Application of deep learning and machine learning models with enhanced feature extraction for the prediction of plant extraction yields using supercritical CO2: An optimization and comparative analysis

Abstract

The efficient extraction of essential oils (EOs), particularly volatile compounds, from medicinal, aromatic, or oil-rich crop plants using supercritical carbon dioxide extraction (scCO₂) is crucial for industries such as pharmaceuticals, cosmetics, and food. However, optimizing this process presents challenges due to the intricate molecular diversity of the compounds and the complex interplay of scCO₂ parameters. To address these limitations, this study introduces a hybrid predictive framework that combines deep learning and machine learning, utilizing 694 scCO₂ experimental data points sourced from the literature across 21 plant species. Four major molecular compounds per plant were selected as input features, alongside key process parameters, including temperature, pressure, extraction time, co-solvent ratio, and CO₂ flow rate. Morgan fingerprints were computed for these compounds, and a convolutional neural network (CNN) was utilized to extract their high-level representations into compact vectors. These vectors were integrated with normalized process parameters and fed into a CNN-Multilayer Perceptron (CNN-MLP) hybrid architecture. Performance was compared with Support Vector Regression (SVR), Random Forest (RF), Gaussian Process Regression (GPR), and XGBoost, all optimized using OPTUNA. The CNN-MLP achieved the best performance, with an R² of 0.974 and a Root Mean Squared Error (RMSE) of 1.431 on the test set. A paired t-test (p = 0.810) and Bland–Altman analysis (mean difference: 9.35 %) confirmed the model's robustness. To further assess generalizability, external validations were conducted using unseen experimental conditions. The CNN-MLP was tested on three extraction profiles and demonstrated strong predictive performance, with Pearson correlations ranging from 0.95 to 0.98.