DeepUMQA-PA is a single-model method based on physical-aware features combined with the deep learning network. We constructed the Voronoi contact surface of the heavy atoms in the protein complex to aware the physical tendencies, and characterized the contact properties between the residues based on the molecular hydrophobicity and the orientation of the contact surface. Additionally, ultrafast shape recognition features, voxelization features, and large language model features are combined with the invariant point attention mechanism network to predict the per-residue quality of protein complexes.