What question did this study set out to answer?

The aim is to develop a machine learning framework that predicts the mechanical performance of concrete with recycled concrete aggregate.

February 11, 2026Open Access

Physics-guided machine learning framework for RCA concrete by experimental database, modelling, and statistical validation

Key Points

The aim is to develop a machine learning framework that predicts the mechanical performance of concrete with recycled concrete aggregate.
Utilized RCA from six parent strength classes and three particle-size fractions at various replacement levels.
Established a comprehensive experimental database for RCA performance.
Trained gradient-boosting models (XGBoost and LightGBM) using a physics-guided data augmentation strategy.
Assessed model performance using metrics like R², RMSE, and MAE.
The framework achieved strong predictive accuracy for compressive strength and stable performance.
RMSE values remained within acceptable experimental ranges, even at higher RCA replacement levels.
Statistical analyses confirmed significant strength reductions with increased RCA content.

Abstract

The search for sustainable construction materials has increased the use of recycled concrete aggregate (RCA); however, variability in mechanical performance and the absence of accurate forecasting tools limit its utilization. A complete laboratory program utilized RCA from six parent strength classes (C5–C50) and three particle-size fractions (0–4 mm, 5–11 mm, 12–22 mm) at 10%–50% replacement levels. The mechanical performance of concrete incorporating RCA is strongly influenced by the aforementioned parameters. This study proposes a physics-guided machine learning (ML) framework to predict the compressive strength (CS) and splitting tensile strength (STS) of RCA concrete, explicitly addressing data sparsity and RCA-induced variability. A comprehensive experimental database covering multiple RCA replacement levels, particle-size fractions, and parent concrete strength classes was established, and gradient-boosting models (XGBoost and LightGBM) were trained using a physics-guided data augmentation strategy in which synthetic samples were generated within physically admissible bounds derived from experimental variability. The model’s performance was assessed using R², RMSE, and MAE, with supplementary error stratification across RCA regimes. The results show that the proposed framework achieves strong predictive accuracy and stable performance, with RMSE values remaining within ranges considered acceptable experimentally, even at higher RCA replacement levels. The accompanying statistical analyses confirm that the observed reductions in strength associated with increasing RCA content are both statistically significant and physically meaningful. Compared with conventional augmentation approaches, the physics-guided strategy improves model robustness in sparsely populated regions of the dataset without introducing non-physical relationships. As a result, the framework provides a transparent, physically consistent machine-learning approach for data-driven mixture design and the development of sustainable concrete systems.

Read Full Paperexternally

Perguntar à IA

Bookmark

View Full Paper