Machine learning models for NMR prediction typically target a single scalar — the isotropic chemical shift. But a nucleus in a solid experiences the full anisotropy of its magnetic environment: the shielding is a tensor, not a number. The chemical shift, its anisotropy, and the asymmetry parameter are three independent observables extracted from a single mathematical object.
The paper on tensorial machine learning for zeolite NMR (arXiv: 2603.22268) trains a model that predicts the complete magnetic shielding and electric field gradient tensors for multiple NMR-active nuclei — aluminum, silicon, oxygen, sodium, hydrogen — across a diverse dataset of zeolitic materials.
The method leverages the equivariant structure of the tensors: they transform in known ways under rotation, and the model respects this by construction. The predictions recover not just isotropic shifts but the full spectral line shapes, including the quadrupolar broadening from electric field gradients that dominates aluminum and sodium spectra.
For the RTH zeolite, the model generates complete predicted spectra for aluminum-27 and silicon-29 directly from the predicted tensors — bypassing the usual pipeline of first-principles calculation entirely.
The through-claim: predicting the tensor instead of the scalar doesn't just add information — it changes what the model must learn. A scalar prediction can succeed by capturing local coordination statistics. A tensor prediction must capture the directional character of the electronic environment — the way bonding geometry, framework topology, and long-range order conspire to orient the principal axes of the shielding. The model that predicts the full object is forced to understand the structure more deeply than the model that predicts only its trace.
2603.22268. Solid-state NMR / machine learning / zeolites / tensor prediction / magnetic shielding.