Advanced analysis on the correlation of salicylic acid solubility to solvent composition, temperature and pressure via machine learning approach.

Abstract

This work aims to use powerful machine learning methods to predict salicylic acid solubility in various solvents as function of pressure and temperature. Using a dataset consisting of 217 data points and 15 input features, the analysis was performed using variables including pressure, temperature, and 13 different solvents as integral aspects. The considered solvents for this study included: ethanol, water, methanol, ethyl acetate, PEG 300, 1,4-dioxane, 1-propanol, 1-butanol, 1-pentanol, 1-hexanol, 1-heptanol, acetonitrile, and acetone. Temperature between 243.15 and 323.15 K, and pressure between 90 and 101.32 kPa were used in the models. The study commenced with a comprehensive data pre-processing phase, which involved normalizing the data using a Min-Max Scaler. This was followed by the removal of outliers using the k-Nearest Neighbors Outlier Detection (KNNOD) technique. Several models, including Convolutional Neural Networks (CNNs), Polynomial Regression (PR), and Kernel Ridge Regression (KRR), were employed to predict the solubility of salicylic acid. The Hyperband method was utilized for hyper-parameter optimization, ensuring optimal performance for each model by dynamically allocating computational resources. The effectiveness of these models was evaluated using metrics such as R² scores, MSE, and MAE. The results revealed that CNNs outperformed the other models with a high degree of accuracy (R² score of 0.989, MSE of 4.161203E-05, and MAE of 3.760119 E-03), while KRR achieved an R² score of 0.913873. The results of the study underline the robustness of preprocessing methods, model selection, and hyper-parameter tuning for the attainment of accurate predictions, making useful contributions to the area of solubility prediction by salicylic acid in various solvent environments.