ALBUQUERQUE, N.M. — In 2023, a team of physicists at Sandia National Laboratories announced a major discovery: a way to control LED light. If improved, lasers could someday be replaced by cheaper, smaller, and more energy-efficient LEDs in countless technologies, from UPC scanners and holographic projectors to self-driving cars. The research team believed that it would take years of careful experimentation to hone the technique.
Now, the same researchers report that three artificial intelligence labmates improved their best results by a factor of four. It took about 5 hours.
The resulting paper, published in Nature Communications, shows that AI is evolving beyond just an automation tool into a powerful engine for clear and understandable scientific discovery.
“We are one of the leading examples of how self-driving labs can be established to support and augment human knowledge,” said Sandia’s Prasad Iyer, author of the new paper and 2023 announcement.
This research was funded by the Department of Energy’s Office of Basic Energy Sciences and Sandia’s Institute-Directed Research and Development Program. Portions of this experiment were conducted at the Center for Integrated Nanotechnology, a DOE Office of Science user facility jointly operated by Sandia National Laboratories and Los Alamos National Laboratories.
Researchers team up across disciplines to modernize labs
A lucky coincidence spurred the research. Mr. Iyer has a new officemate.
Sarketh Desai came to Sandia as a postdoctoral fellow. While Iyer is an expert in optics, Desai was familiar with machine learning, a type of artificial intelligence, and was experimenting with ways to use it in scientific research. Together, they modernized Ayer’s optics laboratory.
First, we used generative AI models to learn and simplify complex data. They then fed this simple data set to a second AI called an active learning agent, which connected it to an optical instrument. They asked them to design an experiment based on the data they learned, run it on the device, analyze the results, and come up with a new experiment based on the results, repeating the process.
300 onwardsth The experiment took about five hours, but was a huge improvement over what researchers had spent years developing.
Addressing the AI black box problem
Iyer said it was his idea to involve Desai in the project, but he had some concerns about handing over lab equipment to an AI agent.
“You could do an infinite number of meaningless experiments without any meaningful results,” Iyer said.
Because AI has a black box problem. A query is entered, an answer is obtained, but it is often difficult for users to understand how the AI came up with the answer.
But science requires explanation. When scientists make a discovery, they share why they think the discovery works or makes sense. That’s the only way to move science forward. Because other scientists can test the idea and develop or disprove it.
Desai also recognized the importance of ensuring that AI-based conclusions are understandable.
“We strive to find good experiments that advance our understanding of the field,” he said. “So there’s a lot of emphasis on interpreting why something worked or didn’t work.”
Team prioritizes verifiable AI augmented research
Iyer and Desai agreed that AI automata alone will not be enough to advance their field. To deal with the black box problem, they introduced a type of fact checker. Perhaps unsurprisingly, it was a different AI. However, this individual was trained differently. The task was to find equations that explained complex data trends.
The researchers connected this third AI with the second AI in a loop. Active learning agents generate data and perform subsequent experiments, while equation learners try to devise equations that fit the data.
Immediately after the experiment was completed, the researchers had a new equation to validate that the autonomous driving lab had discovered a systematic way to control spontaneous emission, a type of light produced by LEDs, on average 2.2 times more effectively than previously achieved at a 74-degree angle. The best results at a given angle showed a 4x improvement.
Remarkably, the AI accomplished this in a way that Sandia’s team never thought possible. It was based on a fundamentally new way of thinking about how light and materials interact at the nanoscale.
Desai said the AI platform’s success is promising for science, but it also relies on a lot of computing power, which may not be available to all labs. Learning from data was powered by a Lambda Labs workstation equipped with three high-end NVIDIA RTX A6000 GPUs.
Still, Desai said he wants to see how far he can take the process. “As a next step, we are generally interested in using AI to arrive at interpretable optimization schemes and explainable decisions. We are interested in applying this to steering problems and other materials science problems in general.”
