Utilization of sequential model-based optimizer integrated machine learning models in correlation of famotidine solubility in supercritical carbon dioxide.

Abstract

We investigated solubility variations of a medication in supercritical carbon dioxide with an insight into preparation of nanomedicines with improved aqueous solubility. As the case study, the solubility of famotidine (FAM) medicine in sc-CO₂ (supercritical carbon dioxide) was computed as a function of temperature and pressure, with a particular focus on modeling and predicting solubility and sc-CO₂ density. Three regression models with machine learning behavior including Quadratic Polynomial Regression (QPR), Weighted Least Squares (WLS), and Orthogonal Matching Pursuit (OMP) were employed to analyze the data, and Sequential Model-Based Optimization (SMBO) was utilized for hyper-parameter tuning. Among these models, the best-performing model for predicting FAM solubility was the QPR model, with an impressive coefficient of determination (R²) of 0.95858 for all sets including training and validation. Additionally, QPR exhibited low MAPE of 1.64278E + 00, RMSE of 9.6833E-02, and a maximum error of 1.49480E-01, while exhibiting a higher maximum error of 18.99 kg/m³ for density predictions, indicating areas for potential improvement. These results highlight the accuracy and precision of the QPR model in predicting FAM solubility in sc-CO₂. For the prediction of sc-CO₂ density, QPR again proved to be the most effective model with a remarkable R² score of 0.99733. This model achieved a low MAPE of 1.06004E-02, RMSE of 8.4072E + 00, and a maximum error of 1.89894E + 01. The QPR model demonstrates its exceptional capability in accurately predicting sc-CO₂ density in terms of temperature and pressure.