |
| Explainability Tool |
Used to interpret ML and DL models for anomaly detection, defect classification, OEE prediction, and quality monitoring. It supported global feature importance, local explanation fidelity, and trust in productionized ML models. |
| Device-Level Quality Detection Tool |
Used to link device-level defects and production-quality issues to machine-signal-related features and root causes, supporting defect analysis and quality improvement. |