Carbon nanotube-induced magnetic shielding effects on 129Xe NMR from equivariant neural networks.

Equivariant graph neural networks (EGNNs) have been successfully used in building efficient and accurate machine learning interatomic potentials (MLIPs) for molecular dynamics (MD) and to predict nuclear magnetic resonance (NMR) parameters in complex systems. In this work, EGNNs are used to fine-tune an atomistic foundation model, which serves as an MLIP for large-scale molecular dynamics simulations of xenon in carbon nanotubes. In addition, EGNNs are employed to develop a machine-learning model for the 129Xe NMR magnetic shielding tensor, σ, enabling its direct prediction from the MLMD snapshots. Both models are data-efficient and remain accurate well beyond their training data. Using this dual-model framework, we show that xenon encapsulated in carbon nanotubes experiences a nanotube-induced, negative 129Xe isotropic chemical shift that varies non-monotonically with nanotube diameter and depends further on temperature and Xe loading, establishing this dual-model approach as an efficient route to interpret confinement-driven 129Xe NMR response.