Bayesian exploration of the composition space of CuZrAl metallic glasses for mechanical properties

Abstract

Designing metallic glasses in silico is a major challenge in materials science given their disordered atomic structure and the vast compositional space to explore. Here, we tackle this challenge by finding optimal compositions for target mechanical properties. We apply Bayesian exploration for the CuZrAl composition, a paradigmatic metallic glass known for its good glass forming ability. We exploit an automated loop with an online database, a Bayesian optimization algorithm, and molecular dynamics simulations. From the ubiquitous 50/50 CuZr starting point, we map the composition landscape, changing the ratio of elements and adding aluminum, to characterize the yield stress and the shear modulus. This approach demonstrates with relatively modest effort that the system has an optimal composition window for the yield stress around aluminum concentration c_Al = 15% and zirconium concentration c_Zr = 30%. We also explore several cooling rates (“process parameters”) and find that the best mechanical properties for a composition result from being most affected by the cooling procedure. Our Bayesian approach paves the novel way for the design of metallic glasses with “small data”, with an eye toward both future in silico design and experimental applications exploiting this toolbox.