Challenges with hard-to-learn data in developing machine learning models for predicting the strength of multi-recycled aggregate concrete
Jeonghyun Kim
Abstract
Research on multi-recycled aggregate concrete (MRAC), which involves reusing recycled concrete, has been actively pursued to promote sustainable practices. However, studying the properties of MRAC often requires significant time and resources. Machine learning (ML)-based predictive methods offer a promising solution to overcome these challenges. This study developed and evaluated ML models to predict the compressive strength of MRAC using 197 samples, 8 input features, grid search, cross-validation, and 9 algorithms. The results demonstrated that ML models could achieve high accuracy (R² > 0.9) even without the application of advanced techniques. However, certain data points consistently exhibited high error rates under different ML algorithms , cross-validation methods and data split ratios, suggesting the presence of inherently difficult-to-learn data. The high error likely result from the integration of supplementary cementitious materials , which have contradictory effects on compressive strength , as well as issues arising from small sample sizes of compressive strength at early-age. These results highlight the importance of incorporating domain knowledge when constructing the database to improve model reliability. Finally, a post-analysis was conducted using Shapley Additive Explanations to clarify the relationship between the inputs and output, and recommendations were provided for improving MRAC properties for future research. This study provides valuable insights into the application of ML for predicting concrete properties.