Development of SVM-based machine learning model for estimating lornoxicam solubility in supercritical solvent

Abstract

This paper investigates the application of Support Vector Regression with Quadratic Kernel (QSVR) for modeling the solubility of lornoxicam in supercritical carbon dioxide. The dataset comprises temperature (T) and pressure (P) as input variables, while the solubility (Y) of lornoxicam serves as the output variable entire the modeling. The temperature is measured in Kelvin (K), and the pressure is measured in bar as the inputs of models. To improve the predictive performance of the QSVR model, three distinct hyper-parameter optimization techniques, namely Genetic Algorithm (GA), Tabu Search (TS), and Bayesian Hyperparameter Optimization (BHO) are employed. These optimization methods are utilized to fine-tune the hyper-parameters of the QSVR model and enhance its predictive accuracy. The BHO-QSVR model achieved an impressive R2 score of 0.96725, indicating a strong fit between the predicted and actual solubility values. Additionally, it exhibited a Mean Absolute Error (MAE) of 1.75666E-05 and a maximum error of 3.02849E-05. Comparatively, the GA-QSVR and TS-QSVR models also performed well, achieving R2 scores of 0.95346 and 0.95882, respectively. The GA-QSVR model achieved an MAE of 1.56725E-05 and a maximum error of 4.92382E-05, while the TS-QSVR model exhibited an MAE of 1.84075E-05 and a maximum error of 5.02443E-05.

Keywords

• Pharmaceutics
• Nanomedicine
• Support vector regression
• Quadratic kernel
• Hyperparameter tuning