At a glance
- Company
- Hitachi EnergyMaker of power semiconductors; at the time of the study the plant belonged to Hitachi ABB Power Grids
- Industry
- Manufacturing
- Where and when
- Switzerland. Data to April 2019, rollout February – August 2020, published 2021.
- Who produced the figures
- Measured by researchersThe figures come from a study by university researchers, published in a peer-reviewed journal. The company supplied the production data and reported the 21.7% figure to the researchers. Yields in the paper are rescaled to protect confidential data.
What they did
The plant in Lenzburg makes transistor chips for power equipment. A chip passes through 200 process steps. At the end each chip is tested, and those that fail are scrapped. The share that passes is called the yield.
With so many steps, it is hard to tell which one causes the failures. The plant’s usual tools were correlation analysis and the judgement of its experts.
Researchers took the records of 1,197 past production batches, about 1.8 million chips. Each batch came with 3,614 recorded settings and measurements. They trained an AI model to predict a batch’s yield from these. Then they used a method that shows how much each input contributed to the prediction. This ranked the 200 steps by how much they affected yield.
Two steps stood out. In both, the model found that the result depended on which machine had done the work. It suggested sending material through one particular machine in each step. The plant confirmed that this would not create a capacity problem.
The numbers
The test. The plant made one new batch of 24 wafers, the silicon discs that chips are cut from. It split them into four groups of six, 372 chips in each. On the paper’s 0–100 scale:
| Group | Yield score |
|---|---|
| Made the usual way | 38.2 |
| First change only | 50.6 |
| Second change only | 85.7 |
| Both changes | 86.0 |
- Yield loss down 21.7%. The company compared the group with both changes against its average yield and reported this figure to the researchers.
The rollout. The plant then used the method on a different chip, one with a low yield nobody had been able to explain. The model pointed to two etching steps. Engineers changed both.
- Yield loss down 51.3%, comparing production in July and August 2020 with February to June 2020. The gain held for the whole period studied.
Why the researchers think it worked
- The usual tools missed it. The company confirmed that its standard methods had not found the two steps in the second product. A conventional statistical model, tried for comparison, found nothing either.
- The model’s output could be read by engineers. It named steps and machines. When the results were shown to production experts, they found the causes: one machine type let particles onto the wafers, and another scratched them while loading.
- The fix cost little. The weaker machines were moved to steps where they did no harm.
What the case shows
- AI on the factory floor can pay off without new equipment. The gain came from data the plant already recorded and from a change in routing.
- It needs a lot of clean history. The model learned from more than a thousand batches with thousands of recorded values each.
- The AI proposed; engineers decided and tested. The plant ran a controlled test on one batch before it changed anything in normal production.
What this says about the practices
Practices the company applied
- Start with one costly process, not a company-wide rollout The work was aimed at one product and one cost: transistor chips that fail the final test are scrapped. Do first · Effort: medium (weeks, a few people)
- Measure time saved per task against a baseline One production batch was split into four groups. One group was made the usual way, and the other three with one or both of the suggested changes. Do first · Effort: low (days, one person)
- Keep a person checking anything that goes to a customer or carries legal weight The model only pointed to where the losses came from. The plant’s process engineers supplied the possible actions, checked that they could be carried out, and worked out the physical causes. Do first · Effort: medium (weeks, a few people)
- Make the AI show what its answer is based on The researchers used a method that shows how much each process step contributed to the model’s prediction, and production experts used that to find the physical causes. Do next · Effort: medium (weeks, a few people)
Sources
- Senoner, Netland and Feuerriegel (2022). Using Explainable Artificial Intelligence to Improve Process Quality: Evidence from Semiconductor Manufacturing. Management Science, 68(8), 5704–5723. doi.org
- The same paper, free to read in the ETH Zurich Research Collection. doi.org
- ETH Zurich (15 December 2021). Improving quality through artificial intelligence. News article naming the plant as Hitachi Energy, Lenzburg. ethz.ch