Decoding Crystallographic Surface Chirality from Real Space and Momentum Space Images by Machine Learning
Intrinsically chiral metal surfaces, where handedness arises from the asymmetric step-kink-terrace topology of high-Miller index planes, are model systems for enantiospecific catalysis, sensing, and spintronics. Yet no consistent method exists to classify their handedness directly from experimental observables, without prior knowledge of crystallographic indices. Here, we report a dual-domain machine learning framework that recognizes handedness in chiral metal surfaces from two independent image representations: atomic structure models in real space and simulated momentum-resolved photoemission maps of Fermi surface projections in reciprocal space. A ResNet18 model pretrained on general image data is fine-tuned for chirality classification independently on both image representations and achieves meaningful classification accuracy for both real space and reciprocal space images. However, the Fermi surface classifier significantly outperforms the real space classifier. Critically, the reciprocal space classifier, trained only on synthetic images, correctly identifies synchrotron-acquired experimental Angle-Resolved Photoemission Spectroscopy (ARPES) maps of Cu(643)$^R$ and Cu(643)$^S$. The stronger performance in momentum space, as well as the transfer to experimental data, shows that handedness is encoded more globally and robustly in the electronic structure than in the spatially localized kink-site geometry. These results establish ARPES as a quantitative readout of crystallographic surface chirality and suggest a scalable route to identify chiral metal surfaces relevant to spin-selective phenomena.