AIMS5.0 Use Case Spotlight · UC9: Energy-aware scheduling of batch processing machines in wafer fabs

In a wafer fab, a small number of process steps account for a large share of the electricity bill. Diffusion and oxidation furnaces are a typical example. They process many wafers in one batch, at high temperatures and often for several hours. Around them, cooling, air handling and ultrapure water supply run day and night to keep the cleanroom stable.

For production planning, these batch machines bring a recurring trade-off. A furnace can start with a partly loaded batch, which keeps urgent lots on schedule but uses energy and capacity less efficiently. Or it can wait for more wafers, which fills the batch but delays lots that are already due. Planners face this decision many times a day, and in most cases the delivery date decides.

Use case 9 (UC9) of AIMS5.0 developed AI methods that bring energy into this decision. FernUniversität in Hagen, Infineon, Systema and Virtual Vehicle worked on the use case within work package 6, which covered twelve industrial use cases along the semiconductor value chain.

Dispatching rules generated by genetic programming

Most of the work in a fab is scheduled with dispatching rules: priority rules that decide which lot is processed next on which machine. Designing good rules by hand is difficult, because product mix, due dates and machine constraints interact in ways that are hard to oversee.

The UC9 team used genetic programming to generate dispatching rules automatically. Candidate rules are assembled from basic building blocks, evaluated in a simulation model of the fab, and developed further if they perform well. Performance was measured as total weighted tardiness, which adds up how late each lot is and weights the delay by the lot's priority. The generated rules reduced total weighted tardiness by up to 20% compared with the reference rules.

Energy as a second objective

In the next step, the team extended scheduling to two objectives: delivery performance and energy consumption. The energy-aware scheduler takes into account how much energy a batch will need and when demand and electricity prices are high. It can then shift or combine batches so that the load is spread more evenly over the day, while lots keep moving through the line.

In situations with high energy peaks, this can save about 750 kWh per day. The energy-aware scheduling and energy management methods are documented among the 54 AI technologies collected by work package 2 of AIMS5.0, so that other projects and companies can build on them.

Measuring where energy is used

Scheduling around energy requires data on where and when energy is consumed. Instead of installing meters at every machine, the team identified the most relevant points in the fab's energy network and placed smart meters there. This keeps the cost of measurement low and still gives a reliable picture of the main energy flows. The same knowledge is useful in an energy shortage, when a fab has to decide quickly which processes get priority.

Predicting problems in the cooling network

UC9 also looked beyond the production tools. The cooling network is critical infrastructure in a fab, and a failure can force tools to stop. The team trained an AI model that detects and predicts anomalies in the cooling network with an accuracy of 98.6%. Early warnings give maintenance staff time to act before a fault leads to downtime or energy losses.

Links to other use cases

The scheduling results of UC9 feed into use case 10, an AI-enabled manufacturing execution system that connects several WP6 use cases, so that planning, material flow and maintenance work with the same data. FernUniversität in Hagen presented its simulation-based planning methods from AIMS5.0 at the Winter Simulation Conference 2025.

Relevance beyond semiconductors

Energy-intensive batch processes are common in other industries as well, for example in the heat treatment of metals or in glass and ceramics production. As energy prices and emission targets put more pressure on manufacturers, methods that plan production with both delivery dates and energy use in view will become more widely needed.

The UC9 partners will present the results 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.