Beyond Black-Box: Human-Centric and Physics-Informed AI for Integrated and Autonomous Construction

As global demand for urban infrastructure collides with severe labor shortages and urgent decarbonization goals, industrialized construction has emerged as a critical paradigm shift for the AEC industry. However, the transition toward fully integrated and autonomous systems remains stalled by a fundamental technical hurdle: distributional shift. Unlike highly structured digital environments, construction and manufacturing are defined by extreme customization, where AI models trained on specific projects often fail to generalize to the unique data distributions of new, unseen scenarios. This talk introduces a research framework designed to bridge this customization gap through the integration of Human-Centric and Physics-Informed Machine Learning. The approach simultaneously advances three parallel research frontiers: the development of Physics-Informed Machine Learning that embeds IFC-based building knowledge graphs, causal relationships, and physical laws into algorithms to ensure robustness and explainability; the advancement of computational frameworks to capture, explain, and transfer tacit human knowledge, formalizing and embedding high-dimensional heuristic knowledge into machine learning algorithms to stabilize customized manufacturing; and the establishment of a Human-AI Co-evolution model where autonomous systems and human practitioners form a symbiotic, adaptive relationship. By synthesizing structured building data with physical laws and human knowledge, this research provides the theoretical and practical foundations for a future of integrated and autonomous construction that is not only resilient and scalable, but also fundamentally aligned with societal needs and the sustainable evolution of global infrastructure.

Speaker: Miaosi Dong, UC Berkeley