34 Microsc. Microanal. 27 (Suppl 1), 2021 doi:10.1017/S1431927621000714 © Microscopy Society of America 2021 Fast Automated Phase Differentiation in Industrial Stainless Steel by Combining Low-Loss EELS Experiments with Machine Learning-based Algorithms Luc Lajaunie 1 , Beatriz Amaya Dolores 2 , Ashwin Ramasubramaniam 3 , Lorena González Souto 2 , Rafael Sanchez 4 , Juan Almagro 4 , Javier Botana 2 and José Calvino 5 1 Departamento de Ciencia de los Materiales e Ingeniería Metalúrgica y Química Inorgánica, Facultad de Ciencias, Universidad de Cádiz, Cádiz, Spain, 2 Departamento de Ciencia de los Materiales e Ingeniería Metalúrgica y Química Inorgánica, Facultad de Ciencias, Universidad de Cádiz, Spain, 3 Department of Mechanical and Industrial Engineering, University of Massachusetts Amherst, Massachusetts, United States, 4 ACERINOX EUROPA SAU, Technical Department, Spain, 5 Departamento de Ciencia de los Materiales e Ingeniería Metalúrgica y Química Inorgánica, Facultad de Ciencias, Universidad de Cádiz, United States Introduction. Duplex stainless steels (DSSs) constitute a family of steels made of chromium-nickel- molybdenum-iron bi-phased alloys in which α ferrite and γ austenite fractions are present in relatively large separate volumes. Due to their bi-phased microstructure, they possess higher mechanical strength and better corrosion resistance than standard austenitic stainless steels and are used for a wide range of applications including thermal desalination plants, pipes and storage tanks for the oil & gas industry [1]. However, during aging, a large variety of phases, including the Cr-Mo-rich σ phase which is often observed at the α/γ interface boundary, are known to precipitate in DSS and lead to a dramatic deterioration of their mechanical and corrosion properties [2]. Characterizing and mapping in a timely manner the phases present in aged DSS for industrial applications if thus of high-importance. Because of the high intensity of the signal, low-loss EELS allows us to obtain large dataset with short acquisition time. However, interpretation and analysis of such data is not straightforward. Low-loss EELS spectra contain many excitation processes including volume plasmon which can be used for fingerprinting approaches but requires to use a catalogue of reference spectra and laborious fitting procedures [3]. In the present work, we developed a new and fast method based on low-loss EELS experiments to automatically separate the phases present in as-cast and aged industrial DSS. It allows us, not only to map α and γ phases, but also intermetallic phases such as the σ phase. Experimental. As-cast (2205, 2304 and 2001) and aged DSS (after thermomechanical treatment at 1090 or 1270ºC) were fabricated at the ACERINOX EUROPA SAU plant of Campo de Gibraltar. TEM samples were prepared by electropolishing and were studied by using a FEI Titan Cubed Themis 60-300 microscope at the University of Cádiz which was operated at 200 kV. The Themis is equipped with a double Cs aberration- corrector, a monochromator, an X-FEG gun, an Ultra High Resolution Energy Filter (Gatan Quantum ERS) which allows working in dual-EELS mode and a Super X EDS detector. Phase determination was based on the prior results of EDS quantification. Absorption correction for EDS quantification was performed by taking into account the thickness of the probed area. For this purpose, low-loss EELS measurements were used to determine the t/λ ratio (t the thickness of the analyzed crystal and λ the inelastic mean free path) and the modified Iakoubovskii formula [4] was used to determine λ. EELS data were acquired with a dispersion of 0.1eV/pixel and an energy resolution of 0.8 eV. Dataset of about 18 μm × 18 μm (100 pixels × 100 pixels) were acquired with a dwell time of 0.05s. After multiple scattering removal, EELS datasets were clustered by using the “kmeans++” algorithm. The only parameter needed was the number of clusters which was determined by using the silhouette metric [5] before clustering. To compare with more classical approaches, EELS dataset were also fitted pixel per pixel by using the Drude model in order to map the plasmon energy and plasmon width parameters. In order to interpret the results, EELS experimental spectra were also compared with calculated spectra based on the density-functional theory. https://www.cambridge.org/core/terms. https://doi.org/10.1017/S1431927621000714 Downloaded from https://www.cambridge.org/core. IP address: 54.196.76.231, on 05 Nov 2021 at 00:50:58, subject to the Cambridge Core terms of use, available at