• Information
  • Production


In the AISSI project, researchers of KIT are collaborating with industry partners to show how the use of artificial intelligence could decide the race in the semiconductor crisis: they are working on making chip production in Germany more agile and competitive.

Microchip - product of the hour

"Our world is noticeably changing, moving towards more digitalization and electromobility. Semiconductor chips are the basis for implementing new technologies such as electric cars, 5G or Industry 4.0," says Christoph Jacobi from the Institute of Materials Handling and Logistics Systems (IFL) at KIT. Today, chip production is mostly dominated by Asian or American giga fabs. Mass production of identical chips is the credo of these companies in order to keep production costs and end customer prices for semiconductor chips as low as possible. The industry leaders are responding to the continuing high demand for chips by massively expanding their production capacities. European manufacturers are losing out in this competition.


A dark factory hall with robots.
Example of semiconductor production in Germany: In the AISSI project, new AI-based approaches are intended to improve production planning for chip manufacturing and make it more competitive. (Image: IM Imagery / Shutterstock.com)

Class instead of mass

Catching up on the technological and systems engineering lead is only conceivable with the greatest effort. "The European semiconductor industry is taking a different approach to asserting itself on the chip market: with top performance in a niche market. The production of highly specialized and customer-specific microchips for complex technical applications is a European specialty," explains Dr. Andrej Gisbrecht from Robert Bosch GmbH. "Such special orders in small quantities generate an immense planning effort in the production processes of the chip factories. Depending on which material, which systems and work steps or how many personnel are used, there are very large differences in production performance. If you plan inefficiently, you don't make full use of the available resources and therefore produce less," continues Gisbrecht. This is why every professional production facility has experts for the comprehensive planning of all production parameters, who define the short and medium-term work plan in a so-called schedule - with the aim of achieving high productivity (i.e. high output) while meeting the promised delivery dates (high level of service for customers).



Semiconductor production meets AI

In the AISSI (Autonomous Integrated Scheduling for Semiconductor Industry) project, the project partners around IFL and Bosch have set themselves the task of taking production planning to a higher level with artificial intelligence and thus improving the performance of semiconductor production. Jacobi gives an insight into the project work: "We have developed data and decision models based on deep reinforcement learning (DRL) for the integrated planning of production orders and maintenance activities, which are intended to support the planning experts as a decision-making aid. To do this, we use the data from our partners' real production and map it in a digital twin. This digital fab is currently serving as a simulation environment and training ground for the AI and we are evaluating its performance using various comparative models." To make the AI applicable in a production environment, the project partners are building a scheduling agent - software that receives the current status in production as input and returns a scheduling decision. An initial demonstrator is currently nearing completion.



Close-up of robot processing chips
A disk-shaped wafer passes through the semiconductor factory and is processed several times as a workpiece. After several production steps, the wafer is cut into several thousand small microchips. (Image: IM Imagery / Shutterstock.com)

From data master to decision tree

Manufacturing processes in the semiconductor industry are among the most complex industrial processes: Parallel process sequences with hundreds of processing steps, some of which are repeated, make production planning a decision-making problem. This results in a large number of decision options, which are mapped in the AI-supported scheduling agent in the form of a branched decision tree. The connected branches in the tree-like graph correspond to the possible decision paths. "With the help of the decision tree, we enable our AI model to narrow down the search for the most sensible solution in the overall solution tree. A complete search of the solution space is not possible due to the large number of decision paths. By combining the tree search with artificial neural networks, we obtain a prediction of the area of the search tree in which the schedule can be found with which the output of the entire factory will ultimately be as stable and maximum as possible," says Jacobi, describing the decision-making process.


Step ahead with AI

Jacobi is optimistic about the future: "At the moment, global capacities do not cover the actual demand for chips, but there will come a time when it turns into overproduction. An important point is not simply to produce a lot, but also to produce as efficiently as possible. This peak will determine which manufacturer can offer its chips at the best conditions and the lowest price. Many people do not realize how crucial logistics are in such cases. It is only when there is a hitch in the system, such as during the semiconductor crisis, that it becomes clear how important functioning processes are." With the scheduling agent, the AISSI partners want to offer a logistical tool in the future to make production planning as efficient as possible.


Images: IM Imagery / Shutterstock.com

Diese Seite nutzt Website-Tracking-Technologien von Dritten, um ihre Dienste anzubieten. Ich bin damit einverstanden und kann meine Einwilligung jederzeit mit Wirkung für die Zukunft widerrufen oder ändern.

Alle akzeptieren Einstellungen Nur notwendige akzeptierenImpressumDatenschutz