May/June_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 | M A Y / J U N E 2 0 2 0 7 MACHINE LEARNING | AI TAGGING MICROSTRUCTURES WITH MACHINE LEARNING Scientists at the National Insti- tute for Materials Science (NIMS), Ja- pan, and the University of Tokyo are us- ing machine learning to rapidly analyze large amounts of data in order to recog- nize and label microscopic structures inside welded steel. The team prepared steel alloys made of carbon, silicon, manganese, phosphorus, and sulfur by cooling them from 1400°C at differ- ent rates: 0.3°, 1°, 3°, or 10°C per second. The variable cooling rates led to forma- tion of different microstructures with- in the steel. Metallurgists then manual- ly identified three types of microstruc- tures within the alloys: ferrite/perlite, ferrite/perlite/bainite, and bainite/mar- tensite. Sub-phases of ferrite were also identified. The images were run through sev- eral machine learning models, using al- gorithms to train them to recognize and label the images. The team found a ma- chine learning classification method called “random forest” made the most accurate predic- tions of alloy mi- crostructure. This method could be applied to a wide range of metals in both research and industry set- tings, according to the team. “There is much hope this machine learning method will assist in automating microstructure analysis us- ing large datasets and in the develop- ment of newmaterials with desired me- chanical properties,” says Dmitry Bul- garevich of NIMS. www.nims.go.jp/eng. USING MACHINE LEARNING TO STUDY INTERFACES In a new study from the DOE’s Ar- gonne National Laboratory, Lemont, Ill., researchers used a new technique based on machine learning to discov- er the details of buried interfaces and edges in a material. By using machine learning as an image processing tech- nique, the scientists dramatically ac- celerated the laborious manual process of looking at interfaces without hav- ing to sacrifice accuracy. Atom probe tomography (APT) was used to gener- ate the data. Time-of-flight measure- ments and mass spectrometry were then performed to identify where a par- ticular atom originated in the material. The new process generates a very large dataset of positions of atoms in the sample. To analyze it, researchers seg- ment the data into 2D slices. Each slice is represented as an image, from which the machine learning algorithm can de- termine the edges and interfaces. In training the algorithm to recog- nize interfaces, the scientists took an unconventional approach: Rather than using images from a library of materi- als that might have had poorly defined boundaries, they began with pictures of cats and dogs to help the algorithm learn about edges. The team was then able to prove the algorithm’s accura- cy by compiling a set of molecular dy- namics simulations. They used these to make synthetic datasets in which the composition of the simulated sample was completely known. Then, by going back to the machine learning method, they were able to extract composition profiles and compare them to reality. The team believes they could dramati- cally speed the analysis of a wide range of material interfaces by pairing the ma- chine learning algorithmwith newly de- veloped quantitative analysis software. anl.gov. Three types of alloy microstructures were identified in the micro- scopic images. From left, ferrite/pearlite, ferrite/pearlite/bainite, and bainite/martensite. Courtesy of NIMS. 3D point cloud reconstruction of a cobalt superalloy atom probe tomography spec- imen (left) and resulting interface from the edge detection method (right). Courtesy of Argonne National Laboratory. The Camille and Henry Dreyfus Foundation announced a new program for machine learning in the chemical scienc- es and engineering. The program is open to U.S. academic institutions that grant a bachelor’s or higher degree in the chemical sciences, including materials chemistry and chemical engineering. dreyfus.org . BRIEF