Akademik Veri Yönetim Sistemi
International Journal of Hydrogen Energy, cilt.201, 2026 (SCI-Expanded, Scopus)
The hydrogen storage behavior of copper (Cu) nanoclusters (NCs) with magic sizes is critically influenced by quantum size effects arising from the interplay between atomic geometry, electronic structure, and thermodynamic stability. In this study, we investigate a representative series of hydrogen-interacting Cu NCs (Cux–H2, x = 0.7–1.0 nm) using a physics-guided workflow, enhanced by a multimodal explainable artificial intelligence (MXAI) framework. This approach integrates three complementary representations: atomic coordinates embedded via continuous-filter interaction layers, simulated electron-microscopy images processed through a residual vision encoder, and tabulated compositional descriptors analyzed using graph convolutional networks. A temperature-scaled attention mechanism fuses these streams to predict key electronic and energetic properties associated with hydrogen adsorption, including HOMO–LUMO levels, band gap, total interaction energy, and formation energy. Additionally, we calculate the hydrogen binding energy per atom (Ebind) and entropy-corrected heat of adsorption (ΔHads) to assess thermodynamic feasibility under ambient conditions, along with the equilibrium Cu–H2 distance (d) to capture structural interaction features. Our results reveal a strong size-dependent trend: smaller NCs exhibit enhanced hydrogen binding and larger electronic rearrangements, consistent with quantum confinement effects. The MXAI model not only replicates these trends with millisecond-level inference speed, referring specifically to the neural-network forward pass once geometric, tabular, and image inputs have been constructed, but also provides interpretable attribution maps that identify Cu content and Cu–H2 coordination as dominant predictive factors. Because all DFTB calculations required to generate the target properties and tabular descriptors are performed offline during dataset construction, the MXAI framework enables genuine high-throughput prediction: once trained, it can evaluate large numbers of unseen nanocluster candidates without invoking any additional electronic-structure calculations. By enabling rapid, explainable screening of NC configurations, this hybrid framework offers a powerful tool for guiding the rational design of next-generation hydrogen storage materials.