A high-throughput framework for predicting surface density of states from bulk electronic structure

Understanding and predicting the electronic structure at material surfaces is critical for the design of catalysts, semiconductors, and energy interfaces. However, obtaining accurate surface density of states (DOS) remains computationally expensive, requiring slab-based density functional theory (DFT) calculations that are infeasible for high-throughput exploration. We present a data-efficient and physically grounded framework for predicting surface DOS directly from bulk electronic structure. Bulk and surface DOS can be compactly represented using principal component analysis (PCA), revealing aligned low-dimensional manifolds that reflect shared chemical and orbital trends. We train a linear transformation matrix using only three compounds—CuNbS, CuTaS, and CuVS—to map bulk latent features to their surface counterparts. This model is then applied to unseen compositions, including CuCrS, CuMoS, CuTiS, and CuWS and captures key surface DOS features for unseen compositions. This approach enables high-throughput screening of surface electronic properties across chemically diverse design spaces without requiring explicit surface calculations. Additionally, by providing a new approach to spectral data prediction based solely on unsupervised approaches, a new framework for modeling complex data with limited training data is provided.