A GAN-physical simulation coupled framework for joint optimization of building energy prediction and spatial configuration

Building energy optimization faces challenges in simultaneously addressing spatial design and energy performance prediction. This paper proposes a novel framework integrating generative adversarial networks (GANs) with physics-based simulation engines to achieve joint optimization of building energy consumption prediction and spatial configuration. The framework establishes bidirectional coupling between conditional GAN architectures and EnergyPlus thermodynamic calculations, embedding physical constraints directly into the generative process through a physics-informed discriminator. Experimental validation across office, residential, and educational building typologies demonstrates superior performance, achieving 6.8% mean absolute percentage error in energy prediction and 23.7% average energy consumption reduction compared to code-compliant baselines. The model generates well-distributed Pareto fronts containing 42-56 non-dominated solutions, outperforming conventional sequential optimization approaches by 8-12 percentage points while maintaining computational efficiency. This hybrid methodology advances sustainable building design by transcending limitations inherent to purely data-driven or physics-only approaches, providing a scalable solution for intelligent architectural optimization.