November_December_2021_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 | N O V E M B E R / D E C E M B E R 2 0 2 1 8 MACHINE LEARNING | AI USING NEURAL NETWORKS TO PREDICT STRAIN Scientists at the University of Illi- nois Urbana-Champaign recently used data from high-resolution images of stainless steel samples to train neural networks that make predictions about how the material will behave under strain at its grain boundaries. “Rather than using an extremely detailed and cumbersome physics model with a lot of fitting parameters, we used machine learning to train a neural network to make these predictions,” explains re- searcher John Lambros. This is the first time the technique was applied to learn what happens in a metal microstruc- ture under various loading conditions. In this experiment, the goal was to see how much strain accumulated at grain boundaries in a polycrystalline metal during creep. “We believed the physical differ- ences between the two grains adjacent to the boundary would be more import- ant, or at least an equally important pa- rameter. So, the most remarkable find- ing for me was that one single geomet- ric parameter was able to predict the results 80% percent of the time,” says Lambros. “It’s the geometry—the angle at which you’re loading it that made the most difference. It was surprising, be- cause it means that all this sophisticat- ed, multiscale modeling that we do to understand all the physics may be only about 20% percent important.” Lam- bros notes that the current model only works near grain boundaries, and that a different set of inputs will be needed to work in the interior. illinois.edu . MACHINE LEARNING SUPPORTS HYDROGEN STORAGE A team of materials scientists and computer scientists from Sandia Na- tional Laboratories, Angstrom Labora- tory in Sweden, and Nottingham Uni- versity in the U.K. spent more than a year creating 12 new alloys that demon- strate how machine learning can help accelerate the future of hydrogen ener- gy by making it easier to create hydro- gen infrastructure for consumers. Such machine learning models only take sec- onds to execute and can rapidly screen new chemical spaces: In this case, 600 materials show promise for hydrogen storage and transmission. The team demonstrated that machine learning techniques could indeed model the physics and chemistry of complex phe- nomena that occur when hydrogen in- teracts with metals. The researchers also found something else, results that have dramatic implications for small- scale hydrogen generation at hydrogen fuel-cell filling stations. “These high-entropy alloy hy- drides could enable a natural cascade compression of hydrogen as it moves through the different materials,” says Vitalie Stavila of Sandia, adding that compressing hydrogen is traditionally done through a mechanical process. He describes building a storage tank with multiple layers of these different alloys. As hydrogen is pumped into the tank, the first layer compresses the gas as it moves through the material. The sec- ond layer compresses it even further and so on through all of the layers of dif- fering alloys, naturally making the hy- drogen usable in motors that generate electricity. “As hydrogenmoves through those layers, it gets more and more pressurized with no mechanical effort,” says Stavila. “You could theoretically pump in 1 bar of hydrogen and get 800 bar out—the pressure needed for hydro- gen charging stations.” sandia.gov. Top row shows experimental measurements of three different strain components; bottom row shows corresponding network predictions of a neural network trained from a different set of experiments, illustrating that the model works. Researchers are using machine learning models to discover new high-entropy alloys with attractive hydrogen storage properties. Courtesy of MatthewWitman.