New AI Tool Can Help Researchers Understand Quantum Materials

Researchers at the Department of Energy’s SLAC National Accelerator Laboratory have developed a new machine learning tool that is able to interpret the swinging of atomic spins in a system, enabling more efficient studies of complex quantum materials.

Published in Nature Communications, the study details the creation and training of the new artificial intelligence (AI) platform using “neural implicit representations,” a machine learning tool that derives unknown parameters from experimental data. This approach has already been implemented in other scientific fields, such as medical imaging and cryo-electron microscopy.

The researchers say that their new tool could help to accelerate the investigation of novel materials, as it helps automate some of the more labor-intensive experimental data analysis within materials research. 

An alternative approach to AI assistance

To understand the rules governing unusual systems, such as magnetic materials, researchers often want to examine their collective excitations. In simple terms, collective excitations are phenomena that arise at very small scales, in which a solid behaves more like a set of weakly interacting particles in a vacuum. One example of this might be tiny changes in the pattern of atomic spins in a material, which can affect its magnetic properties. Understanding such properties is key for the development of novel technologies, such as advanced spintronics devices that have the potential to revolutionize how we send and store data.

Current methods for studying collective excitations are incredibly intricate and often require a significant investment of resources. Excitation spectra are normally obtained using inelastic neutron or X-ray scattering techniques. These are analyzed by comparing experimental results against calculated predictions, but a paucity of available neutron sources has many reconsidering how such experiments can be made more efficient.

To reduce the extensive wait times for quantum materials analysis at neutron scattering facilities, there is a real appetite for new techniques for real-time modeling and analysis of experimental spectral data.

Some headway has been made with the application of conventional machine learning algorithms in speeding up the collection and processing of neutron scattering data. In a bid to improve upon these efforts, the researchers from the SLAC National Accelerator Laboratory set out to construct a similar system that makes use of neural implicit representations, rather than traditional learning algorithms.

Neural implicit representations “see” inputs as being coordinates on a map. For imaging applications, where a traditional image-based learning system would store an image directly, neural implicit representations create a recipe for how to interpret the image by connecting a pixel’s coordinate to its color. With enough of these data points, such a model can learn to make very detailed predictions, even between pixels in an image.

“Our motivation was to understand the underlying physics of the sample we were studying. While neutron scattering can provide invaluable insights, it requires sifting through massive data sets, of which only a fraction is pertinent," said co-author Alexander Petsch, a postdoctoral research associate at SLAC’s Linac Coherent Light Source (LCLS) and Stanford Institute for Materials and Energy Sciences (SIMES).

“By simulating thousands of potential results, we constructed a machine learning model trained to discern nuanced differences in data curves that are virtually indistinguishable to the human eye,” he said.

Predictive algorithm can assist with analysis

The research team wanted to design an AI tool that could interpret the measurements made by the team at the LCLS, and recover the microscopic details of the material in near real-time as it was measured.

After thousands of simulations covering a range of different parameters that would normally be used to study quantum materials, the researchers fed this spectral data into a machine learning algorithm. This step was done so that the team could predict answers from theory as soon as they measured real spectra.

For the real spectra component, they fed in inelastic neutron scattering data that had been generated by Petsch as part of his doctoral thesis.

When the team applied their machine learning model to the real-world data, they found that it was able to successfully overcome a number of crucial challenges with data analysis, including background noise and missing data points. The researchers also demonstrated how their approach could be used to continuously analyze data in real time – highlighting the potential of such a technique to streamline current materials analysis approaches.

"Our machine learning model, trained before the experiment even begins, can rapidly guide the experimental process," said SLAC lead scientist Josh Turner, who oversaw the research. “It could change the way experiments are conducted at facilities like LCLS.”

Crucially, the researchers say that this machine learning model design is not just limited to inelastic neutron scattering experiments. This technique could also help bypass the need for complex peak-fitting algorithms or user-intensive post-processing in other scattering measurements.

“Machine learning and artificial intelligence are influencing many different areas of science,” said co-author Sathya Chitturi, a PhD student at Stanford University. “Applying new cutting-edge machine learning methods to physics research can enable us to make faster advancements and streamline experiments. It's exciting to consider what we can tackle next based on these foundations. It opens up many new potential avenues of research."

Reference: Chitturi SR, Ji Z, Petsch AN, et al. Capturing dynamical correlations using implicit neural representations. Nat Commun. 2023;14(1):5852. doi: 10.1038/s41467-023-41378-4

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