Site-Selective Initial Hydration of Boron at Multicomponent Borosilicate Glass–Water Interfaces Revealed Using Machine Learning Molecular Dynamics

# DeePMD Potential for Multicomponent Oxide Glass / Water SystemDate: 2026-04-29Author: Ryuki Kayano A machine learning potential (DeePMD-kit) for simulating oxide glass surfaces in contact with water.Elements covered: Al, B, Ca, H, Li, Mg, Na, O, Si, Zr. --- ## Directory Structure ### 01-modelTrained DeePMD-kit model files. | File | Description ||------|-------------|| `graph.pb` | Full TensorFlow frozen graph of the trained potential || `compressed.pb` | Compressed version for faster inference. This potential file was employed for all of simulations in this study. || `out.json` | Full description for training model (descriptor: `se_atten_v2`, r_cut = 8.0 Å, fitting: energy network) | ### 02-datasetDFT reference data (VASP OUTCAR / DeePMD `npy` format) used to train and test the developed MLP.`INCAR.aimd` and `INCAR.scf` are the VASP input templates for AIMD and single-point SCF calculations, respectively. | Subdirectory | Content ||---|---|| `bulk-glass/` | Bulk glass configurations (Albite, CJ, ISG, NBS, NSAC, SBNA, S655, VP, …) at 300 K, 1400 K, and 3000 K. Each composition has one OUTCAR per temperature. Compiled into `dataset.train/` and `dataset.test/`. || `bulk-water/` | Bulk liquid water at densities 0.8–1.2 g cm^{-3} and temperatures 200–600 K, generated by classical MD and labeled by DFT. || `bulk-water-aimd/` | AIMD bulk water trajectories at densities 0.8–1.2 g cm^{-3} and temperatures 280–400 K. || `gel/` | Hydrated glass-gel configurations for a wide range of glass compositions used to capture glass–water interface chemistry. || `surface-glass-cmd/` | Glass slab configurations produced by classical MD annealing (400 K, 1400 K, 2400 K) for ~40 compositions.|| `surface-glass-mlmd/` | Dry glass slab configurations generated by the preliminary MLP for ~20 compositions. | ### 03-lammps-isgLAMMPS trajectory files for ISG-1 (International Simple Glass) water–glass interface simulations at 330 K under NPT ensemble for 4 ns.Six independent runs (`ISG-1_01` – `ISG-1_06`) starting from different initial configurations, each containing the last 4000 frames at 1 ps intervals.Note that the analyzed trajectories in this study were dumped with an interval of 5 fs, but it's too huge to upload.