Water plays a pivotal role in life and industry, yet its complex hydrogen bonding networks and reaction mechanisms under extreme conditions remain incompletely understood. This research aims to overcome the limitations of existing computational method...
Water plays a pivotal role in life and industry, yet its complex hydrogen bonding networks and reaction mechanisms under extreme conditions remain incompletely understood. This research aims to overcome the limitations of existing computational methodologies by employing neural network potentials to elucidate dielectric relaxation phenomena in liquid water and reaction networks in supercritical water oxidation at the molecular level.
In the dielectric relaxation study of liquid water, neural network potentials were utilized to quantitatively reproduce the wide dielectric spectra from 0.1 GHz to 40 THz for the first time. Unlike previous studies that were limited to calculating the static dielectric constant (ε_0), this research captured both primary and secondary Debye relaxations and damped harmonic oscillator components, reproducing even hydrogen bond stretching resonances. A newly developed Markovian approach was established to describe hydrogen bond coordination state transitions through transition probability matrices, directly connecting activation barrier differences between models. The neural network model exhibited lower transition barriers and faster rearrangement dynamics compared to conventional pairwise potentials, which was identified as originating from accurate representation of many-body interactions.
In the study of acetic acid reaction networks in supercritical water oxidation, the superiority of neural network potentials was demonstrated through systematic comparison with reactive force fields. Molecular lifetime-based kinetics was introduced to overcome the pathway overlap limitations of conventional concentration-based models, precisely determining activation barriers for individual reaction steps. The neural network potential reproduced mechanisms consistent with experimental trends, including initial radical pathways, complete oxidation behavior, methane generation, and hydrocarboxyl-carbonic acid intermediate pathways, while the reactive force field showed systematic bias in overestimating thermal dissociation pathways due to overstabilization of acetyl and hydroxyl radicals. The superior agreement of neural network potentials with experimental networks was quantitatively validated through graph theory-based similarity analysis (Jaccard similarity 0.59, Wasserstein similarity 0.35).
This research methodologically presented innovative analytical frameworks including complete dielectric spectrum analysis using neural network potentials, Markovian hydrogen bond dynamics analysis, molecular lifetime-based reaction kinetics, and graph theory-based network validation. Scientifically, it established quantitative foundations that can contribute to energy storage, sensor applications, and environmentally friendly process optimization by elucidating water's dielectric relaxation mechanisms and molecular-level reaction pathways in supercritical water oxidation. These achievements present a new paradigm for understanding complex chemical phenomena under extreme conditions and provide an important foundation for advancing sustainable chemical technologies.