Machine learning used to discover plasmonic metals

9/16/2022

Jenny Applequist

Materials scientists have been on the hunt for plasmonic metals — ones whose electrons oscillate when light strikes their surface — because of their importance in a range of applications, such as nano-antennas, subwavelength imaging and biosensors. But identifying plasmonic metals is no easy task, in part because of the immense numbers of potential candidate materials. Calculation of the optical properties of even one simple material can consume 100 core hours on a supercomputer. So how can researchers hope to assess the suitability of thousands of metals?

A new paper published in the Advanced Optical Materials journal offers a big shortcut: a machine-learning solution that can sift through vast databases of materials data to find small sets of promising candidate materials. Furthermore, it can do so on an ordinary laptop in just seconds.

The work takes advantage of a recent development in materials science: the emergence of online databases that provide extensive data on the properties of tens or even hundreds of thousands of materials.

André Schleife
André Schleife 

Co-author André Schleife says, “We don’t necessarily need to run all simulations ourselves because we can instead rely on data provided by others through those databases.” Unfortunately, the databases aren’t ideal: “They have a certain type of data in them, or they have different types of data in them — and it’s not necessarily always precisely what we need for a given type of research.” 

So how can one extract the needed insights from these diverse and imperfect databases? To do so, Schleife and his co-author, Ethan Shapera, created an approach in which a simplified, machine-learning-based model does an initial screening pass of database contents to identify a small number of potentially suitable materials. Then, simulations based on density functional theory (DFT) — a popular method for investigating materials’ electronic structures — are used for more careful assessment of the identified materials.

The enormous advantage of their approach is that crude results can be obtained very fast, dramatically shrinking the pool of materials that require further consideration. Because the models are simplified, Schleife says, the results “are never exact and are approximations.” Hence it is crucial to scrutinize the selected materials more closely through the subsequent use of more accurate techniques, such as DFT. 

As the paper reports, Shapera and Schleife successfully used their new approach to identify three specific materials — AlCu3, ZnCu, and ZnGa— as “excellent potential new plasmonic metals.” They did so via detailed DFT analysis of the electronic structure and optical properties of the materials found by the machine-learning screening.

Ethan Shapera
Ethan Shapera

Shapera says it was when the constructed machine-learning models started to match DFT results that he realized the approach was a success. “The persistent challenge in machine learning is to construct models which are able to learn complex patterns, and then to correctly apply the discovered patterns to making predictions. When I compared the machine learning model’s prediction of the plasmonic quality factors — which are the criteria we used to judge how good a material is — versus the DFT calculated, the machine learning model agreed extremely well. That was our indication that this approach we developed is viable.” 

The code and data produced in this work have been published on the Materials Data Facility website, from which interested researchers can obtain them. 

Schleife is an associate professor in Materials Science & Engineering, MRL and NCSA and was Shapera’s Ph.D. advisor. Shapera graduated from UIUC in 2022 and is now a postdoc at the Graz University of Technology in Austria, where he is developing approaches to accelerate material design by combining high-throughput calculations with machine learning.

The paper is Ethan P. Shapera and André Schleife, “Discovery of New Plasmonic Metals via High-Throughput Machine Learning.” It is available at https://doi.org/10.1002/adom.202200158