Classificação preditiva de fases para ligas multicomponentes CrCoFeMnNi utilizando machine learning

Abstract

When we refer to multicomponent alloys, or high entropy alloys as they are commonly called, it is inevitable to discuss the challenge of exploring new compositions that may be of scientific interest. This difficulty is related to a range of possible compositions, given the amount of different elements and their percentages that can be combined to originate new materials. Since its discovery, alternative methods have been studied to try to predict some characteristics of these alloys without the need to experimentally produce them, or to use empirical methods, this, in addition to saving time, can optimize resources and makes it economically more feasible to investigate a greater number of combinations, until a material is reached that justifies its manufacture for further analysis. Among the existing ways to carry out this exploration, we can highlight methods based on the functional theory of density, or even thermodynamic simulations, which can use different methods, such as CALPHAD, which uses the Gibbs energy functions. However, these are methods that still have relatively long development cycles. Based on this, this work is intended to use science guided by big data, more specifically machine learning, where from a database, originated through thermodynamic simulation, using CALPHAD as a method, three different algorithms to predictively classify phases in multicomponent alloys, formed by the elements nickel (Ni), manganese (Mn), iron (Fe), chromium (Cr) and cobalt (Co). This classification consists of predicting whether a given composition at a constant temperature of 1000ºC presents a face-centered cubic, body-centered cubic and sigma structure as a phase, or whether none of these are present and are grouped under “others”. With that, through a database with 1000 different compositions, it was possible to carry out supervised training of three different types of algorithms, where, after properly trained and optimized, they performed the classification of about 494 new combinations that did not exist in the initial base. As a result, an accuracy of 91.72% was reached for the decision tree algorithm, 94.95% for the k-nearest neighbors and 96.36% for the support vector machine.