The new approach determines the design of optimal materials with minimal data

Northwestern University researchers have developed a novel computational approach to accelerate the design of materials exhibiting metal-insulating transitions (MITs), a rare class of electronic materials that have shown the potential to initiate future design and delivery of microelectronics and systems. faster quantum information: core technologies behind large-scale Internet of Things devices and data centers that power the way humans work and interact with others.

The new strategy, a collaboration between Professors James Rondinelli and Wei Chen, integrated statistical inference techniques, optimization theory and physics of computational materials. The approach combines multi-objective Bayesian optimization with latent variable Gaussian processes to optimize ideal characteristics in a family of MIT materials called complex lacunar spinels.

When researchers look for new materials, they typically look in places where existing data on similar materials already exists. The design of many material property classes has been accelerated in existing works with data-driven methods aided by high-throughput data generation coupled with methods such as machine learning.

Such approaches, however, were not available for MIT materials, which are classified according to their ability to reversibly transition from an electrically conductive to an insulating state. Most MIT models are built to describe a single material, making model generation often challenging. At the same time, conventional machine learning methods have shown limited predictive capabilities due to the lack of available data, making it difficult to design new MIT materials.

“Researchers understand how to distill information from large datasets on materials where they exist and when appropriate characteristics are available,” said Rondinelli, professor of materials science and engineering and Morris E. Fine Professor in Materials and Manufacturing at McCormick School. of Engineering, and corresponding author of the study. “But what do you do when you don’t have large datasets or the necessary functionality? Our work breaks this status quo by building predicative and exploratory models without requiring large datasets or functionality from a small dataset.”

A paper describing the work, titled “Featureless Adaptive Optimization Accelerates Design of Functional Electronic Materials,” was published November 6 in the journal Review of Applied Physics.

The research team’s method, called the Advanced Optimization Engine (AOE), bypasses traditional machine learning-based discovery models by using a latent variable Gaussian process modeling approach, which requires only the chemical compositions of materials to discern the their optimal nature. This allowed AOE based on Bayesian optimization to efficiently search for materials with optimal tuning of the band gap (resistivity / electrical conductivity) and thermal stability (synthesizability), two characteristics that define useful materials.

To validate their approach, the team analyzed hundreds of chemical combinations using simulations based on density function theory and found 12 previously unidentified compositions of complex lacunar spinels that showed optimal functionality and synthesizability. These MIT materials are known for hosting unique spin textures, a feature necessary to power the future Internet of Things and other resource-intensive technologies.

“This advancement overcomes the traditional limitations imposed by material designs based on chemical intuition,” said Chen, Wilson-Cook Professor of Engineering Design and Professor and President of Mechanical Engineering and co-author of the study. “By re-formulating functional material design as an optimization problem, we have not only found a solution to the challenge of working with limited data, but we have also demonstrated the ability to efficiently discover optimal new materials for future electronics.”

Although the researchers tested their method on inorganic materials, they believe the approach could also be applied to organic materials, such as designing protein sequences in biomaterials or monomeric sequences in polymeric materials. The model also offers guidance on how to make better decisions towards optimal material design by choosing ideal candidate compounds to simulate.

“Our method paves the way for the optimization of multiple properties and the co-design of complex multifunctional materials where data and prior knowledge are scarce,” said Rondinelli.

The work on this study grew out of a project exploring Bayesian optimization in materials discovery as part of the Graduate School at Northwestern-sponsored Predictive Science and Engineering Design (PSED) interdisciplinary cluster program. It was supported by funding from the National Science Foundation and the Agency’s DIFFERENTIATE program for Advanced Research Projects (ARPA-E), which seeks to use emerging artificial intelligence technologies to address key energy and environmental challenges.

“This work highlights the impact of the collaborative PSED interdisciplinary design cluster,” said Chen. “It also highlights the crucial advances occurring in AI and machine learning at Northwestern in design and optimization.”