Elemental segregation in alloy materials critically determines their surface and bulk elemental compositions, which fundamentally govern key physicochemical properties, such as catalytic performance, stability, and mechanical characteristics. The decoding of such segregation phenomena is a prerequisite for the rational design of advanced alloy catalysts. While X-ray absorption fine structure (XAFS) serves as a powerful probe for material microstructures, its conventional application suffers from ensemble-averaging limitations in resolving nanoscale compositional heterogeneity. Taking the surface segregation in Pt-Ru alloy nanoparticles as a representative case, we have developed a machine-learning-assisted X-ray absorption near-edge structure (XANES) interpretation framework. This framework enables quantitative deconvolution of surface and bulk phase signals from experimental spectra. In particular, a cascade multiscale convolutional neural network was trained to predict the Pt coordination numbers from XANES fingerprints. Through further computational-experimental spectral deconvolution, surface and bulk XANES profiles were discriminatively resolved by subtracting the theoretically calculated bulk spectrum from the experimental data, according to their respective weights. This process successfully extracted the surface-specific spectral profile. The predictions made by the neural network confirmed that the surface Pt-Pt coordination numbers could be significantly higher than those of the bulk phase. This was accompanied by a reverse trend in Pt-Ru coordination, which conclusively verified the occurrence of Pt surface segregation. This methodology, which integrates machine learning, computational modeling, and experimental probes, provides new paradigms for atomistic-level decoding of physicochemical properties in nanostructured alloys while also demonstrating a new possibility for analyzing localized physicochemical variations within nanomaterials via XAFS.
Zhou et al. (Thu,) studied this question.