Machine Learning Techniques for Low-Carbon Concrete Strength Prediction

Abstract

The strength of Low-Carbon Concrete (LCC) is influenced by multiple factors such as material composition, curing conditions, and Carbon Emission (CE) parameters, making it difficult to capture the variation patterns of strength under the influence of multiple factors. To address this, this study proposes a Low-Carbon Concrete Strength Prediction (LCCSP) method based on Machine Learning (ML) techniques. By modeling the CEs of concrete and identifying three key factors affecting the performance of LCC—age, Water-Cement Ratio (W/C), and the number of Cementitious Materials (CMs)—these factors are used as input samples for an LCCSP model based on Self-Organizing Map-Radial Basis Function Neural Networks (SOM-RBF NNs). First, the SOM network is employed to cluster and extract features from multi-dimensional input data (age, W/C, number of CMs). Then, the cluster centers and radii are used as the Hidden Layer (HL) parameters of the Radial Basis Function Neural Network (RBFNN) to enhance its ability to handle complex nonlinear relationships. The model predicts the strength of LCC based on the input samples. In experiments, the Mean Absolute Error (MAE) and the Coefficient of Determination (R²) are used as evaluation metrics. The results show that the MAE remains relatively small at various time points, and the R² consistently approaches 1, demonstrating good adaptability and robustness for both single-factor and multi-factor inputs, providing more accurate and stable technical support for LCCSP.