AI-based forecasting of groundwater corrosion and scaling indices in semi-arid regions using 25-year data analysis.

Assessing water corrosivity indices is vital for sustainable management, since it damages infrastructure, increase costs, and threaten public health. In this study, the corrosive and scaling behavior of groundwater was modeled and predicted using, Artificial Neural Networks (ANN), Support Vector Machines (SVM), Multivariate Adaptive Regression Splines (MARS), and Random Forest (RF). Three indices were employed: the Langelier Saturation Index (LSI), the Ryznar Stability Index (RSI), and the Puckorius Scaling Index (PSI). The models were developed using 25 years of daily groundwater data from the Dezful-Andimeshk plain in southwestern Iran. In the study area, LSI values ranged from - 8.91 to 0.27, RSI from 8.46 to 18.72, and PSI from - 5.83 to 3.62, indicating that groundwater exhibits both corrosive and scaling tendencies depending on location. SAR (Sodium Adsorption Ratio), pH, and TDS (Total Dissolved Solids) were used as input variables for model development. Among the tested algorithms. The SVM model, performance metrics during the testing phase were as follows: LSI (R² = 0.92, RMSE = 0.11), RSI (R² = 0.81, RMSE = 0.21), and PSI (R² = 0.82, RMSE = 0.21). Overall, model comparisons indicated that all four algorithms achieved acceptable accuracy (R² = 0.80-0.93). However, MARS and ANN consistently provided superior and more stable performance, effectively capturing nonlinear and interactive relationships among the predictors. RF produced competitive results but did not show clear dominance over the other models in this dataset.