Simulation-Based Transfer Learning for Concrete Strength Prediction

Data-driven techniques, such as machine learning, have shown their superiority in the design and development of concrete materials by accounting for their inherent complexity autonomously. Their wider adoption, however, has been hindered by the availability of large, reliable data and the model generalization performance to new data. In this paper, we propose a simulation-based transfer learning framework, where machine learning models are pre-trained on simulated results from physics-based models and then fine-tuned on target datasets. Compared with existing transfer learning approaches, the proposed method does not require real datasets for pre-training but instead takes advantage of prior domain knowledge embedded in existing physics-based models. The effectiveness of the proposed framework is investigated in the context of concrete compressive strength prediction. Results demonstrate that our framework enables desirable prediction accuracy, speeds up the learning process, and improves generalization ability to real-world data, even with a small amount of training samples. The proposed framework offers a practical solution to data scarcity in many scientific and engineering domains, where various physical models can be utilized as an additional knowledge source.