Comparative evaluation of machine learning algorithms for predicting the speed of sound in electrolyte solutions
Abstract
Accurate prediction of the speed of sound in electrolyte solutions is important for understanding physicochemical behavior and improving industrial process design. In this study, several advanced machine learning models, including Decision Tree (DT), AdaBoost, Random Forest (RF), K-Nearest Neighbors (KNN), Gradient Boosting (GB), Ensemble Learning (EL), Support Vector Machine (SVM), XGBoost, and CatBoost, were developed and evaluated to predict the speed of sound in electrolyte solutions. Eight physicochemical parameters, namely molar mass, density, cation radius, anion radius, lattice energy, hydration energy, electrical conductivity, and molarity, were used as input variables based on a dataset of 125 experimental observations. Pearson and Spearman correlation analyses showed that lattice energy and molarity had the strongest positive relationships with the speed of sound, whereas hydration energy, conductivity, and density exhibited inverse relationships. Model performance was assessed using the correlation coefficient (R), RMSE, MAE, MBE, DR, and SI. Among all models, Gradient Boosting demonstrated the most balanced predictive performance during the testing stage with R2 = 0.9624, RMSE = 4.4884, and MAE = 2.4064, indicating excellent prediction accuracy and generalization capability. SHAP analysis revealed that lattice energy was the most influential parameter affecting the predicted speed of sound, followed by molarity and anion radius. Overall, the proposed machine learning framework provides an accurate and interpretable data-driven tool for estimating the speed of sound in electrolyte solutions and offers valuable insight into the physicochemical factors governing acoustic behavior in aqueous electrolyte systems.