Deep learning reveals a correlation between local composition and Majorana zero modes in FeTe_1-xSe_x

Majorana zero modes (MZMs) in topological superconductors offer a promising route towards faulttolerant quantum computation. Among candidate materials, the iron-based superconductor FeTe_1-xSe_x has recently emerged as a robust and relatively clean platform for realizing and exploring MZMs. Yet MZMs in this material emerge only in a subset of magnetic vortices, and the origin of this spatially inhomogeneous yield remains unresolved. Here we develop a deep-learning framework for the in situ, high-throughput and standardized mapping of the top-layer Te/Se composition directly from scanning tunnelling microscopy topographs. By achieving atomically resolved compositional mapping with high fidelity, we quantify the local chemical disorder surrounding vortex cores. Combining these AI-derived atomic maps with MZM classifications obtained independently from vortex-core spectroscopy, we statistically analyse 61 vortices and find that MZMs are strongly enriched within a narrow window of local Te concentration, whereas vortices outside this window predominantly host Caroli-de Gennes-Matricon bound states without a zero-bias peak. These results identify local composition as a key statistical factor governing the uneven MZM yield in FeTe_1-xSe_x, and establish an AI-assisted strategy for identifying and optimizing topological superconductors for device applications.