Researchers Develop AI “Scientist” To Self-Drive Materials Discovery Experiments

A new AI-based method for efficient data collection could help scientists overcome complex challenges in materials discovery and design, allowing for greater precision and speed than ever before.

The method may also pave the way for “self-driving experiments,” where the intelligent algorithm can take a dataset and then define the parameters for the next set of measurements carried out. Doing so could enable the rapid discovery of new materials, the researchers say, allowing for greater advancements in combatting climate change, advancing quantum computing and accelerating drug design.

The research, published in npj Computational Materials, is the result of a collaboration between computer science and materials science researchers at Stanford University and the Department of Energy’s SLAC National Accelerator Laboratory.

Enabling the rapid discovery of new materials

The advent of high-power computers and computational modeling software capable of running large-scale molecular dynamics simulations was a huge boost to materials scientists. Suddenly, instead of spending thousands of hours of human manpower on manual trial-and-error experiments, it was possible to model these materials and simulate how they might behave using computers and deep theoretical calculations.

Despite this leap, materials discovery is still a very time-consuming and expensive process. The sheer scale of the possible materials out there is mind-bending – it is thought that with just four elements, it is possible to create over 10 billion possible materials.

For scientists interested in targeted materials discovery – i.e. the search for materials bearing certain desired properties – traditional discovery techniques are still very slow, especially if the researcher has a more complex goal than maximizing or minimizing one simple property.

In their new publication, the Stanford and SLAC researchers present a new approach that can satisfy more complex design goals, such as discovering the conditions to synthesize nanoparticles of a certain size, shape or composition. It can also learn and improve from each experiment it sees, using AI to suggest next steps based on the data it has read so far.

The approach is based on something called Bayesian algorithm execution (BAX), developed by study author Willie Neiswanger, who was a postdoctoral fellow in computer science at Stanford during the research period and is now an assistant professor of computer science at the University of Southern California.

Using this approach, a researcher can turn their complex design goal into a “recipe” or a “shopping list” type of filtering algorithm which is automatically translated into one of three BAX-based data collection strategies. This bypasses many of the time-consuming difficulties needed with previous methods, resulting in a process that excels in situations where multiple physical properties need to be considered.

"Our method allows you to specify complex objectives, enabling automatic optimization over a large design space, which increases the likelihood of finding new, amazing materials," said Sathya Chitturi, a PhD student at SLAC and Stanford who led the research. "The Bayesian algorithm execution framework lets you capture the intricacies of materials design tasks in an elegant and simple way."

Better materials for a better world

To demonstrate the use of this method, the research team applied it to a variety of custom goals for nanomaterials synthesis and magnetic materials characterization. The results suggest that this new method is significantly more efficient than other popular modern techniques, especially in complex scenarios.

"By combining advanced algorithms with targeted experimental strategies, our method makes the process of discovering new materials easier and faster,” said collaborator Chris Tassone, director of the Materials Science Division at the Stanford Synchrotron Radiation Lightsource (SSRL) at SLAC. “This can lead to new innovations and applications in many industries.”

Continue reading below...

Industry Insight

Researchers Want To Embrace AI But Still Have Major Concerns, Survey Finds

Read more

The team suggests that this method could make it more efficient to design new materials with specific catalytic properties that could improve the chemical processes used in manufacturing, making them more energy-efficient and sustainable while also producing less waste. Similarly, the method could be applied to create tailored drug delivery systems that can improve the targeting and release of therapeutics, which could result in improved efficacy and reduced side effects.

The researchers say that their new method is designed to be user-friendly and open-source, to allow diverse groups of scientists worldwide to easily use and adapt it to their own research.

While the research team continues to look for ways to incorporate this framework into more experimental and simulation-based research to further demonstrate its effectiveness, scientists from the SLAC’s Machine Learning Initiative (MLI) have begun to investigate its application in larger-scale materials simulations. Neiswanger, alongside collaborators from the MLI, has already published an additional paper outlining how the application of BAX could also help to optimize the performance of particle accelerators.

Reference: Chitturi SR, Ramdas A, Wu Y, et al. Targeted materials discovery using Bayesian algorithm execution. npj Comput Mater. 2024;10(1):156. doi: 10.1038/s41524-024-01326-2

This article is a rework of a press release issued by the SLAC National Accelerator Laboratory . Material has been edited for length and content.