A machine-learning approach to high-throughput spatially resolved ARPES measurements

Spatially resolved angle-resolved photoemission spectroscopy (ARPES) is a powerful technique for mapping local electronic structures, but it requires long measurement times to achieve sufficient signal-to-noise ratio (SNR) at each spatial point. This constraint limits beamtime efficiency and exacerbates sample degradation. In samples with domain structures, however, similar spectra are repeatedly measured across spatial positions, creating substantial redundancy in spatially resolved ARPES datasets. Here we systematically evaluate how SNR degradation from reduced acquisition times affects unsupervised domain classification. PCA-based dimensionality reduction maintained high agreement with the high-SNR reference classification even under low-SNR conditions, whereas direct clustering of spectral vectors showed rapid degradation beginning at a relatively high-SNR regime. Because dimensionality reduction is applied solely for domain classification, the representative spectrum of each domain is obtained as a direct average of the measured data. These results demonstrate that combining short-acquisition measurements with intra-domain spectral averaging through the co-addition of spectra within each classified domain can substantially accelerate spatially resolved ARPES while preserving classification accuracy.