Compressive strength prediction of metakaolin mortar using CATBoost enhanced with genetic algorithm and particle swarm optimization

Metakaolin is a widely recognized supplementary cementitious material (SCM) among researchers and practitioners due to its ability to improve concrete qualities while reducing carbon emissions. The accurate estimation of the compressive strength (CS) of metakaolin mortar and finding the complex relationship among the mixture proportions and other variables is a difficult task for such a blended mixture. Conventional experimental methods are time-intensive and cost-intensive in nature, which limits the availability of comprehensive datasets. The main aim of the present study was to explore advanced machine learning techniques, namely, the CATBoost algorithm coupled with two metaheuristic optimization algorithms, the genetic algorithm (GA) and particle swarm optimization (PSO), for the prediction of CS of metakaolin mortar mix. For the construction of ML models, a dataset comprising 424 experimental records was collected from published literature, which included parameters such as cement grade, curing age, water-to-binder ratio, sand gradation, and plasticizer dosage. The performance of the base CATBoost model and its optimized variants (CATBoost-GA and CATBoost-PSO) was assessed using several performance indicators: R², WMAPE, NSE, RMSE, MAPE, and MAE. Among all developed models, CATBoost-PSO revealed the highest predictive performance on the test set (R2 = 0.895, RMSE = 6.892 MPa, MAE = 4.421 MPa), outperforming both the CATBoost and CATBoost-GA models. The developed ML models had robust generalization capabilities in predicting the CS of MK mortar. However, the CATBoost-PSO model consistently produced the lowest prediction errors, showing its superior capacity to optimize base CATBoost model hyperparameters. The COM analysis results also demonstrate the resilience and versatility of ensemble learning methods, especially when integrated with optimization techniques. This study not only emphasizes the predictive efficacy of the base and hybrid CATBoost models for precise strength forecasting but also offers a user-friendly graphical user interface for predicting the CS of MK mortar based on the selected input variables of mix design factors.