Discovery of flat-band 2D materials via physics-informed scoring and structure-based learning.

Flat electronic bands are a fertile ground for exotic quantum phenomena, from unconventional superconductivity to fractional topology. However, their rarity and the reliance on density functional theory (DFT) calculations for identification have limited systematic exploration across large material spaces. Here, we report the discovery of multiple previously unrecognized flat-band two-dimensional materials through a structure-informed, data-driven framework. We introduce a physically motivated flatness score that combines band dispersion and density-of-states features, enabling algorithmic labeling of known materials without requiring manual inspection. This score is then used to train a multimodal deep learning model that predicts flat-band propensity directly from atomic structure. Applied to more than 10,000 unlabeled materials, our framework identifies numerous flat-band candidates. Among the top-scoring structures (flatness score higher than 0.9), DFT calculations on materials with kagome-like lattices confirm a flat-band prediction accuracy of 98%, demonstrating the ability to discover topologically nontrivial systems. By bypassing the need for precomputed band structures, our approach enables large-scale, interpretable screening and expands the design space for correlated quantum phases in low-dimensional systems.