Argonne Lab’s ‘AI-NERD’ predicts material behavior with unprecedented accuracy

What if scientific breakthroughs could happen without a scientist running every experiment? A new AI development at Argonne National Laboratory takes a concrete step toward that vision, “marking a concrete step towards autonomous materials discovery,” as a recent Nature Communications’ paper notes. The field of material science is facing something of a quandary as the demand for new, better-performing materials in fields ranging from renewable energy to aerospace is outpacing researchers’ ability to discover and characterize them with traditional methods.

Traditionally, understanding how materials change at the atomic level has relied on painstakingly slow and labor-intensive experiments. Observing such microscopic dynamics in action is proven challenging. While techniques like X-ray photon correlation spectroscopy (XPCS) offer a glimpse into atomic behavior, the data they generate are incredibly complex.

Introducing AI-NERD

Enter AI-NERD (Artificial Intelligence for Non-Equilibrium Relaxation Dynamics), a novel technology from Argonne National Laboratory that promises to accelerate materials research. This approach leverages unsupervised deep learning, specifically an autoencoder neural network, to analyze complex X-ray data and unveil the hidden “fingerprints” of material behavior. By teaching itself to recognize patterns in XPCS data without expert training, AI-NERD creates condensed material “fingerprints” from intricate X-ray scattering patterns. This breakthrough enables researchers to map and analyze material behavior in previously impossible ways.

“The goal of the AI is just to treat the scattering patterns as regular images or pictures and digest them to figure out what are the repeating patterns. The AI is a pattern recognition expert,” said James (Jay) Horwath Argonne National Laboratory, the first author of the study in a press release.

Decoding material genomes

These fingerprints are more than just patterns—they are condensed representations of a material’s structure and behavior, distilling volumes of  of XPCS data into essential features. As Horwath explained, “You can think of it like having the material’s genome, it has all the information necessary to reconstruct the entire picture.”

What sets AI-NERD apart is its ability to learn and identify patterns without expert guidance. This unsupervised learning approach allows the system to discover hidden relationships and trends in material behavior that might elude scientists. By processing and categorizing X-ray scattering images, AI-NERD creates a comprehensive map of material dynamics, offering researchers a new lens for viewing atoms and molecules.

In the research, the scientists used a technique called Uniform Manifold Approximation and Projection (UMAP) to transform their complex dataset into a simple two-dimensional picture, as shown in the visual above. UMAP is similar to another popular method called t-distributed Stochastic Neighbor Embedding (tSNE). Both of these methods try to preserve the relationships between data points when reducing dimensions. For a brief overview of word embeddings, check out the article “What embeddings are and how to explore them in R&D.”

The broader landscape of AI in materials science has picked up steam in recent years

This development in AI-assisted materials research at Argonne National Laboratory is part of a broader trend of artificial intelligence revolutionizing materials science. In recent years, several notable advancements have paved the way for AI-NERD’s development.

In 2023, researchers made significant strides in using AI for materials synthesis and characterization. For instance, one team developed an AI system capable of extracting “recipes” for producing materials from scientific papers. This system could identify correlations between precursor chemicals and resulting crystal structures, streamlining the material discovery process.

Another group created an AI system that recognizes patterns across different materials recipes. This innovation allows the AI to suggest alternative recipes for known materials, potentially opening up new avenues for synthesis.

While machine learning continues to be something of a trending subject, its use in material science is not new. In 2018, researchers at Virginia Tech developed a machine learning framework that trains “on the fly” to accelerate the development of computational models for materials design.

Similarly, AI system called ARTIST from Aalto University and the Technical University of Denmark debuted in 2019 to instantly determine how a molecule will react to light, potentially accelerating the development of flexible electronics and other technologies.

In nuclear materials research, University of Wisconsin-Madison and Oak Ridge National Laboratory unveiled an AI system was trained to detect and analyze microscopic radiation damage in potential nuclear reactor materials in 2018. It outperformed human experts in both accuracy and speed.

AI-NERD’s potential

In particular, the development of AI-NERD promises to boost the analysis of XPCS data in particular. As the upgraded Advanced Photon Source comes online, generating 500 times brighter X-ray beams than its predecessor, the need for efficient data processing becomes even more critical. “The data we get from the upgraded APS will need the power of AI to sort through it,” Horwath emphasizes. AI-NERD’s ability to create material “fingerprints” and identify patterns in large datasets could advance researchers’ understanding of material dynamics for an array range of applications.