Predicting Hydrogen Uptake in Carbon Nanotubes Using Graph Neural Networks and Molecular Simulations

Hydrogen storage in carbon nanomaterials remains a critical challenge for clean energy applications. In this work, we develop a machine learning framework to accurately predict hydrogen uptake in single-walled carbon nanotubes (SWCNTs) by combining Grand Canonical Monte Carlo (GCMC) simulations with a graph neural network (GNN). A dataset of 270 configurations spanning diameters from 0.68 to 1.36 nm, temperatures of 77–350 K, and pressures up to 100 bar was generated using classical Lennard-Jones potentials validated against established force fields. The GNN, operating directly on atomic graph representations, achieves excellent predictive performance with a test R² of 0.98 and RMSE below 0.55 mmol/g. Interpretability analysis via attention mapping reveals that the model preferentially weights carbon atoms on the inner nanotube surface, consistent with the physics of nano confinement-driven physisorption. Our simulations confirm an optimal SWCNT diameter of ~0.81 nm for hydrogen storage at 77 K (~3.2 wt%), while ambient-temperature capacities remain below 0.5 wt%, highlighting the intrinsic limitations of pristine nanotubes. This work demonstrates that physics-informed machine learning, trained on modest yet high-fidelity simulation data, can enable rapid, interpretable, and reliable virtual screening of nanomaterials for hydrogen storage.