AIMS5.0 Use Case Spotlight · UC12: Human-centred AI for the optimisation of robust and competitive semiconductor manufacturing networks

A semiconductor fab is one of the most complex production systems in industry. Wafers pass through hundreds of process steps, many tools serve several products at once, and an automated transport system moves material between them around the clock. Every shift, planners decide which lot runs next, which tool needs attention and when material leaves the warehouse for the cleanroom. Many of these decisions still rely on long-term models and on the experience of the people on the shop floor.

Use case 12 (UC12) of AIMS5.0 set out to improve planning, decision-making, digitalisation and automation in semiconductor manufacturing and intra-logistics. WU Vienna, Hochschule Zittau/Görlitz, TU Dresden, Systema and Infineon worked together in three sub-use cases: short-term forecasting of work in progress, fast simulation of production and transport, and a digital Kanban system for material delivery. UC12 is part of work package 6, which brought AI into twelve use cases across the European semiconductor value chain.

Where the team started

At the start of the project, forecasts that support decisions in the fab were based on static long-term models. They could not provide short-term predictions, especially when data was incomplete. Historical data showed a high variation in how individual tools were used, which makes planning harder. Holistic simulation models that include the automated material handling system existed, but they were impractical for daily use. They either took too long to run or lacked the necessary detail, which limited the options for planning and control. Material delivery between warehouse and cleanroom still worked with physical Kanban cards.

Forecasting the next hours

The amount of work in progress waiting at a tool can change a lot within a few hours. The team developed an AI-based decision support system that predicts work-in-progress levels in the short term and shows them in an interactive dashboard. Planners can see where a queue is likely to build up and act during the same shift, for example by changing priorities or moving resources. The forecasting algorithms and the dashboard tools are two of the results that UC12 hands over for further use.

A year of simulation in two minutes

Capacity planning in a fab has to take the automated material handling system into account, because transport times and congestion directly affect throughput. The partners developed LogiFabSim, a simplified but accurate model of production and transport. It approximates the dynamics of the transport system with a correction term instead of simulating every movement in full detail.

The result is a model that simulates one year of operation in about two minutes. Compared with benchmark simulation frameworks, this reduces computation time by 75% to 99.9%. Because LogiFabSim is fast and highly adjustable, it can be applied to a wide range of scenarios and can also generate data to train AI algorithms. A holistic information model links the data from production and logistics, and rapid modelling methods based on queueing theory answer quick what-if questions about the production flow.

Digital Kanban

In the third sub-use case, the team digitalised and automated the material delivery process between warehouse and cleanroom. The physical Kanban cards were replaced by a digital Kanban system. This reduces variability and manual interventions and makes the process more transparent. It also lowers the risk of delays, handling errors and contamination, which matters in a cleanroom environment.

Trust in AI advice

The term "human-centred" in the title of UC12 refers to the people who work with these tools. Whether planners act on AI advice depends on more than accuracy. UC12 therefore also looked at how transparency affects the way people use AI recommendations. A first pre-test with 73 participants compared a "black box" AI with one that explains its reasoning, and both groups followed the advice to a similar degree. The team reads this as an early sign that explanations alone do not build trust, and that the way they are presented matters.

Results and where they apply

Each UC12 result has a defined field of application. Rapid modelling of production systems and process flows can be applied in any industry and is also used in teaching. LogiFabSim serves the semiconductor industry and automation, and a version of the factory simulation was developed for education. The algorithms for short-term work-in-progress forecasting and the dashboard tools support planning on the shop floor. The digital Kanban approach was developed for semiconductor back-end manufacturing and can act as an enabler for further automation. Course material based on these tools brings the results into teaching.

The UC12 partners will be at the AIMS5.0 booth at EFECS 2026 in Helsinki on 2 and 3 December. An overview of all nine work packages is available in the AIMS5.0 final results article.

The AIMS5.0 project is supported by the Chips Joint Undertaking and its members, including top-up funding by National Funding Authorities from involved countries under grant agreement no. 101112089.