Learning Work Function via Implicit Reasoning on Electrostatic Potential Landscapes

ABSTRACT Precise work function engineering in two‐dimensional (2D) materials is pivotal for next‐generation nanoelectronic devices. However, current data‐driven approaches are often hampered by the scarcity of high‐precision data and a lack of physical interpretability. We propose a Graph‐Potential Cross‐Modal Contrastive Learning framework designed to uncover correlations between crystal and electronic structures. Rather than performing a direct scalar mapping, our approach respects the strict thermodynamic definition of the work function (Φ = E vac − E Fermi ). By extracting the vacuum level from 1D PAEP morphology and predicting the Fermi level via an auxiliary head, the model accurately predicts work functions ( R 2 = 0.902). This indicates an automatic extraction of features governing electron escape barriers. Additionally, the model demonstrates exceptional fidelity in morphological reconstruction; predicted skewness and kurtosis of the potential surface show near‐perfect linear correlation with DFT data ( R 2 > 0.98), proving that it successfully decodes microscopic charge distribution details. This cross‐modal alignment paradigm drives artificial intelligence to transcend simple numerical fitting and learn physically informative representations, facilitating future potential‐contour‐based inverse material design.