arXiv subjects

Meguru Yamazaki

Publications and source records attributed to Meguru Yamazaki.

3 recordsLinked to original sources

arXiv2025

Knowledge Distillation Framework for Accelerating High-Accuracy Neural Network-Based Molecular Dynamics Simulations

Neural network potentials (NNPs) offer a powerful alternative to traditional force fields for molecular dynamics (MD) simulations. Accurate and stable MD simulations, crucial for evaluating material properties, require training data encompassing both low-energy stable structures and high-energy structures. Conventional knowledge distillation (KD) methods fine-tune a pre-trained NNP as a teacher model to generate training data for a student model. However, in material-specific models, this fine-tuning process increases energy barriers, making it difficult to create training data containing high-energy structures. To address this, we propose a novel KD framework that leverages a non-fine-tuned, off-the-shelf pre-trained NNP as a teacher. Its gentler energy landscape facilitates the exploration of a wider range of structures, including the high-energy structures crucial for stable MD simulations. Our framework employs a two-stage training process: first, the student NNP is trained with a dataset generated by the off-the-shelf teacher; then, it is fine-tuned with a smaller, high-accuracy density functional theory (DFT) dataset. We demonstrate the effectiveness of our framework by applying it to both organic (polyethylene glycol) and inorganic (L$_{10}$GeP$_{2}$S$_{12}$) materials, achieving comparable or superior accuracy in reproducing physical properties compared to existing methods. Importantly, our method reduces the number of expensive DFT calculations by 10x compared to existing NNP generation methods, without sacrificing accuracy. Furthermore, the resulting student NNP achieves up to 106x speedup in inference compared to the teacher NNP, enabling significantly faster and more efficient MD simulations.

cs.LG↗

arXiv2025

Molecular Dynamics Simulations of SrTiO$_3$ with Oxygen Vacancies using Neural Network Potentials

A precise analysis of point defects in solids requires accurate molecular dynamics (MD) simulations of large-scale systems. However, ab initio MD simulations based on density functional theory (DFT) incur high computational cost, while classical MD simulations lack accuracy. We perform MD simulations using a neural network potential (NNP) model (NNP-MD) to predict the physical quantities of both pristine SrTiO$_3$ and SrTiO$_3$ in the presence of oxygen vacancies (V$_{\text{O}}$). To verify the accuracy of the NNP models trained on different data sets, their NNP-MD predictions are compared with the results obtained from DFT calculations. The predictions of the total energy show good agreement with the DFT results for all these NNP models, and the NNP models can also predict the formation energy once SrTiO$_3$:V$_{\text{O}}$ data are included in the training data sets. Even for larger supercell sizes that are difficult to calculate using first-principles calculations, the formation energies evaluated from the NNP-MD simulations well reproduce the extrapolated DFT values. This study offer important knowledge for constructing accurate NNP models to describe point-defect systems including SrTiO$_3$:V$_{\text{O}}$.

cond-mat.mtrl-sci↗

arXiv2025

Active Learning of a Neural Network Potential for Large-Scale Atomistic Simulations of Polymer Electrolyte Membranes

Machine learning interatomic potentials (MLIPs) can achieve near density-functional-theory (DFT) accuracy at force-field computational cost; however, long-time, large-scale molecular dynamics (MD) simulations often fail when trajectories sample local atomic environments that are underrepresented in the training set. Here, we employ an active-learning workflow to develop a robust neural network potential (NNP) for perfluorinated ionomer membranes (Nafion) across a wide range of hydration levels ($\lambda$). A reliable deep potential (DP) model is constructed through iterative dataset expansion within active-learning loops. Specifically, off-equilibrium configurations are generated via non-equilibrium DPMD simulations and selected using an ensemble force-deviation criterion combined with a three-dimensional structural feature space augmented by minimum interatomic distances, which significantly enhances the DP model's robustness. The trained DP model enables stable MD simulations of large Nafion systems containing approximately 10,000-20,000 atoms for an extended duration of 31 ns. Our DPMD simulations reproduce the qualitative hydration dependence of density and yield self-diffusion coefficients of hydrogen atoms and hydronium ions in quantitative agreement with experimental values. Compared with previous ab initio MD and MLIP-MD studies, our simulations show improved agreement with experimental transport properties and remain predictive up to $\lambda = 24$, well beyond the training range ($\lambda \leq 13$). This work provides an efficient and scalable approach for achieving stable, large-scale NNP-MD simulations of heterogeneous polymer electrolyte membranes and related disordered materials.

cond-mat.mtrl-sci↗