Particle packing theory-guided multi-fidelity deep learning for discovering low-carbon cost-effective high-performance concrete

Data-driven methods for concrete design offer time and cost efficiencies but lack integration of concrete-specific knowledge, compromising model reliability and generalizability. This paper presents a framework to integrate data and concrete theories into the design of low-carbon, cost-effective, high-performance concrete. This research has three primary novelties: (1) Morphology of concrete ingredients, such as solid wastes, which largely affect the fresh and hardened properties of concrete, is considered based on a particle packing theory. (2) Data scarce challenge is addressed using a multi-fidelity strategy, which integrates a Compressible Packing Model, Discrete Element Model, and experimental data. (3) Compliance with concrete theories is enforced using a Physics-Informed Neural Network. The proposed approach achieves prediction accuracy of 98 %, and reduces cost and carbon emission by 29 % and 50 %, respectively, compared with traditional methods. These results demonstrate the framework's potential to accelerate sustainable concrete design in real-world applications.