Episode 2: AI steers autonomous x-ray scattering experiments

Prachi Patel of MRS Bulletin interviews Kevin Yager and Masafumi Fukuto of Brookhaven National Laboratory about an artificial intelligence algorithm they designed that analyzes data and then decides what should be measured next. In their first autonomous experiment, the researchers used x-ray scattering to map the boundaries of a droplet where nanoparticles segregate. Read the article in Scientific Reports.

PRACHI PATEL: Discovering new materials takes an enormous amount of time. You make a material, measure its properties, analyze the data, and then repeat the process all over again. Automation has sped things up. But now scientists have made this automation smarter. In a recent paper published in the journal Scientific Reports, researchers presented an artificial intelligence algorithm that can analyze data and then decide what to measure next. Here’s Kevin Yager, a scientist at Brookhaven National Laboratory. 

KEVIN YAGER: I think its important to make a distinction between automation and autonomous. It’s autonomous in the sense that you tell it a goal and it, you know, starts conducting the experiments and updating its experimental plan on each iteration.  

PATEL: The goal is to speed up every step in the materials discovery process, improve those steps, and couple them better to each other. And eventually, make the entire experimental workflow autonomous.  

YAGER: Not to replace the human experimenter but really to liberate the scientist to think about the data at a higher level because the tool is automatically making decisions about what to measure, doing that measurement, and then updating its experimental plan in a loop. So the human can think about the meaning of the data as its being collected and intervene as necessary. 

PATEL: The researchers start by defining a set of goals for their experiment. The algorithm then works in a multidimensional parameter space. Those parameters can be things like material composition, temperature, and pressure. And the algorithm explores how material properties vary throughout that space, Yager says. 

YAGER: The algorithm treats it as a very abstract mathematical problem. Which is saying ok I have some data points in this space and what I’m going to do is I’m going to interpolate between the existing data to create what’s called a surrogate model that sort of tries to represent the data. And then along with that surrogate model I can compute a corresponding uncertainty. So how certain or uncertain my model is across that space. Where I’ve measured a lot of data my model is pretty certain. Where I’ve measured not very much data my model is very uncertain. So the algorithm essentially says wherever my uncertainty is high, that’s probably where I should measure next because I’m going to gain the most information. 

PATEL: For their first autonomous experiment, the team used x-ray scattering to map the boundaries of a droplet where nanoparticles segregate. They compared the standard approach with the new AI algorithm explains Masafumi Fukuto, a scientist at Brookhaven and co-author on the paper. 

MASAFUMI FUKUTO: The first test that we did was to compare a simple grid search, a grid scan of this material as a function of spatial coordinates vs AI-driven search of these spatial coordinates. We found features like the boundaries of this heterogeneous material much more quickly than you do with a simple grid scanning method. 

PATEL: The algorithm could be applied to any other materials research and discovery method. 

FUKUTO: The brain part, the AI part, the decision algorithm part is completely independent of the technique that you use. 

PATEL: This is Prachi Patel for MRS Bulletin’s Materials News Podcast.