An interpretable probabilistic machine learning model for forecasting compressive strength of oil palm shell-based lightweight aggregate concrete containing fly ash or silica fume

Abstract

Forecasting the compressive strength (CS) of oil palm shell (OPS)-based lightweight aggregate concrete (LWAC) heavily relies on machine learning (ML) models. However, these models were primarily trained on limited data, and did not provide any uncertainties or explanations for CS predictions. To overcome these limitations, we compile a dataset of 813 experimental records with 9 input variables for OPS-based LWAC containing fly ash or silica fume. Natural Gradient Boosting (NGBoost), a notable probabilistic machine learning model, is utilized in this study to predict the CS value of OPS-based LWAC. The model is fine-tuned using the collected dataset, and normal distribution-based prediction intervals (PIs) are simultaneously determined to quantify the uncertainties of CS estimates. The results indicate that NGBoost has superb generalization capabilities within cross-validation (average R2/NLL: 0.943/6.111) and satisfactory probabilistic prediction performance on the testing dataset (R2/NLL: 0.969/6.790). To further enhance model transparency, both the mean estimate and its uncertainty are innovatively interpreted using Shapley additive explanation (SHAP). Analyses show that incorporating a reasonable amount of fly ash (<75 kg/m3) does not significantly affect the CS value. However, the addition of silica fume (<150 kg/m3) and superplasticizer (<7 kg/m3) can improve it. Despite its relatively low feature importance for mean predictions, OPS is the most influential factor positively correlated with PIs, primarily due to variations in its size, surface characteristic, and gradation. Additionally, an online graphical user interface based on NGBoost has been created to aid in the probabilistic estimation of CS value and mixture design for OPS-based LWAC. © 2024 Elsevier Ltd