Towards nano-mechanical simulations of ceramics containing realistic defects via machine-learning potentials: the example of TiB2.

Transition metal diboride (TMB2) ceramics combine high hardness with excellent thermal and chemical stability, making them attractive for protective coating applications. TMB2s commonly grow as largely off-stoichiometric-containing vacancies or other simple crystallographic defects-and a particularly essential question is how such defects alter the response to mechanical loads at the nanoscale. We exploit molecular dynamics (MD) equipped with here-trained machine-learning interatomic potential (MLIP) to reveal effects of point and planar defects during nano-mechanical tests of TiB2, being a representative of the most common AlB2-type TMB2s. MLIP training consists of active learning on configurations from finite temperature ab initio molecular dynamics, including equilibrium and uniaxially loaded structures as well as various defective and/or extremely strained environments. Transferability to near-indenter-tip environments is achieved via on-the-fly training on extrapolative clusters from simple nanoindentation runs. Following MLIP's validation, we simulate room-temperature nanoindentation of TiB2-x structures, where B sub-stoichiommetry is realized by disordered B vacancies, single- and double-planar defects previously revealed by electron microscopy. A somewhat non-intuitive prediction is that TiB2-x structures can exhibit hardness comparable to the ideal TiB2, challenging traditional assumptions about weakening effects of sub-stoichiometry. This behavior is ascribed to Ti-rich planar defects, contrarily to vacancies which, at the same chemistry, notably deteriorate mechanical properties. We hypothesize that similar effects may be expected for other group 5-6 transition metal diborides.