Oxygen grain-boundary segregation in HCP Ti — Computational investigations using an atomic cluster expansion potential

The segregation energy of oxygen interstitial solutes to grain boundaries (GBs) in hexagonal close-packed (HCP) titanium is investigated through atomistic simulations based on a machine-learning interatomic potential (MLIP). The Ti–O MLIP is developed for titanium with interstitial oxygen solutes up to a concentration of 20 at.%. It is based on the formalism of the atomic cluster expansion (ACE), trained on an extensive dataset of density functional theory calculations exploring over 200,000 atomic environments for Ti and O interstitials. The ACE MLIP is used to compute oxygen GB segregation energies in 4685 different symmetric tilt GBs with [ 0001 ] , [ 1 1 ̄ 00 ] and [ 1 2 ̄ 10 ] tilt axes. The segregation energies span a range of -0.2 eV to 2.0 eV, and over 90 % of the 4685 GBs display one or more sites with negative (attractive) segregation energies. The lowest-energy sites are found to be those with coordination numbers and Voronoi indices most similar to that of the equilibrium octahedral site in the bulk HCP Ti structure. We further explore the efficacy of crystal-graph convolutional neural network ML models for predicting segregation energies based solely on information about the local atomic environment. • The Ti–O MLIP is developed for titanium with interstitial oxygen solutes up to a concentration of 20 at.% based on atomic cluster expansion (ACE) format. • The MLIP is shown to reproduce with reasonable accuracy properties including elastic constants, twin boundary energies, and grain-boundary segregation energies of oxygen solutes. • Important features and CGCNN prediction models are investigated for oxygen GB segregation energies in 4685 different symmetric tilt GBs.