From non-centrosymmetry to surface Janus design: Sparse GW/BSE learning for excitonic two-dimensional photocatalysis candidates

Janus two-dimensional semiconductors are promising excitonic and photocatalytic platforms because chemically distinct top and bottom surfaces can generate out-of-plane polarity, tune band alignment, and promote charge separation. Database-scale discovery, however, faces two linked obstacles. Crystallographic non-centrosymmetry does not imply true surface Janus asymmetry, so broad non-centrosymmetric searches can substantially overcount relevant materials. In parallel, sparse GW and Bethe-Salpeter equation (BSE) labels leave high-level excitonic screening incomplete. Here we develop a boundary-aware C2DB workflow that integrates structural surface classification, sparse GW/BSE label completion, formula-group validation, uncertainty-aware band-edge screening, and candidate audit. From 805 non-centrosymmetric ternary two-dimensional semiconductors, the structural classifier identifies 229 strict surface-Janus-like entries, while 506 are surface-symmetric despite lacking inversion symmetry. Within the strict subset, formula-group validation yields mean absolute errors of 0.087 eV for GW band gaps, 0.113 eV for GW valence-band maxima, 0.081 eV for GW conduction-band minima, and 0.091 eV for BSE exciton binding energies. Uncertainty-aware pH 0 screening retains 48 candidates, and a practical 1.6-3.2 eV gap window further narrows this set to 18 priority entries, evenly divided between direct GW/BSE-supported candidates and machine-learning-completed validation targets. The strongest direct candidates are AsISe and AsITe, while IrClS stands out as a clean machine-learning-priority target for future GW-BSE validation. Rather than replacing high-level electronic-structure calculations, this atlas defines a validation-readiness map for surface-Janus-like two-dimensional semiconductors and clarifies the boundary between structural polarity, sparse excitonic labels, and uncertainty-aware photocatalytic band-edge screening.