ASML pilots AI for supply chain resilience

A pilot at Dutch high-tech company ASML showed that an AI machine trained based on Deep Reinforcement Learning makes smarter supply chain decisions than traditional planning methods. The challenge now is to make the suggestions generated by artificial intelligence understandable for planners and to incorporate them into IT systems, according to Head of Planning & Delivery Strategy Maarten Hendriks at a data science conference in the Netherlands.

By Harm Beerens

The event was organized by JADS, a data academy that stems from a collaboration between TU Eindhoven and Tilburg University. The theme of the event was ‘Make your supply chain triple R with data and AI’. “Triple R stands for Responsiveness, Robustness and Resilience: three things that are desperately needed for a company to deal with all the disruptions coming your way,” said Academic Director Jos van Hillegersberg in his welcome address. He named a whole host of data tools companies can use to achieve this triple R status. These ranged from building data spaces and digital twins, to doing simulations and deploying AI. A lot of knowledge is being developed and shared about all these things within JADS.

Last-minute changes

“When I heard what today’s theme was, I thought: that is exactly what we’re doing too,” said keynote speaker Hendriks (pictured) at the start of his talk. “Within ASML, we are dealing with disruptions and uncertainty in the market. Sales are growing steadily, but there are constant fluctuations in demand that are difficult to predict in advance. Customers order their machines well in advance, but changes can come right up until the last minute. To be able to respond optimally to this, we need to be resilient in the control of our inventory and assembly processes. In the department I head, we are continuously researching this and, of course, we are also looking at AI.”

AI bypasses bottlenecks

During an exploratory study by one of his department’s PhD students, they looked at whether AI could lead to higher resilience than a traditional way of planning. “This led to promising results,” Hendriks said. “We built an AI-controlled agent that uses Deep Reinforcement Learning to determine the optimal orders for each link in the production process.” The test showed that this leads to significantly higher delivery performance than with traditional ordering rules, and also with less buffer stock. The AI agent proved especially good at preventing potential bottlenecks in the assembly process, by cleverly bringing production orders forward slightly just in time.

Incorporate into planning process

According to Hendriks, the findings from the PhD research are extremely promising, but this does not mean that the current way of planning can now be jettisoned and all supply chain decisions will be made with AI from now on. “The challenge lies not in the quality of the decisions AI makes,” Hendriks claimed. “This research concerned a simplified model of reality, and I expect the performance improvement of the AI agent with our real supply chain data will be even higher. After all, you have to deal with even more complexity then. No, the challenge will lie mainly in how we incorporate this AI technology into the planning process and how we ensure that our planners embrace it.”

The latter is quite a tricky issue. While the AI agent’s ordering decisions were superior, no one understands exactly how they came about. Hendriks: “AI should be supportive of the ordering process, so planners should have confidence in what an agent comes up with. Planners should also always be able to override or modify an order suggestion. In addition, AI tools need to be integrated into the rest of the IT infrastructure we work with within ASML, and that too is easier said than done. In short, it will take some time before we can fully exploit the benefits of AI. But it will happen, I am sure.”

ASML wants scalable growth

For ASML, the research project around Deep Reinforcement Learning is just one of the improvement projects deployed to make the supply chain triple R. Also in other departments, such as the service supply chain, a lot is already being done with AI at ASML. But there is another important reason why he and his department will continue to invest heavily in AI in the near future, according to Hendriks: scalability. “So far, the number of supply chain planners has increased roughly in line with ASML’s turnover growth. We need to change that. We will continue to grow fast, but it is becoming increasingly difficult to get good people. I see AI as a tool to help planners do their jobs more efficiently, so that we can grow as a company in a scalable way.”