Optimization algorithm for improving the prediction accuracy of API solubility in green solvent.

Area of Science:

• Pharmaceutical Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Supercritical carbon dioxide (SC-CO₂) is a key green solvent in pharmaceutical processing.
• Predicting drug solubility in SC-CO₂ is crucial but challenging due to complex thermodynamic and molecular interactions.
• Accurate solubility data is vital for formulation design, extraction, and process optimization.

Purpose of the Study:

• To develop a hybrid AI framework for predicting active pharmaceutical ingredient (API) solubility in SC-CO₂.
• To enhance hyperparameter tuning and predictive performance using metaheuristic optimization algorithms.
• To provide a robust computational tool for supercritical system analysis.

Main Methods:

• A hybrid AI framework integrating TabNet and histogram-based gradient boosting (HGB).
• Incorporation of metaheuristic optimizers: Attack-Leave Optimizer (ALO), Energy Valley Optimizer (EVO), and Botox Optimization Algorithm (BOA).
• Model evaluation using statistical indicators, cross-validation, prediction interval bootstrapping, and Pareto front analysis.

Main Results:

• The EVO-tuned TabNet model achieved the highest predictive accuracy (R² value).
• The model demonstrated strong generalization capability with narrow prediction intervals.
• Statistical analysis confirmed significant performance differences among optimizers (Kruskal-Wallis test).

Conclusions:

• The proposed hybrid AI framework improves predictive accuracy and interpretability for API solubility in SC-CO₂.
• The computational tool assists in solvent selection and formulation analysis for supercritical systems.
• Future work should expand experimental datasets for broader model generalization.