Predictive quantum vibrational spectra through active learning 4G-NNPs

Abstract

Predictive simulation of vibrational spectra of complex condensed-phase and interface systems with thousands of degrees of freedom has long been a challenging task of modern condensed matter theory. In this work, fourth-generation high-dimensional committee neural network potentials (4G-HDCNNPs) are developed using active learning and query-by-committee, and introduced to the essential nuclear quantum effects (NQEs) as well as conformational entropy and anharmonicities from path integral (PI) molecular dynamics simulations. Using representative bulk water and air-water interface test cases, we demonstrate the accuracy of the developed framework in infrared spectral simulations. Specifically, by seamlessly integrating non-local charge transfer effects from 4G-HDCNNPs with the NQEs from PI methods, our introduced methodology yields accurate infrared spectra using predicted charges from the 4G-HDCNNP architecture without explicit training of dipole moments. The framework introduced in this work is simple and general, offering a practical paradigm for predictive spectral simulations of complex condensed phases and interfaces, free from empirical parameterizations and ad hoc fitting.

Figures (7)

• Figure 1: The adopted workflow for 4G-HDCNNPs training and vibrational spectral simulations using modified AMLschran2020committee with NQEs incorporated.
• Figure 2: Calculated force mean absolute errors (MAEs) of the 4G-HDCNNP models across all three generations. The MAEe are determined using independently generated validation sets calculated with revPBE0-D3, both for classical and quantum configurations averaged from bulk and interface performances.
• Figure 3: (a) Calculated correlations between the reference (DDEC6) and 4G-HDCNNP predicted dipole moments along $x$, $y$, and $z$ directions. (b) Calculated errors ($\Delta\boldsymbol{\mu} = \boldsymbol{\mu}_{\text{NN}} - \boldsymbol{\mu}_{\text{ref}}$) are also shown.
• Figure 4: Calculated RDFs for (a) $O_W$--$O_W$ (oxygen of water), (b) $O_W$--$H_W$ (hydrogen of water), and (c) $H_W$--$H_W$ pairs from classical 4G-HDCNNP simulations compared to the reference revPBE0-D3 AIMD data. Calculated errors are shown in the bottom panels.
• Figure 5: 4G-HDCNNP calculated MD, TRPMD, and PA-CMD simulated IR spectra of bulk water compared to the experiment.augBertie1996 The dashed line represents the OH-stretch maximum peak position from the experiment.