TACC resources are helping NIST develop a neural network for materials prediction


February 15, 2022 – When Tony Stark must travel to space in the original Iron Man film, he asks his artificial intelligence (AI) assistant, JARVIS, to craft a suit that can withstand harsh conditions.

As AI specialist Kamal Choudhary explains, “The way I see it, JARVIS created a materials database, scanned the database, found a suitable material, tested it, and then synthesized an alloy that can withstand space conditions.

“That’s what we want from our system, and that’s why we called it JARVIS.”

For a given material performance metric, multiple JARVIS components can work together to develop optimized or entirely new materials. Ref: npj computer materials

Choudhary, a researcher at the National Institute of Standards and Technology (NIST), is the founder and developer of JARVIS (Joint Automated Repository for Various Integrated Simulations) – an open data set for automating materials discovery and optimization.

registered mail Computational Natural Materials in December 2021, Choudhary and Brian DeCost (NIST) described the latest improvements to JARVIS that apply AI to accelerate discovery. Their Atomistic Line Graph Neural Network (ALIGNN) combines graphical neural networks with chemical and structural knowledge of materials and outperforms previously reported models for atomistic prediction tasks with very high accuracy and better or comparable model training speed.

“ALIGNN can predict features in seconds instead of months,” Choudhary said.

Along with the inspiration of Iron Man, there was the Materials Genome Initiative. Launched in 2011 under President Obama, the initiative is a multi-federal initiative to discover, manufacture and deploy advanced materials twice as fast and at a fraction of the cost of traditional methods.

NIST’s original contribution to this initiative was the creation of a database of materials and their properties, which were rigorously created using standardized, state-of-the-art computational methods.

Several such databases have been set up, but “what’s unique about the JARVIS database is that it contains modules for different types of computational approaches,” according to David Vanderbilt, a professor of physics at Rutgers University, a member of the National Academy of Sciences, and a contributor on the project. “There are many different theoretical levels at which to approach the field. JARVIS is unusual in that it has more layers than other databases.”

ALIGNN predicted some examples of CO₂-adsorbing MOFs from CoREMOF DB. Photo credits: Kamal Choudhary, Taner Yildirim, Daniel Siderius, A Gilad Kusne, Austin McDannald, Diana L Ortiz-Montalvo

The original data for JARVIS were drawn from density function theory (or DFT) calculations. “DFT is the standard method most people use to calculate the properties of a material at the atomic level,” Vanderbilt explained. “They are first-principal calculations, where there is no experimental input and the results are derived from theory from scratch according to the laws of quantum mechanics.”

This paradigm was incredibly effective, “but if you look at the periodic table, there are billions of possible combinations of elements — more than we can ever generate data,” Choudhary said. “This is where machine learning comes in.”

If quantum mechanical calculations can serve as a screening tool for physical experiments, Choudhary argued, then machine learning can serve as a screening tool for expensive calculations.

But first such a system must be trained. Neural networks like ALIGNN require huge amounts of training data to be effective. Choudhary’s state-of-the-art AI model is powered by DFT simulations of 70,000 materials and counting. This growing database was used to train the neural network, which in turn can quickly characterize new materials or search for materials with specific properties.

“It’s the Materials Genome Initiative’s dream come to life,” said Choudhary.

Writing in arXiv, Choudhary and his collaborators provided an example of how the system can speed discovery. They used ALIGNN to predict the CO₂ adsorption properties of metal organic frameworks, a class of porous materials that can remove CO₂ from the atmosphere, and to computationally rank leading candidates for experimental synthesis.

The JARVIS dataset was generated primarily on supercomputers at NIST, which have been working on this effort for almost five years. More recently, Choudhary gained access to the Frontera and Stampede2 supercomputers at the Texas Advanced Computing Center (TACC), which also contributed to the dataset.

“The field of machine learning has been around since the 1980s, but the main problem has been well-curated datasets,” Choudhary said. “We are now approaching 100,000 materials in our database and this was only possible thanks to Frontera and NIST. That helped us close that gap.”

With a large number of available training examples and knowledge of chemistry and physics hard-coded into the neural network, Choudhary was able to significantly improve the accuracy of his machine learning model. “The more domain knowledge you can leverage, the better. I think physics and AI shouldn’t be competitors; they should be friends and collaborators.”

The ALIGNN tool, as well as those for DFT calculations and other machine learning methods, will be integrated into JARVIS and made available to researchers worldwide. Choudhary estimates that 8,000 chemists and biologists use the repository each year. Recently, it has enabled scientists at Argonne National Laboratory to study topological magnetic materials and helped researchers at Northwestern University study transfer learning for materials.

Schematic of a non-directional crystal diagram representation and corresponding line diagram construction for a SiO4 polyhedron. Ref: npj computer materials

Choudhary is also collaborating with David Vanderbilt to develop “Beyond-DFT” methods, apply them to quantum materials, and integrate these methods and datasets into JARVIS.

“DFT contains some significant approximations,” said Vanderbilt. “Because electrons are treated as independent, they miss some of the very peculiar and interesting behaviors in quantum materials that lead to effects beyond the normal expectation of ordinary theory.”

These include, but are not limited to, unconventional superconductivity, the quantum Hall effect, and topological magnetic structure. “Ordinary DFT doesn’t work well enough for these classes of materials,” he continued. “Our database uses three or four higher Beyond DFT approaches to give the community a sense of how the answers can differ depending on the underlying approach.”

By establishing a database of possible materials and developing tools to automate the screening, Choudhary hopes to accelerate the discovery pipeline and bring Iron Man-like abilities closer to reality.

“Imagine the day when a model that can predict a new material, a new drug — and says, ‘Out of a million molecules, try this one first,'” Choudhary said. “This is the golden age of materials science.”

Source: Aaron Dubrow, TACC


Comments are closed.