Machine learning modeling for the prediction of plastic properties in metallic glasses

Abstract

Metallic glasses are one of the most interesting mechanical materials studied in the last years, but as amorphous solids, they differ strongly from their crystalline counterparts. This matter can be addressed with the development and application of predictive techniques capable to describe the plastic regime. Here, machine learning models were employed for the prediction of plastic properties in CuZr metallic glasses. To this aim, 100 different samples were subjected to tensile tests by means of molecular dynamics simulations. A total of 17 materials properties were calculated and explored using statistical analysis. Strong correlations were found for stoichiometry, temperature, structural, and elastic properties with plastic properties. Three regression models were employed for the prediction of six plastic properties. Linear and Ridge regressions delivered the better prediction capability, with coefficients of determination above similar to 80% for three plastic properties, whereas Lasso regression rendered lower performance, with coefficients of determination above similar to 60% for two plastic properties. Overall, our work shows that molecular dynamics simulations together with machine learning models can provide a framework for the prediction of plastic behavior of complex materials.

Metallic glasses are one of the most interesting mechanical materials studied in the last years, but as amorphous solids, they differ strongly from their crystalline counterparts. This matter can be addressed with the development and application of predictive techniques capable to describe the plastic regime. Here, machine learning models were employed for the prediction of plastic properties in CuZr metallic glasses. To this aim, 100 different samples were subjected to tensile tests by means of molecular dynamics simulations. A total of 17 materials properties were calculated and explored using statistical analysis. Strong correlations were found for stoichiometry, temperature, structural, and elastic properties with plastic properties. Three regression models were employed for the prediction of six plastic properties. Linear and Ridge regressions delivered the better prediction capability, with coefficients of determination above similar to 80% for three plastic properties, whereas Lasso regression rendered lower performance, with coefficients of determination above similar to 60% for two plastic properties. Overall, our work shows that molecular dynamics simulations together with machine learning models can provide a framework for the prediction of plastic behavior of complex materials.