PSI - Issue 70

Santhosh Kumar N V et al. / Procedia Structural Integrity 70 (2025) 440–446

444

4. Results and discussion This segment compares machine learning models MLR, RF, SVM, XGBoost, and KNN performance in their ability to predict lime mortar compressive strength with using a database containing 381 entries. The statistical information of the inputs i.e. Lime, Ggbs, Redmud, Fine aggregate, Water, Age are already given in Table 1. Models are assessed for accuracy using scatter plots of actual versus predicted value for training and test sets, with comparisons being made on both raw and normalized data ( Fig 1. (a) (b) (c) (d) (e) ). The plots show each model's predictability and ability to generalize, making it possible to determine the best approach to estimate compressive strength. A comparison of Mean Squared Error (MSE), Mean Absolute Error (MAE), and Adjusted R² for prediction using 6 different ml models is presented (Table 2, Table 3). The models based on raw data showed that Multiple Linear Regression (MLR) gave the best results with the smallest MSE (0.8997) and a large Adjusted R² (0.6852), showing its power in identifying linear relationships. Random Forest (RF) presented an average performance with MSE = 1.6490 and an Adjusted R2 of 0.4231. However, template boosting (XGBoost) showed the worst performance with the highest MSE of 2.1968 and an alarming low of 0.2315 for Adjusted R2, suggesting overfitting. Support Vector Machine (SVM) outperformed both RF and XGboost with values of 1.3907 for MSE and 0.5135 for Adjusted R2, indicating reasonable forecasting ability. K-Nearest Neighbors (KNN) offered comparable results, achieving a lower MSE of 1.222 and a higher value of 0.573 for Adjusted R2. Upon normalization, SVM significantly improved, achieving the highest Adjusted R² (0.7994) and the lowest MSE (0.5734), indicating that normalization enhanced its ability to capture relationships. MLR’s performance remained consistent (MSE = 0.8997, Adjusted R² = 0.6842), while RF and XGBoost showed marginal changes. KNN’s performance also improved (MSE = 1.1282, Adjusted R² = 0.6053), with lower MAE values indicating better accuracy. These results underscore the importance of data preprocessing, i.e. normalization notably benefiting non-parametric models like SVM as it uses quadratic optimization solver, normalization speeds the convergence and KNN uses Euclidian distance for prediction, normalization ease the computation. SVM emerged as the top performer, while MLR remained a reliable baseline. Future research should explore hyperparameter tuning and feature engineering to further enhance model performance. Table 2. ML model performances with raw lime mortar dataset 4.1 Prediction of lime-based mortars compressive strength

Model MLR

MSE 0.8997 1.6490 2.1968 1.3907 1.2220

MAE 0.6447 0.7138 0.7902 0.7880 0.6360 MAE 0.6447 0.7099 0.7902 0.4795 0.6147

Adjusted R 2

0.6852 0.4231 0.2315 0.5135 0.5730

RF

XG Boost

SVM KNN

Table 3. ML model performances with normalized lime mortar dataset

Model MLR

MSE 0.8997 1.6643 2.1968 0.5734 1.1282

Adjusted R 2

0.6842 0.4252 0.2315 0.7994 0.6053

RF

XG Boost

SVM KNN