AI gets a bad rap. And it’s easy to see why—a lot of the headlines we read tend to be about the robots taking human jobs, questionable levels of bot customer service, the amount of power it consumes, and the eventual moment when Skynet becomes self-aware and decides to do away with messy humans. It can be a bit hysterical.

What Does AI Do?

Artificial intelligence writes things. It makes pictures and videos. It summarizes documents. It can look through huge reams of information and pick out the answers to questions we pose it. But one of the major criticisms of AI in recent years has been its “black box” nature; it can sometimes come up with answers to questions without us being able to understand how or why.

What do we want AI to actually do? Can it help us discover new materials? Build better electronics? Use energy more efficiently? Help scientists solve problems that might otherwise take them a really long time?

That’s the idea behind “AI for Science,” in which AI is used as part of the scientific process itself. Researchers at NTT have demonstrated one way this could work.

Collaborating With AI

A recent experiment involved a semiconductor called beta-gallium oxide, or β-Ga₂O₃. It’s being studied as a potential material for future power semiconductors, which control and convert electrical power in electronic equipment. To make electronic devices from materials such as β-Ga₂O₃, researchers need to find ways of depositing them as very thin layers, or films. Ideally, the atoms in the film should form a single, orderly crystal structure, because this gives much greater control over the material’s electronic properties.

Sputtering

One way to produce thin films is “sputtering,” a process that can cover large areas at relatively low cost. Sputtering is essentially a way of coating a surface with an extremely thin layer of material: atoms are knocked from a source material inside a vacuum chamber and deposited onto another surface. The problem was that until now nobody had managed to use sputtering to produce a single-crystal β-Ga₂O₃ film.

NTT chose this as a test for its AI for Science approach.

Finding the best way to make the film meant getting several conditions right at the same time. In NTT’s experiment, there were four things to consider: temperature, sputtering power, and the flow of argon and oxygen gases into the chamber. Argon helps knock atoms from the source material so they can settle on the surface and form the film, while changing the amount of oxygen entering the chamber can affect the quality of the resulting film.

Research Is Complicated

Change one and you might affect the result. Change several together and the possibilities multiply. A researcher can conduct an experiment, measure what happened, choose another combination and try again. Then again. And again. Get the idea? It’s a potentially long and repetitive process.

NTT decided to hand much of that process to AI and automation.

… But AI Can Help

NTT’s system made a film, automatically measured its quality, then used AI to study the result and choose which combination of settings to try next. It made another film, checked that one, learned from the result and tried again. Each experiment helped the AI decide what to do in the next one. And it turned out that the system was able to carry out the deposition and evaluation cycle about three times faster than conventional experiments run by engineers and researchers.

After 56 experiments, it had found conditions that enabled NTT to produce high-quality β-Ga₂O₃ films. NTT believes this is the world’s first single-crystal β-Ga₂O₃ thin film made using sputtering, a process suited to large-area, lower-cost production.

Which is great. But finding the right recipe was only part of NTT’s research. In general, there’s a problem with letting AI search through combinations until it finds one that works. The machine might be able to give you the best settings, after many iterations, but it may not be able to tell you why those settings worked. That’s when we come back to the black box problem.

The Black Box Problem

A new method discovered for one piece of equipment might not work when the equipment changes. Science depends on humans being able to understand relationships: which factors matter, how they affect one another and what you can learn from an experiment and then apply somewhere else. That’s why NTT then analyzed the data generated during the AI experiments using another machine-learning technique called a “random forest.” A random forest looks for patterns in data and helps identify which factors had the greatest effect on the result.

Doing that analysis showed what was going on inside the recipe. It revealed how the four experimental parameters affected the quality of the film and identified one relationship that deserved particular attention: the interaction between temperature and oxygen flow. From this, NTT researchers were able to come up with a rule they could understand and follow themselves: adjust the individual parameters in sequence, then concentrate on fine-tuning temperature and oxygen flow together.

So they tried it. When the researchers took what the AI had learned and refined the process themselves, they produced an even higher-quality film.

Understanding What AI Can Do

Here’s one thing machines can do better than us: they can work their way through experimental possibilities without some of the assumptions we might bring to the problem. They can collect data as they go, then help us identify patterns in the results and turn them into rules we can understand, reproduce and potentially use with other equipment.

And doing it this way means that humans remain part of the process. In NTT’s work, the knowledge gained from the AI-led experiments gave researchers a way to improve the result themselves.

Modern life relies on materials developed through years of research. Semiconductors, communications equipment, power electronics and many other technologies began with scientists trying to understand how materials behaved and how they could be improved. But experiments take time and when several variables interact, the number of possible combinations can become enormous. And not every possibility is worth pursuing.

An AI-driven laboratory can carry some of that load and help scientists understand what the search has revealed, so that discoveries made in one set of experiments can become knowledge that humans might use for the next.

NTT plans to broaden its AI for Science approach with other materials and equipment, including semiconductor, oxide and quantum materials, automating more of the experimental process and developing AI models that can use existing knowledge about materials and deposition processes.

What do we actually want AI to do for us?

Helping scientists run more experiments, understand what those experiments have taught us and use that knowledge to make something better is a pretty good start.

Innovating a Sustainable Future for People and Planet

For further information, please see this link:
https://group.ntt/en/newsrelease/2026/08/25/260825a.html

If you have any questions on the content of this article, please contact:

Public Relations
NTT Science and Core Technology Laboratory Group
https://tools.group.ntt/en/news/contact/index.php

Daniel O'Connor

Daniel O'Connor joined the NTT Group in 1999 when he began work as the Public Relations Manager of NTT Europe. While in London, he liaised with the local press, created the company's intranet site, wrote technical copy for industry magazines and managed exhibition stands from initial design to finished displays.

Later seconded to the headquarters of NTT Communications in Tokyo, he contributed to the company's first-ever winning of global telecoms awards and the digitalisation of internal company information exchange.

Since 2015 Daniel has created content for the Group's Global Leadership Institute, the One NTT Network and is currently working with NTT R&D teams to grow public understanding of the cutting-edge research undertaken by the NTT Group.