Multi-agent collaboration for knowledge-guided data-driven design of ultra-high-performance concrete (UHPC) incorporating solid wastes

Data-driven design of concrete attracts increasing interests in waste valorization and decarbonization but lacks generalizability and reliability without concrete domain knowledge. Recent research suggests that knowledge graphs are promising for imparting concrete knowledge into data-driven design, yet manual construction of knowledge graphs is inefficient and hard to scale. This paper presents a multi-agent collaboration framework to streamline knowledge-guided data-driven design of green concrete. The framework decentralize design tasks among specialized agents, and a large language model-based approach is developed to automate the extraction of concrete knowledge for constructing concrete knowledge graphs. The framework has been applied to create a knowledge graph and design green ultra-high-performance concrete (UHPC). The primary novelties of this research involve the multi-agent collaboration framework for designing UHPC and the automatic extraction of UHPC knowledge for constructing the knowledge graph. Results show that concrete knowledge is imparted into data-driven design of UHPC and enables explicit interpretation of machine learning outcomes regarding physical and chemical mechanisms, advancing the transition from purely data-driven to knowledge-guided design of eco-friendly composite materials.