Decoding Crystallographic Surface Chirality with Machine Learning: From Atomic Geometry to Fermi Surface Projections
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. We report a dual-domain machine learning framework that decodes crystallographic surface chirality from two independent image representations: atomic structure models in real space and simulated momentum-resolved photoemission maps of the Fermi surface projections in reciprocal space. ResNet18, a deep convolutional neural network, fine-tuned on a database of labeled images achieves ~73% classification accuracy on atomic models and ~99% on Fermi surface projections. We show that the latter transfers directly to synchrotron-acquired experimental images after fine-tuning on just two labeled frames. We identify a working correspondence between the two representations: just as the kink site geometry fixes the orientation of crystallographic planes in real space, the surface normal position in a momentum-resolved photoemission map anchors the orientation of the Fermi surface polygons in reciprocal space. It is precisely this relative orientation that encodes handedness into the map topology with high accuracy. The pronounced difference in accuracy shows that handedness is more readily recovered from the momentum-space electronic pattern than from the local atomic geometry of the kinked surface. This finding has direct implications for the disorder resilience of geometric chiral-induced spin selectivity (CISS) at realistic metal surfaces.