April_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 | A P R I L 2 0 2 0 8 MACHINE LEARNING | AI NEURAL NETWORKS ENHANCE JET ENGINE DESIGN A team of researchers at the DOE’s Argonne National Laboratory (ANL), Lemont, Ill., is combining unique x-ray experiments with novel computer sim- ulations to help aerospace companies save time and money on jet engine de- sign. Using the lab’s 7-BM x-ray beam- line at the Advanced Photon Source (APS), engineer Brandon Sforzo and his colleagues can explore the steel of an engine’s fuel injector. “Visualizing through steel with this detail is not possible with any oth- er diagnostic technique,” says engineer Prithwish Kundu. In 2019, the team in- vestigated the fluid dynamics within a gas turbine engine and found behav- ior that surprised Sforzo. These types of revelations help scientists under- stand the fundamental physics that af- fect engine performance. They also give scientists like Kundu, who feed this in- formation into the lab’s supercomput- ers, building blocks that enable high- fidelity simulations. With the right boundary conditions, scientists can build models that predict a host of en- gine behavior that may be unmeasur- able during experiments. While engineering thrives on high- fidelity 3D models, those models often run for months on su- percomputers, a scarce resource. To solve this challenge, Kundu and others are exploring deep neural networks, a type of artificial intel- ligence. They have al- ready developed neu- ral-network algorithms that significantly re- duce the time it takes to optimize models and help understand the inner workings of combustion engines. Using ANL’s computers, Kundu and his team created a high-fidelity model that measures how two different jet fuels behave in the combustor section of a gas turbine engine. The researchers are now working with NASA Langley to sim- ulate supersonic combustion. anl.gov. MACHINE LEARNING REVEALS CRYSTAL STRUCTURES Nanoengineers at the Universi- ty of California, San Diego (UCSD) de- veloped a computer-based method that could make it less labor-intensive to determine the crystal structures of various ma- terials. The tech- nique uses a ma- chine learning al- gorithm to inde- pendently analyze electron diffraction patterns. Profes- sor Kenneth Vec- chio and his team developed the new approach, which involves using a scanning electron microscope (SEM) to collect electron backscatter diffrac- tion (EBSD) patterns. SEM-based EBSD can be performed on large samples and analyzed at multiple length scales, providing local submicron information mapped to centimeter scales. However, the drawback of com- mercial EBSD systems is the software’s inability to determine the atomic struc- ture of the crystalline lattices present within the material being analyzed. This means a user must select up to five crystal structures presumed to be in the sample before the software attempts to find probable matches to the diffraction pattern. The complex nature of the pat- tern often causes the software to find false structure matches in the user-de- fined list, with accuracy dependent on operator experience and prior knowl- edge of the sample. In contrast, the method that Vecchio’s team developed does this autonomously. The deep neu- ral network independently analyzes each diffraction pattern to determine the crystal lattice out of all possible lat- tice structure types, with a degree of ac- curacy greater than 95%. ucsd.edu. Argonne researchers prepare an experiment to investigate fuel injector design at the lab’s Advanced Photon Source. Illustration of the inner workings of a neural network that computes the probability that the input diffraction pattern belongs to a given class.