Since there are many unknown stable materials with ternary or higher (multinary) composition, crystal structure prediction is necessary to accelerate the rate of material discovery. This demands fast and accurate evaluation of free energies in exploring a vast number of atomic configurations. The neural network potential (NNP) can meet this requirement but a scarcity of information on the crystal structure poses a challenge in choosing training sets. In this work, we propose a method of constructing training sets from density functional theory (DFT)–based dynamical trajectories of disordered structures, which does not require any preceding information on material structures except for the chemical composition. With this method, we find strong correlation of NNP and DFT energies, ensuring that the NNPs can properly rank energies among low-energy crystalline structures. We also find that the evolutionary search using the NNPs is more efficient than the DFT-based approach. This work was published at Physical Review B 2020, 102, 224104.
Despite the great potential of a-CuI:Sn as a new class of transparent p-type semiconductors, the fundamental understanding is still incomplete. To reveal the microscopic origin, structural and electrical properties of a-CuI are investigated. Despite the amorphous structure, states at valence band maximum are extended in linear ways due to the hybridization between I-5p state, explaining the high hole mobilities in the experiment. This work will help design new wide-band-gap p-type semiconductors. This work was published at Physica Status Solidi B, accepted.
This work proposes a highly efficient and atomic-resolution prediction uncertainty indicator for neural network potential (NNP) using a “replica” ensemble. Compared to previous approaches using model ensembles, the present method excels in the training speed and spatial resolution. This method is applied to simulating silicidation processes of Ni contacts in semiconductor devices, which is a highly complicated process and requires close monitoring of prediction uncertainty to judge the soundness of simulations. By analyzing the problematic structures, NNP can be improved with a repairing data set, thereby obtaining a reliable simulation up to several nanoseconds. This work was published at The Journal of Physical Chemistry Letters 2020, 11, 6090.
Among various emerging memories, phase change memory has attracted wide interests to compete against conventional memories. In spite of recent commercialization of 3D XPoint by Intel and Micron, slow writing speed (crystallization) is still problematic. Molecular dynamics simulation helps to understand crystallization kinetics in atomic scale. First-principles molecular dynamics gives good insight, but it is difficult to quantitatively compare with experiments due to finite cell. Recently, nueral network potential (NNP) fitted to DFT data enables accurate large-scale crystallization simulation. In this study, we develop reliable NNP for GeTe and simulate crystallization process considering device condition. Nucleation-limited crystallizations are well reproduced,leading to incubation periods of 7 or 17 ns at 500 K. This is in good agreements with the minimum incubation time of 30 ns in experiments. This work was published at Computational Materials Science 2020, 181, 109725.
Ab initio calculations based on the density functional theory (DFT) become a vital tool in materials science for understanding and predicting material properties. However, it requires in-depth knowledge on underlying theories and enough experience to produce reliable data. Recently, several automation utilities have been developed to accelerate data production but they still assume that users are familiar with technical details. Here, we introduce a full-fledged automation code running a DFT program. The package requires only structure information from the user and provides a highly accurate band structure, band gap, effective mass, density of states and dielectric constant for the given structure. As a result, anyone can run DFT program without any background knowledge using the package. This work was published at Computer Physics Communications 2020, 256, 107450.