September_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 | S E P T E M B E R 2 0 2 0 7 MACHINE LEARNING | AI Researchers at the University of Jyväskylä, Finland, developed distance-based machine learning methods capable of reliably predicting the structures and atom- ic dynamics of nanoparticles. The new techniques should facilitate more efficient explorations of particle-particle reactions. www.jyu.fi/en. BRIEF PROBING POROUS MATERIALS WITH MACHINE LEARNING A group of international research- ers led by a team at the University of Cambridge, U.K., recently used ma- chine learning techniques to accurate- ly predict the mechanical properties of metal organic frameworks (MOFs), which could be useful in applications from storing dangerous gases to power- ing hydrogen-based cars. The scientists used their machine learning algorithm to predict the properties of thousands of existing MOFs, as well as those yet to be synthesized in the laboratory. Their results could significantly speed up the way materials are characterized and de- signed at the molecular scale. The crystalline structure of MOFs means that they can be made like build- ing blocks, where individual atoms or molecules can be switched in or out of the structure, a level of precision not possible with materials such as plas- tics. The structures are highly porous with massive surface area: An MOF the size of a sugar cube laid flat would cov- er an area the size of six football fields. In addition, they make highly effective storage devices. The pores in any given MOF can be customized to form a per- fectly-shaped storage pocket for dif- ferent molecules, just by changing the building blocks. MOFs are synthesized in powder form, then put under pres- sure and formed into larger pellets. Due to their porosity, many MOFs are de- stroyed in the process. To address this problem, an al- gorithm was developed to predict the mechanical properties of MOFs, so that only those with the necessary mechan- ical stability are manufactured. The re- searchers used a multi-level compu- tational approach to build an interac- tive map of the structural and mechan- ical landscape of MOFs. First, they used high-throughput molecular simulations for 3385 MOFs. Then they developed a free machine learning algorithm to automatically predict the mechanical properties of existing and future MOFs, along with an interactive website where other scientists can design and predict the performance of their own versions. www.cam.ac.uk . SCIENTIFIC MACHINE LEARNING TAKES OFF To address challenges involved in jet engine design, researchers at The University of Texas at Austin (UT) are developing scientific machine learn- ing , a new field that blends scientif- ic computing with machine learning. By combining physics modeling and data-driven learning, it is possible to create reduced-order models—simula- tions that can run in a fraction of the time required by traditional methods. The new technique has been ap- plied to a combustion code used by the Air Force known as General Equation and Mesh Solver (GEMS). The UT group received “snapshots” generated by run- ning the GEMS code for a particular sce- nario that modeled a single injector of a rocket engine combustor. These snap- shots represent pressure, velocity, tem- perature, and chemical content in the combustor and serve as the training data from which the researchers derive the reduced-order models. Generating that training data in GEMS takes about 200 hours of computer processing time. Once trained, the reduced-order mod- els can run the same simulation in sec- onds, as well as simulate into the fu- ture. Although not perfect, the models are particularly effective at capturing the phase and amplitude of the pres- sure signals, key elements for making accurate engine stability predictions. utexas.edu . Crystalline MOF. Courtesy of David Fairen- Jimenez. Researchers are developing a faster modeling technique to test jet engine performance under various conditions. Thiol-covered gold nanoparticles. Courtesy of Antti Pihlajamäki.