Optimization of drug solubility inside the supercritical CO₂ system via numerical simulation based on artificial...

Area of Science:

• Chemical Engineering
• Computational Chemistry
• Machine Learning Applications

Background:

• Supercritical carbon dioxide (SC-CO2) is a promising solvent for pharmaceutical processes.
• Accurate prediction of SC-CO2 density and drug solubility is crucial for process design.
• Machine learning offers a powerful approach for modeling complex physical properties.

Purpose of the Study:

• To evaluate Polynomial Regression (PR), Extreme Gradient Boosting (XGB), and LASSO models for predicting SC-CO2 density.
• To assess the performance of these models in estimating niflumic acid solubility in SC-CO2.
• To optimize model hyperparameters using the Barnacles Mating Optimizer (BMO).

Main Methods:

• Utilized PR, XGB, and LASSO regression models.
• Employed the Barnacles Mating Optimizer (BMO) for hyperparameter tuning.
• Validated model performance using R-squared values for density and solubility predictions.

Main Results:

• PR achieved the highest accuracy for SC-CO2 density (R²=0.99207) and niflumic acid solubility (R²=0.96949).
• XGB demonstrated strong predictive performance for both density (R²=0.92673) and solubility (R²=0.92961).
• LASSO provided good predictive capabilities, with R² values of 0.81917 for density and 0.82094 for solubility.

Conclusions:

• Machine learning models, particularly PR and XGB, show high accuracy in predicting SC-CO2 density and niflumic acid solubility.
• The BMO algorithm is effective for optimizing machine learning models in this context.
• These findings support the application of machine learning for solvent-solute property estimation in the pharmaceutical industry.