Artificial Intelligence (AI) is a game changer both for industry and society. To ride the wave of this technological revolution, High Tech Campus Eindhoven in cooperation with NXP, Signify, ASML and Philips has recently opened the AI Innovation Center  on Campus. In this 1100 square meter hub organisations can develop strategies, talent, knowledge and skills in the vastly evolving field of AI. In this interview series managers from the four founding companies share their insights on the role of AI both in their companies and the world at large. Part 2: Arnaud Hubaux, Senior Technical Program Manager at ASML.

❕ On April 22nd, Arnaud will be one of the keynote speakers during the AI Leadership Forum. See the entire program and register here!

In 2019 the chipmaking industry churned out more than 634 billion computer chips, generating a global turnover of 412 billion euro’s. Although the companies that design and produce the chips, like Intel, Samsung and NXP, are known all over the world, the same cannot be said about the company that makes the incredibly complex machines on which these chips are manufactured. We’re talking of course about ASML, the multinational Brainport-based company that with a sense of irony describes itself as ‘the most important tech company you’ve never heard of’.

Yet ASML’s lithography machines are the backbone of the chipmaking world. Their incredibly accurate lasers, used by large chip manufacturers, print chip patterns onto silicon wafers with resolutions up to 13 nanometers (a nanometer is a millionth of a millimeter). Although ASML is primarily a hardware company, software has become increasingly important to them over the years. This includes artificial intelligence (AI) and machine learning (ML), which the company uses to improve machine performance.

As a senior program manager Arnaud Hubaux drives the adoption of AI at ASML. Hubaux introduced machine learning solutions to detect defects during production and optimise the behaviour of their lithography machines. “We use AI as a complementary technology to our product portfolio, which helps our customers get the most out of their chip manufacturing processes,” he says.

Firstly, can you tell us why ASML decided to co-invest in the AI Innovation Center?

Hubaux: “Many people don’t realise that ASML’s chipmaking machines rely extensively on software to produce and maintain nanometer precision. That includes machine learning and artificial intelligence solutions. The AI Innovation Center is a gateway for us to talk about the challenges of what AI means in that context. Also, with the start of this Center I’m excited that students fresh out of university can immediately learn to apply AI at a scale where a manufacturing process can rely on it. That’s a completely different ballgame than just knowing the mathematical background. The Center is also a way of promoting what we do in the ‘AI for Good’ scope and giving back to the community.”

Since ASML’s technology is driven primarily by physics — mathematical formulas and models lie at the heart of their lithography systems — adding AI or ML to their toolset wasn’t a foregone conclusion, says Arnaud. Yet he saw a possibility for algorithms to optimise the performance of their machines in order to improve the ‘yield’, the number of good wafers produced per day.

Optimisation is crucial, since lithography processes can malfunction and ‘drift’. In a business where every nanometer counts, process drift is a potential showstopper, which is why ASML engineers are always monitoring and recalibrating machines to perform up to par. “We tried to push the optimisation with our standard physical model,” Hubaux says, “but we were limited by the fact that we didn’t know what was happening in the customer environment.”

Since protection of intellectual property (like chip designs and manufacturing processes) is top priority for chipmakers, each lithography machine operates in a black box environment, completely off-grid and unconnected to any cloud or internet. Hubaux: “We provide all the knobs the customer can turn. Customers configure our machines to their specific production process. So we’re like a car dealer selling a car and knowing exactly how it can perform, but the customer actually drives that car.”

Without knowing how the machine is configured and used by their customers, it’s all but impossible to achieve full optimisation. That’s where machine learning came in handy, says Hubaux. “For optimisation you have to understand how the domain works. But for us the customer domain is a no-go area. So if I cannot go there myself, I need someone to tell me when I’m right and when I’m wrong. That’s where machine learning imposed itself as the solution, because it is designed to address this kind of issue. The algorithm was able to learn the parts that we didn’t know, which really helped us to further push the boundary of optimisation.”

But how can you train your algorithm if you do not have actual manufacturing data from the customer?

Hubaux: “Yes, that’s the interesting part. The whole essence of machine learning is data, that’s how you train the model. In our case the data is owned and protected by the customer, it’s their life blood and competitive edge. We have to make clear agreements about that with customers. After all, this intellectual property is the competitive catalyst for the industry. For machine learning it means that we cannot do what most solutions do, which is to train your model first and then deploy it. Instead we’re using a completely untrained model, like a baby that cannot read or write.”


So how does an untrained model learn?

"Our algorithm only needs to know whether each step of the printing process is performed correctly or not. That pass/fail label is supplied to us by the customer. Based on that feedback alone we can tune and train the model, so it slowly adapts its behaviour for the specific context of that customer. Then we can do the finetuning on our machines and help the customers optimise their manufacturing process, without knowing anything about the chip design or manufacturing process itself.”

Apart from machine optimisation ASML is also using AI to detect defects during visual inspections of the wafers. Algorithms can be trained to spot anomalies in high resolution images of parts of the wafer very efficiently. Machine learning is also helping to correct deformations in the processing of laser images on the wafer. Hubaux says these can all be done in-house with ASML’s own data.

According to Hubaux the biggest challenge in applying AI in business is industrialisation or ensuring a solution always performs well at customer locations. And for that, Hubaux says, explainability or being able to reason exactly why the algorithm is performing like it does, whether good or bad, is crucial. “Machine learning is a bit of a black box, you don’t know exactly what goes on under the hood. That is why it is extremely important to always be able to explain what is going on inside the model and link it to physics. We invest in these solutions, but we also invest in explaining how they work and behave.”

Hubaux is critical about hailing AI as the saviour of innovation. Algorithms are useful, but they also have their limits, he says. “For an innovative semiconductor equipment maker like us, you know that new generations of products and systems will never solely depend on AI, because AI needs data to perform and the data is not even there yet. The roadmap for innovation is driven by customer demand and realised by our multidisciplinary engineering teams around the world. So AI will never be the gatekeeper to produce new generations of lithography systems. But AI does help us to squeeze every possible bit of performance out of our existing machines.”

On April 22nd, Arnaud will be one of the keynote speakers during the AI Leadership Forum. See the entire program and register now!

Receive the latest insights, news & events

Subscribe