Machine-learning (ML)-based interatomic potentials can enable simulations of extended systems with an accuracy that is largely comparable to density functional theory (DFT), but with a computational cost that is orders of magnitude lower. In the present contribution we will demonstrate two examples of their application to describing mechanical properties of amorphous materials.

Amorphous silicon nitride (a-SiNx) is a widely studied noncrystalline material, and yet the subtle details of its atomistic structure and mechanical properties are still unclear. Due to small sizes of representative models, DFT cannot reliably predict its structural properties and hence the results lead to strongly anisotropic mechanical properties. Similar issues were present when disentangling mechanical response and local chemical compositions in ternary W-B-C amorphous materials (Fig. 1, left). In both cases, we solved the issues by fitting momentum tensor potentials (MTP) specific for a given material system. Subsequently, we employed simulated annealing procedure to generate an amorphous structural model containing more than 6000 atoms which yielded isotropic mechanical response (Fig. 1, right). The thus obtained values of Young’s modulus were successfully validated against experimental data. The structural models were further validated by comparing mass density as experimentally measured by X-ray reflectivity and by radial distribution function measured by synchrotron X-ray diffraction.