While several studies confirmed that machine-learned potentials can provide accurate free energies for determining phase stabilities, the abilities of MLPs for efficiently constructing a full phase diagram of multicomponent systems are yet to be established. In this work, by employing neural network interatomic potentials, we demonstrate construction of the MgO-CaO eutectic phase diagram including liquid phases based on free energy calculation methods. In particular, the phase diagram predicted by the NNP trained on SCAN functional closely follows the experimental data. We believe that this work paves the way to fully ab initio calculation of phase diagrams. This work was published at Physical Review Materials 2022, 6, 113802

The atomic layer deposition (ALD) process of titanium nitride (TiN) thin films is widely used in microelectronics, but the detailed growth mechanism is still elusive at the atomistic level. In this work, we carry out kinetic Monte Carlo (kMC) simulations on the ALD process using TiCl4 and NH3 precursors. Based on the on-lattice model, we sort out key reactions relevant for the ALD process. Considering the local environments, the reaction energies are calculated at the level of density functional theory (DFT) while the activation barriers are linearly fitted to sampled cases. The resulting kMC model produces the temperature-dependent growth rates and the amounts of Cl residues, which is in reasonable agreement with experiments. Our growth pathway underscores the critical role of surface Cl atoms in the ALD process by generating HCl gas molecules. By revealing the atomistic mechanisms in the TiN-ALD process, this work would help optimize material properties of TiN thin films. This work was published at Computational Materials Science, 213, 111620

With the development of the density functional theory (DFT) and ever-increasing computational capacity, an accurate prediction of lattice thermal conductivity (LTC) based on the Boltzmann transport theory becomes computationally feasible. However, steep computational costs in evaluating interatomic force constants limit the theoretical investigation of LTC to relatively simple crystals. Currently, machine learning potentials (MLPs) are garnering attention as an efficient surrogate model of DFT. However, the applicability of MLPs to a wide range of materials has yet to be demonstrated. Furthermore, establishing a standard training set that provides consistent accuracy and computational efficiencies across a variety of materials would be useful. Herein, we test a various methods of training set construction, and compute LTC of materials with diverse symmetries and a wide range of LTC using MLP. The current work will establish a robust framework for accurately computing LTC with MLPs. This work was published at Computational Materials Science 2022, 211, 111472

The machine learning potentials (MLPs) are garnering large attention, providing the accuracy of ab initio calculations at much lower costs. MLPs guarantee their reliability within the training domain. However, it requires expertise and enormous time to collect all necessary configurations by hand-picking. We suggest a sampling method via metadynamics accumulating on the local atomic environment space, which is called G-metaD. G-metaD is applied to H:Pt(111), GeTe, and Si systems. This work was published at npj Computational Materials 2021, 7, 131