Data-Efficient Hydrogen Adsorption Prediction in Copper Nanoclusters: A Computer Vision-Based Transfer Learning Approach

This study investigates the size-dependent hydrogen $\left(\mathrm{H}_{2}\right)$ storage capacity of magic-sized copper nanoclusters (Cu NCs) by analyzing their structural and electronic properties along with formation energy. Density functional tight binding (DFTB) calculations were performed to determine the HOMO and LUMO energy levels, energy gap ($E_{g}$), Fermi level ($E_{f}$), and formation energy ($E_{F}$) for different $\mathbf{C u}$ NC sizes interacting with H<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>. The results reveal a strong correlation between cluster size and H<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> storage potential, with smaller clusters exhibiting more favorable adsorption characteristics. However, as the system size increases, the computational cost of DFT calculations rises significantly. To address this, an machine learning approach was employed using image-based representations of Cu NCs with transfer learning. This method enabled rapid predictions of electronic properties with DFT-level accuracy while significantly reducing computational time. The model converged in just a few epochs, capturing the size-dependent trends in H<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> adsorption. These findings highlight the potential of AI-driven techniques for accelerating material discovery and optimizing nanoscale hydrogen storage systems.