Feb/March_AMP_Digital

A D V A N C E D M A T E R I A L S & P R O C E S S E S | F E B R U A R Y / M A R C H 2 0 2 1 7 MACHINE LEARNING | AI IMPROVING MATERIAL PROPERTY PREDICTIONS By combining large quantities of low-fidelity data with smaller amounts of high-fidelity data, nanoengineers at the University of California San Diego developed a new machine learning method to more accurately predict ma- terial properties. The team believes their approach is also the first to predict the properties of disordered materials. Researchers looked at the band gaps of various materials as proof of concept. Their multi-fidelity graph networks led to a 22-45% decrease in the mean ab- solute errors of experimental band gap predictions, compared to a traditional single-fidelity approach. “There is no fundamental limita- tion as to what properties this can be applied to,” says Professor Shyue Ping Ong, whose team plans to use the new method to develop better materials for energy storage, photovoltaic cells, and semiconductor devices. “What we show in this work is you can actually adapt a machine learning algorithm to pre- dict the properties of disordered mate- rials. In other words, now we are able to do materials discovery and predic- tion across the entire space of both or- dered and disordered materials rather than just orderedmaterials. As far as we know, that is a first.” ucsd.edu. DEVELOPING SUPERHARD MATERIALS A machine learning mo- del developed at the Universi- ty of Houston (UH) and Man- hattan College, Riverdale, N.Y., can accurately predict the hard- ness of new materials. Super- hard materials—defined as hav- ing a hardness value exceed- ing 40 gigapascals on the Vick- ers scale—are rare, which makes identifying new materials chal- lenging, according to UH profes- sor Jakoah Brgoch. One of the complicating factors is that the hardness of a material may vary depending on the amount of pressure exerted, known as load dependence. That makes testing a material experi- mentally complex and using available computational modeling methods al- most impossible. The new model overcomes this hurdle by predicting the load-depen- dent Vickers hardness based solely on chemical composition. The team says the accuracy of their new model is 97%. “The idea of using machine learn- ing isn’t to say, ‘Here is the next great- est material,’ but to help guide our ex- perimental search,” says Brgoch. The researchers report finding more than 10 new and promising stable borocar- bide phases with work now underway to produce the materials for lab testing. uh.edu . Schematic of a multi-fidelity graph networks approach to more accurately predicting material properties. Courtesy of the Materials Virtual Lab at UC San Diego. Newmachine learning model accurately predicts the hardness of newmaterials. Courtesy of University of Houston. A multi-institutional team including researchers from the National Institute of Standards and Technology, Gaithersburg, Md., developed an artificial intelligence algorithm called CAMEO (Closed-Loop Autonomous System for Materials Exploration) that discovered a potentially useful new material without needing help from scientists. CAMEO determines which experiment to run on a material, conducts the experiment, and collects the data. It can also ask for more information such as crystal structure before running the next iteration. nist.gov. BRIEF

RkJQdWJsaXNoZXIy MjA4MTAy