Compressive strength prediction of coconut fiber reinforced concrete using PSO optimized explainable machine learning

Area of Science:

• Materials Science
• Civil Engineering
• Data Science

Background:

• Natural fibers enhance concrete properties like ductility and crack resistance.
• Coconut fiber in concrete can lead to complex interactions affecting compressive strength prediction.
• Accurate prediction is crucial for optimizing the mix design of coconut fiber reinforced concrete (CFRC).

Purpose of the Study:

• To develop an optimized and explainable machine learning framework for predicting CFRC compressive strength.
• To evaluate the performance of six regression algorithms optimized with particle swarm optimization (PSO).
• To identify key predictors influencing CFRC compressive strength through explainability techniques.

Main Methods:

• Six regression algorithms (SVM, KNN, RF, LGB, XGB, ANN) were optimized using PSO.
• A dataset of 586 experimental samples with eight input parameters was utilized.
• A nested grouped cross-validation (5-fold outer, 3-fold inner) and a 30% test set were employed for model evaluation.
• Explainability methods (SHAP, PDP, ALE, ICE) were used to interpret model predictions.

Main Results:

• The optimized XGBoost (XGB-PSO) model achieved high performance with R²=0.963 and MAPE=7.744% during cross-validation.
• The XGB-PSO model showed strong generalization on the test set (R²=0.953, MAPE=8.875%).
• Model explainability consistently highlighted age and recycled coarse aggregate (RCA) as the most significant predictors.

Conclusions:

• The developed XGB-PSO framework provides a robust and accurate method for predicting CFRC compressive strength.
• The findings support data-driven approaches for CFRC mix design and optimization.
• Age and RCA are identified as critical factors for enhancing CFRC performance.