@Article{cmes.2026.085421,
AUTHOR = {Masood Karamoozian, Zarrin Mahdavipour, Mohammed Ameen, Faisal Binzagr, Abdolraheem Khader, Ahmed Hamza Osman, Ali Ahmed},
TITLE = {Hybrid Physics-Informed Graph Neural Network Surrogate for Multi-Physics Optimization of Net-Zero Prefabricated Modular Buildings Driven by BIM Semantic Data},
JOURNAL = {Computer Modeling in Engineering \& Sciences},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/CMES/online/detail/27712},
ISSN = {1526-1506},
ABSTRACT = {Achieving net-zero carbon emissions in the building sector requires computational tools that are simultaneously efficient, physically rigorous, and scalable to complex multi-domain design spaces. Traditional physics-based simulators such as EnergyPlus and finite-element analysis (FEM) packages deliver high fidelity but are computationally prohibitive for large-scale parametric exploration and real-time multi-objective optimization. This study introduces a hybrid Physics-Informed Graph Neural Network (PI-GNN) surrogate that directly leverages semantic Building Information Modeling (BIM) data to enable rapid, accurate, and physically consistent multi-physics prediction and optimization of prefabricated modular buildings. Building components and their physical and functional relationships are encoded as nodes and typed edges in a heterogeneous directed graph, processed by a multi-layer graph attention network (GAT) backbone augmented with physics-informed loss terms. These constraints explicitly enforce energy and mass conservation, transient heat transfer, structural equilibrium, and carbon balance, ensuring physically plausible predictions even under limited training data. Trained on a dataset of more than 852 BIM-derived parametric models coupled with high-fidelity multi-physics simulations (EnergyPlus, OpenSeesPy FEM, One Click LCA), the surrogate achieves mean absolute percentage errors (MAPE) below 7% across energy demand, embodied carbon, operational carbon, and peak structural stress, while delivering inference speedups of up to 1200<mml:math id="mml-ieqn-1"><mml:mo>×</mml:mo></mml:math> over conventional simulation workflows. Integrated with NSGA-II multi-objective evolutionary optimization, the framework identifies Pareto-optimal designs that reduce operational energy by up to 52% and embodied carbon by 28%–45% relative to baseline configurations. Ablation studies confirm that physics-informed regularization reduces prediction errors by 40%–55% in low-data regimes. The proposed PI-GNN offers a generalizable, interpretable, and physics-consistent surrogate modeling paradigm, bridging BIM semantics with deep learning to accelerate sustainable design and support the global transition toward net-zero prefabricated buildings.},
DOI = {10.32604/cmes.2026.085421}
}