January_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 | J A N U A R Y 2 0 2 0 8 MACHINE LEARNING | AI DEVELOPING NEW MATERIALS VIA MACHINE LEARNING Researchers at Delft University of Technology, the Netherlands, report developing an ultra-compressible yet strong material—without conduct- ing any experimental tests and us- ing only artificial intelligence. Miguel Bessa, assistant professor of materi- als science, says he found the inspira- tion for this project during his time at the California Institute of Technology where he noticed a satellite structure that could open long solar sails from a very small package. He wondered if it would be possible to design a highly compressible but strong material that could be squeezed into a small fraction of its volume. “If this was possible, everyday ob- jects such as bicycles, dinner tables, and umbrellas could be folded into your pocket,” says Bessa. This can be achieved by structure-dominated ma- terials (metamaterials) that exploit new geometries to achieve new properties and functions. “However, metamate- rial design has relied on extensive ex- perimentation and a trial-and-error ap- proach,” Bessa explains. “We argue in favor of inverting the process by using machine learning for exploring new de- sign possibilities, while reducing exper- imentation to an absolute minimum.” Guided by machine learning, the team fabricated two designs at differ- ent length scales that transform brit- tle polymers into lightweight, recover- able, and super-compressible metama- terials. The macroscale design is tuned for maximum compressibility, while the microscale is designed for high strength and stiffness. www.tudelft.nl . MACHINE LEARNING HOLDS POTENTIAL New research from Los Alamos National Laboratory, the University of North Carolina at Chapel Hill, and the University of Florida shows that arti- ficial neural nets can be trained to en- code quantum mechanical laws to de- scribe themotions of molecules, greatly enhancing simulations. The new tech- nique, called the ANI-1ccx potential, will improve the accuracy of machine learn- ing-based potentials in future studies of metal alloys and detonation physics. Quantum mechanical (QM) algo- rithms used on classical computers can accurately describe the mechanical mo- tions of a compound in its operation- al environment. But QM scales poor- ly with different molecular sizes, limit- ing the scope of possible simulations and leaving practitioners to resort to using empirical information. Tradition- ally, empirical potentials have had to strike a tradeoff between accuracy and transferability: When the many parame- ters of the potential are finely tuned for one compound, accuracy decreases on other compounds. Now, the research- ers have developed a machine learning approach called transfer learning that lets them build empirical potentials by learning from data collected on millions Metamaterial created with artificial intel- ligence that transforms a brittle material into a sponge-like material. Courtesy of Delft University of Technology. New deep learning models predict inter- actions between atoms in organic mole- cules. Courtesy of Los Alamos National Lab. Welcome to Machine Learning/AI, a new department focusing on the explosive growth of these disciplines within the materials sci- ence and engineering community. Turn here to learn about the latest research on artificial intelligence and how it impacts computer- aided materials development. BRIEF The U.S. Department of Energy announced $27.6 million in funding over the next three years for specialized research in data science to accelerate discovery in both materials science and chemistry. The 19 awards aim to advance the application of modern data science methods such as artificial intelligence and machine learning to develop new ma- terials and chemical processes. www.science.osti.gov/bes.