In 2023, physicists at Sandia National Laboratories developed a technique to control LED light, which could potentially lead to replacing lasers with LEDs in technologies such as UPC scanners, holographic projectors, and self-driving vehicles. The team anticipated that refining this approach would require years of detailed experimentation.
Recently, the same researchers reported that artificial intelligence has significantly accelerated their progress. Using a trio of AI systems, they improved their previous results by a factor of four within five hours. Their findings were published in Nature Communications.
“We are one of the leading examples of how a self-driving lab could be set up to aid and augment human knowledge,” said Prasad Iyer, an author on the paper and the 2023 announcement from Sandia.
The research was funded by the Department of Energy’s Office of Basic Energy Sciences and Sandia’s Laboratory Directed Research and Development program. Part of the work took place at the Center for Integrated Nanotechnologies, operated jointly by Sandia and Los Alamos national laboratories.
The collaboration began when Saaketh Desai joined Sandia as a postdoctoral researcher. While Iyer specialized in optics, Desai brought expertise in machine learning. Together, they upgraded Iyer’s optics lab with AI-driven tools.
First, they used a generative AI model to process and simplify complex data sets. This streamlined data was then provided to an active learning agent—another AI—which was connected directly to laboratory equipment. The agent designed experiments based on its understanding, ran them using the equipment, analyzed outcomes, and iteratively refined its approach.
After 300 experiments conducted over about five hours, the system had substantially improved upon results previously achieved through years of manual research.
Despite initiating this AI-driven approach, Iyer expressed initial reservations about letting an automated agent control laboratory equipment. “We could potentially do infinite nonsensical experiments without having any meaningful results,” he said.
This concern relates to what is known as AI’s black box problem: while AIs can generate answers from inputs, it is often unclear how those answers are produced—a challenge for scientific transparency and reproducibility.
Desai emphasized that their goal was not just automation but advancing domain understanding with interpretable results. “We are constraining ourselves to finding good experiments that will advance our understanding of the domain,” he said. “Therefore, there is a high emphasis on interpreting why something worked or didn’t work.”
To address interpretability concerns, Iyer and Desai introduced a third AI system trained specifically to derive equations explaining observed data trends. This equation-learning AI worked alongside the active learning agent in a feedback loop: as new experimental data emerged, it sought mathematical formulas fitting those results.
By integrating these approaches, researchers quickly obtained equations verifying that their method allowed for steering spontaneous emission—the light produced by LEDs—on average 2.2 times more effectively than before across a 74-degree angle range; improvements at certain angles reached fourfold increases.
Unexpectedly, the AI discovered methods based on new principles regarding nanoscale interactions between light and materials—approaches not previously considered by the team.
While Desai described these advances as promising for scientific discovery more broadly, he noted that substantial computing resources were required for their work; their data analysis relied on a Lambda Labs workstation equipped with three NVIDIA RTX A6000 GPUs.
