%#!latexmk main.tex
\section*{Acknowledgments}

This study was supported by JSPS KAKENHI (Grant Numbers 23H04096 and
24K01597). Part of the computational resources were provided by Fugaku
at the RIKEN Center for Computational Science (Project IDs: hp250103 and
hp260055) and by Genkai at the Research Institute for Information
Technology, Kyushu University.

% We acknowledge the ChemRxiv preprint server for providing a platform to
% share the preprint version of this work ({\ChemRxiv}).

\section*{Supporting Information}

The Supporting Information is available free of charge.

\begin{itemize}
 \item Model architecture, descriptor settings, and training
       hyperparameters used to construct the machine-learning potential
       (Table~{\SIdeepmd}).
 \item Reference crystalline structures and the linear regression
       procedure used to calibrate GIPAW-calculated {\nP} chemical shifts
       against experimental values (Fig.~{\SInmrref}).
 \item Radial distribution functions used to determine the cutoff
       distances for defining Zn--O coordination environments and
       coordination numbers (Fig.~{\SIgrZnO}).
\end{itemize}

\section*{Data Availability}

The training data generated using VASP, along with the machine learning
potential (MLP) constructed for use in MLMD simulations, and glass
structures obtained by MLMD are available from the Zenodo data
repository (DOI: xxxxxx/zenodo.xxxxxxxx).


\section*{Author Information}
\subsection*{ORCID}

\noindent Takahiro Ohkubo: 0000-0001-8187-1470