AI-empowered prediction of office building energy use from single-view conceptual images for early-stage design

The building sector, particularly office buildings, accounts for a significant share of global energy consumption due to extended operational hours and intensive equipment use. While early-stage design decisions such as form, orientation and spatial layout exert a great influence on energy performance, existing energy prediction tools typically require complete 3D models, making them impractical for early design workflows that heavily rely on fast, iterative 2D conceptual images. This paper addresses the gap between early-stage architectural representation and energy performance feedback by asking: Can accurate energy consumption be predicted directly from a single 2D design image? To address this, we introduce ArchEnergy, a diffusion-based framework that estimates energy use directly from single-view images consisting of two stages: (1) a diffusion model combined with sparse-view reconstruction to generate a 3D mesh from a single conceptual image, and (2) a voxel-based 3D CNN incorporating masked autoencoding and contrastive learning for precise energy estimation. To support this framework, we construct ArchiMeshNet, a dataset of 6784 architectural 3D models with EnergyPlus-simulated consumption data. Experimental results demonstrate that our ArchEnergy achieves high predictive accuracy (R2 = 0.946, MAE = 0.443 kWh/m2·a), outperforming baseline models across both 3D reconstruction and energy prediction. ArchEnergy empowers architects to receive timely, geometry-informed energy feedback early in the design process, without the need for full 3D modeling, bridging a critical gap in existing workflows. By tightly coupling conceptual representation with performance prediction, this study lays the foundation for future research on intuitive, performance-driven design assistance and promotes sustainable architectural practice. Project Page: https://github.com/ThuYinJun/ArchEnergy-main.