¶ UC12 - Human-Centred AI for the Optimization of Robust and Competitive Semiconductor Manufacturing Networks
This use case applies human-centred AI to reduce physical flow and information distortion in semiconductor manufacturing networks. It combines AI methods, operations research, factory physics laws, root-cause analysis, counterfactual reasoning, and advanced visualization to support better decision-making in factory clusters and supply chains. Human-algorithm learning loops are central: AI identifies variability drivers and unstable planning parameters, while human experts remain involved in interpreting and improving decisions. The expected result is more resilient semiconductor production, improved ATP stability, reduced variability, better OEE, higher trust in AI, and lower CO₂ per product through improved material, transport, and energy efficiency.
| # |
Technology Name |
Description |
| 1 |
Rapid modelling |
Rapid modelling is based on queueing theory, where the production process is considered as a network of queues. It helps understand the dynamics between production capacity, inventory, and variability, and supports variability reduction and better capacity planning. |
| 2 |
WIP level prediction using ML |
ML models are used to calculate WIP levels from factory data for ultra-short-term prediction, supporting shop-floor decisions within one shift up to one day. This helps operating workshops make quick decisions to balance WIP flows more effectively. |
| 3 |
AI-based decision support |
ML-based WIP predictions are presented through explainable AI approaches and suitable visualizations so that operators and decision makers can trust the prediction while also integrating their own expertise. |
| 4 |
Integrated transportation and production simulation model |
Production and logistics data are integrated into a combined simulation model. A correction factor is used to simulate dynamic effects and improve the prediction of transport loads based on production workloads. |
| 5 |
Digital Kanban system |
The material delivery process is digitalized and automated through a Digital Kanban system, reducing variability and manual interventions while increasing efficiency, transparency, and productivity. |
| KPI # |
Related Technology |
KPI Target |
| 1 |
Rapid modelling; Integrated transportation and production simulation model; Digital Kanban system |
Reduce variability in specific cases by 30% and improve OEE in specific work centers by 10%. |
| 2 |
WIP level prediction using ML; AI-based decision support |
Improve explainability of AI by 10% and increase acceptance of/trust in AI by 10%. |
| KPI # |
Baseline |
| 1 |
Historical data and rapid models based on it serve as the benchmark for comparing variability reduction actions. The OEE of the ECDP work center at IFD in December 2023 was 42.9%. |
| 2 |
N/A. So far, no AI-based decision support has been implemented to predict WIP levels and support decision-making during a shift. |
| KPI # |
Validation Method |
| 1 |
For variability reduction, potential improvement actions are consulted and investigated with factory experts. Only feasible actions are used to quantify variability reduction, calculated based on factory physics laws. OEE improvements are proven on existing bottleneck situations, considering tools, product mix, actual uptime restrictions, technologies, basic product types, and OEE data. OEE is calculated as Availability × Performance × Quality. |
| 2 |
WIP prediction is validated using standardized regression accuracy and goodness-of-fit metrics, such as RMSE and MAE. Standard regression models are compared with ML models to assess accuracy and robustness. Explainability is tested through continuous involvement and feedback from human decision-makers. Acceptance and trust are evaluated by comparing different representations, visualizations, and interactions with ML-based decision support, using users’ weight on advice as a measure of how strongly decision makers follow AI predictions. |
- Industrial process phases made more effective and precise through integrated and tailored AI algorithms, methods, and tools.
- 20% improvement in prediction precision.
- Improved overall productivity of the manufacturing factory by approximately 20%.
- Self-adjusted/adapted manufacturing processes above 15%.
- Providing documented methods for dynamic application of tools to improve sustainability.
The use case contributes to more effective and precise industrial process phases by combining rapid modelling, ML-based WIP prediction, explainable AI decision support, integrated transportation and production simulation, and a Digital Kanban system. The improvements in OEE and the reduction of process variability directly influence manufacturing productivity. Rapid modelling and simulation help identify sources of variability and support better capacity planning, while ML-based WIP prediction enables short-term shop-floor decision support. The Digital Kanban system increases the degree of logistics digitization and reduces manual interventions, uncertainties, delays, and processing time. Together, these technologies support more robust semiconductor manufacturing networks, improved productivity, and more sustainable planning and logistics decisions.
| AI Toolbox Tool |
Use within the Use Case |
| Tabular Prediction Tool |
Used to predict WIP levels from structured factory data and support ultra-short-term shop-floor decision-making. It helped model relationships between production capacity, inventory, variability, OEE, and operational conditions. |
| Lag-Llama |
Used to support short-term forecasting of WIP levels, transport loads, and production-related time-series indicators. It contributed to prediction tasks within one shift up to one day. |
| Time-Series Anomaly Detection Tool |
Used to identify abnormal temporal patterns in production, logistics, WIP, and transport-load data. It supported the detection of variability drivers and unstable operational conditions in semiconductor manufacturing networks. |
| Explainability Tool |
Used to make WIP predictions and AI-based decision-support outputs interpretable for operators and factory experts. It supported the human-algorithm learning loop by helping users understand which factors influenced AI recommendations. |
| Device-Level Quality Detection Tool |
Used to connect production, logistics, and machine-related signals with quality and OEE-related effects. It supported root-cause-oriented analysis of operational disturbances and manufacturing performance issues. |
| LLM Tool |
Used to support human-centred decision support by generating explanations, summaries, and guidance from production, logistics, and planning information. It helped make AI-supported recommendations easier to interpret by human decision makers. |
| Prompt Context Injection Tool |
Used to provide LLM-based decision-support components with relevant factory, WIP, OEE, transport, and planning context, ensuring that generated explanations and recommendations remained grounded in the use-case data. |