An Accurate Tensorial Model for Prediction of Full Zeolite NMR Spectra

Carlos Bornes, Chiheb Ben Mahmoud, Volker L. Deringer, Christopher J. Heard, Lukáš Grajciar · 2026

Read An Accurate Tensorial Model for Prediction of Full Zeolite NMR Spectra by Carlos Bornes, Chiheb Ben Mahmoud, Volker L. Deringer, Christopher J. Heard, Lukáš Grajciar (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-22268.txt.

Introduction

Bornes and co-authors train a machine learning model that predicts full NMR spectra for zeolites. Solid-state NMR helps read complex crystal structures. But computing NMR tensors from first principles is costly. Their tensorial model predicts magnetic shielding and electric field gradient tensors at once. It trains on a diverse set of zeolite materials. The test nuclei include aluminum, silicon, oxygen, sodium, and hydrogen. The authors report high precision on all NMR observables. They then predict full aluminum and silicon spectra for the zeolite RTH.

Catalog record

CREATORCarlos Bornes, Chiheb Ben Mahmoud, Volker L. Deringer, Christopher J. Heard, Lukáš Grajciar
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47781
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.22268

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

An Accurate Tensorial Model for Prediction of Full Zeolite NMR Spectra Carlos Bornes, Chiheb Ben Mahmoud, Volker L. Deringer, Christopher J. Heard, Lukáš Grajciar Abstract Solid state nuclear magnetic resonance (ss-NMR) is one of the most sensitive and popular techniques for structure elucidation in geometrically complex crystalline materials, such as zeolites. Synergistic support from computational modelling is vital to interpret experimental spectra, and relate ss-NMR to atomistic models. Nevertheless, computational predictions are hindered by the high expense of calculating magnetic shielding (MS) and electric field gradient (EFG) tensors from first principles. In this work, we leverage a novel tensorial machine learning approach to train a general model for predicting complete NMR tensors. We demonstrate the utility of the approach for a diverse dataset of zeolitic materials and NMR-active nuclei ( 27 Al, 29 Si, 17 O, 23 Na and 1 H), predicting all NMR observables to a high degree of precision. These observables are then translated into predictions of the full 27 Al and 29 Si ss-nMR spectra for the exemplary zeolite RTH. Thus, this work opens a pathway to accurate, high-throughput NMR simulation for large-scale and realistic models of chemically complex zeolites. † † affiliation: Department of Physical and Macromolecular Chemistry, Charles University, Hlavova 8, Praha 2, Prague 12800, Czech Republic † † affiliation: Department of Materials, University of Oxford, Oxford OX1 3PH, United Kingdom † † affiliation: Inorganic Chemistry Laboratory, Department of Chemistry, University of Oxford, Oxford OX1 3QR, United Kingdom † † affiliation: Department of Physical and Macromolecular Chemistry, Charles University, Hlavova 8, Praha 2, Prague 12800, Czech Republic † † affiliation: Department of Physical and Macromolecular Chemistry, Charles University, Hlavova 8, Praha 2, Prague 12800, Czech Republic † † email: [email protected] 1 Introduction Zeolites are crystalline nanoporous aluminosilicates widely used in heterogeneous catalysis, gas separation, and water purification. 28 ; 52 Despite decades of academic research and industrial deployment, many fundamental long-standing questions remain, including the preference of specific aluminum distributions and speciation, the effects of cation solvation and dynamics, and the structure and formation mechanism of defects. One reason hindering the advance of a true atomistic understanding of zeolites arises from the disconnect between complex zeolite samples used experimentally and the simplified models typically employed in DFT-based studies. Among experimental techniques capable of bridging this gap, solid-state nuclear magnetic resonance (NMR) spectroscopy stands out as one of the most powerful techniques used for probing zeolite local structure and dynamics, 50 ; 40 with nuclei such as 1 H, 17 O, 23 Na, 27 Al, and 29 Si each providing a unique window into different aspects of zeolite chemistry. However, the interpretation of NMR spectra is often hampered by the complexity of real samples, where multiple overlapping environments, the effects of hydration and temperature, and disorder make unambiguous spectral assignment challenging. 37 First-principles calculations of NMR parameters via the gauge-including projector augmented wave (GIPAW) method 49 ; 64 have become essential tools for bridging this gap, 15 ; 3 but their high computational cost restricts applications to small unit cells, static or short ab initio molecular dynamics (MD) trajectories, and a limited subset of possible configurations. Machine learning interatomic potentials (MLIPs) have partially addressed the computational bottleneck by enabling large-scale atomistic simulations that retain the accuracy of ab initio methods at a fraction of the cost, enabling the exploration of large configurational spaces and long timescales relevant to experimental conditions. MLIPs now span a broad range of specificity, from reaction-specific potentials, 14 ; 1 through class-specific models targeting families of materials such as silicates, ranging from the Si–O system, 25 SiO 2 –water interfaces, 53 and hydrated aluminosilicate zeolites, 26 ; 13 ; 27 ; 17 to foundational models trained on diverse datasets, 8 ; 47 ; 44 each offering different trade-offs between accuracy and transferability. Despite these advances, MLIPs only predict energies and forces for driving MD simulations, while the prediction of experimentally observable properties such as NMR parameters often requires a separate ML model and remains comparatively unexplored, owing in part to the higher cost of generating reference datasets and the need for ML architectures capable of learning tensorial response properties that transform equivariantly under rotations. For molecular solids, the ShiftML framework 48 has demonstrated the power of SOAP-based kernels for predicting chemical shifts of 1 H, 13 C, 15 N, and 17 O, and in its most recent iterations the ShiftML was expanded to other nuclei, 21 and the prediction of chemical shift anisotropy. 35 However, ShiftML is trained on organic molecular solids and does not cover nuclei central to zeolite chemistry such as 23 Na, 27 Al, or 29 Si, nor does it predict electric field gradient (EFG) tensors. The latter is critical from materials science as over 75% of all NMR-active nuclei are quadrupolar. 4 More broadly, early studies using ML for predicting NMR properties focused on predicting scalar quantities (isotropic chemical shifts), and beyond those, it seems timely to make use of the rich structural and dynamical information encoded in tensorial NMR parameters such as chemical shift anisotropy (CSA) and quadrupolar coupling parameters ( C_{Q} , \eta_{Q} ). 3 ; 24 The prediction of these tensorial observables poses a fundamental challenge because, unlike scalar properties, they are not rotationally invariant. Recent progress has been made using rotationally equivariant descriptors such as \lambda -SOAP 19 ; 32 and equivariant graph neural networks 12 ; 60 for predicting NMR observables. In particular, Ben Mahmoud et al. 12 demonstrated that equivariant graph neural networks (eGNN), using the NequIP 10 MLIP architecture at the time, can predict full NMR magnetic shielding and EFG tensors for amorphous SiO 2 with high accuracy. However, whether such approaches can achieve similar accuracy across chemically and structurally diverse material classes, such as aluminosilicate zeolites with varying Si/Al ratios, multiple cation types, and different hydration states has not been demonstrated. In the specific context of zeolites, early works exploited simple correlations between 29 Si and 27 Al isotropic chemical shifts and local geometric parameters such as T – O – T angles. 51 ; 38 Over the years, the method has been used and expanded to predict mainly the 29 Si NMR of siliceous zeolites. Recently, we have employed least absolute shrinkage and selection operator (LASSO) regression to rapidly predict 27 Al and 23 Na chemical shifts in aluminosilicate zeolites, 37 ; 36 which retains structural interpretability but is constrained to low dimensionality and linearity, and requires the user to choose the relevant features. Smooth Overlap of Atomic Positions (SOAP) descriptors combined with non-linear regression methods, such as kernel ridge regression, have been found to sacrifice interpretability but achieve higher accuracy for isotropic shifts across diverse zeolite frameworks. 61 ; 62 ; 18 In this work, we develop a graph-based tensorial model for NMR parameters in aluminosilicate zeolites, capable of predicting complete NMR magnetic shielding and electric field gradient tensors for five nuclei ( 1 H, 17 O, 23 Na, 27 Al, and 29 Si) across the compositional and structural diversity of this highly complex material class. Our model is trained on a diverse dataset spanning pure silica and aluminum-containing zeolites with different Si/Al ratios, protonic and sodium charge-compensating cations, varying water loadings, and multiple framework topologies. We demonstrate

Read the complete text