Ab InitioQuality NMR Parameters in Solid-State Materials Using a High-Dimensional Neural-Network Representation

Abstract

Nuclear magnetic resonance (NMR) spectroscopy is one of the most powerful experimental tools to probe the local atomic order of a wide range of solid-state compounds. However, due to the complexity of the related spectra, in particular for amorphous materials, their interpretation in terms of structural information is often challenging. These difficulties can be overcome by combining molecular dynamics simulations to generate realistic structural models with an ab initio evaluation of the corresponding chemical shift and quadrupolar coupling tensors. However, due to computational constraints, this approach is limited to relatively small system sizes which, for amorphous materials, prevents an adequate statistical sampling of the distribution of the local environments that is required to quantitatively describe the system. In this work, we present an approach to efficiently and accurately predict the NMR parameters of very large systems. This is achieved by using a high-dimensional neural-network representation of NMR parameters that are calculated using an ab initio formalism. To illustrate the potential of this approach, we applied this neural-network NMR (NN-NMR) method on the (17)O and (29)Si quadrupolar coupling and chemical shift parameters of various crystalline silica polymorphs and silica glasses. This approach is, in principal, general and has the potential to be applied to predict the NMR properties of various materials.

Perspective: Current advances in solid-state NMR spectroscopy

“…Obtaining suitable structural models is more of an open research question for glassy-type disorder and polymeric materials. Approaches that have been used successfully in NMR studies include quench-cooling of MD simulations of molten solids, 65 as well as approaches that avoid the need for large-scale DFT calculations, such as training neural networks to predict chemical shifts in glasses 66 or deriving empirical correlations from related compounds. 67,68 In contrast to "static disorder," dynamics tend to average NMR parameters, and so, depending on the time scale of the dynamics, will narrow rather than broaden NMR line shapes.…”

Hochdimensionale neuronale Netze für Potentialhyperflächen großer molekularer und kondensierter Systeme

“…Über diese Anwendungen hinaus wurden viele neue Ideen in Bezug auf HDNNPs vorgeschlagen, wie die Verbesserung von semiempirischen Methoden zur Beschreibung des QM-Teils in QM/MM-Simulationen, [109] die Beschleunigung von Sattelpunktsuchen, [127] eine hierarchische Konstruktion von NNs fürM ehrkomponentensysteme, [133] die Vorhersage von NMR-Parametern von Festkçrpermaterialien, [137] die Vorhersage von thermodynamischen Grçßen [138] und ein neuer Ansatz zur Kombination mit Vielkçrper-Entwicklungen. Iterative Verbesserung des Trainingsdatensatzes mit zwei NNPs.…”

“…In (c) stimmen beide NNPs sehr genau füralle Konfigurationen überein, was zeigt, dass alle relevanten Konfigurationen im Trainingssatz gut vertreten sind. Über diese Anwendungen hinaus wurden viele neue Ideen in Bezug auf HDNNPs vorgeschlagen, wie die Verbesserung von semiempirischen Methoden zur Beschreibung des QM-Teils in QM/MM-Simulationen, [109] die Beschleunigung von Sattelpunktsuchen, [127] eine hierarchische Konstruktion von NNs fürM ehrkomponentensysteme, [133] die Vorhersage von NMR-Parametern von Festkçrpermaterialien, [137] die Vorhersage von thermodynamischen Grçßen [138] und ein neuer Ansatz zur Kombination mit Vielkçrper-Entwicklungen. [139] Viele dieser Entwicklungen wurden erst kürzlich publiziert, was zeigt, dass sich dieser Forschungsbereich rasant entwickelt und viele interessante methodische Verbesserungen, die das Anwendungsspektrum von NNPs und MLPs erweitern werden, in naher Zukunft erwartet werden kçnnen.…”

Quantitative estimation of properties from core-loss spectrum via neural network

“…In recent years, machine learning approaches have widely spread in materials science to predict material properties quantitively and overcome various obstacles with extensive computations [20][21][22][23]. They have also been applied to spectroscopic data, such as infrared spectroscopy [24], nuclear magnetic resonance (NMR) [25], ELNES/XANES [26,27], and extended x-ray absorption fine structure (EXAFS) [28] to extract hidden information.…”