Prediction And Analysis Of Concrete Compressive Strength By Machine Learning Methods
TL;DR
This work presents a systematic multi-model comparison within a unified hyperparameter optimization framework and highlights that boosting-based machine learning models not only achieve high accuracy but also provide interpretable and robust predictions when evaluated through comprehensive error and explainability analyses.
Abstract
Concrete compressive strength (CCS) is a critical parameter directly affecting the load-bearing capacity, durability, and overall safety of engineering structures. Traditional experimental approaches for determining CCS are time-consuming and costly, making predictive models an attractive alternative. In this study, thirteen different machine learning algorithms were applied to a well-established dataset (1030 samples, 8 input parameters) to estimate concrete compressive strength. Unlike many previous studies using the Yeh dataset that primarily emphasize prediction accuracy of individual models, this work presents a systematic multi-model comparison within a unified hyperparameter optimization framework. In addition to conventional performance metrics, permutation importance and SHAP-based explainability analyses are jointly employed, and detailed error evaluations are conducted across curing age and water-to-binder ratio subgroups to enhance engineering interpretability. Among the models tested, the CatBoost algorithm demonstrated the highest predictive performance (R² = 0.9469, RMSE = 3.70), followed closely by XGBoost, Gradient Boosting, and a stacking ensemble model. The results highlight that boosting-based machine learning models not only achieve high accuracy but also provide interpretable and robust predictions when evaluated through comprehensive error and explainability analyses.