Abstract
This study focuses on predicting the acid rain damage using machine learning techniques. First, a database containing 168 datasets was established based on laboratory experiments on seven types of carbonate building stones. Laboratory experiments were conducted at three different pH levels, using sulfuric acid (H₂SO₄) and nitric acid (HNO₃) to simulate acid rain conditions. Furthermore, during experiments, P-wave velocity was used to define the acid rain damage index (D). Then, to predict D efficiently and accurately, the M5 Prime (M5P) algorithm was employed as a white-box and interpretable machine learning technique. The predictive capability of developed M5P model was assessed by several performance assessment metrics. Based on obtained results, the suggested model predicted D with good accuracy and acceptable errors. The feature importance quantification results in M5P model demonstrated that stone's porosity is the most influential variable on acid rain damage. Finally, predictive capability of M5P model was compared with support vector regression (SVR) and random forest (RF) models. This comparison showed that the proposed M5P model outperforms other models in terms of D prediction accuracy, with a determination coefficient (R(2) of 0.891, a root mean square error (RMSE) of 0.021, and a mean absolute error (MAE) of 0.017. Thus, the proposed M5P model offers the construction industry a reliable framework for evaluating acid rain damage of building stones, reducing the need for extensive experimental testing, and facilitating the selection of durable stones for exterior cladding and floor coverings in acid rain-prone areas.