AI-driven optimization of sustainable solvent-based extraction: A deep learning approach for green sample preparation

Abstract

Background

This study introduces a data-driven deep learning framework for optimizing the extraction of plant biomolecules, aiming to improve both efficiency and sustainability in analytical sample preparation. Conventional extraction methods such as maceration, Soxhlet, reflux, and liquid–liquid partition, often paired with solvents of varying profiles are frequently resource-intensive and environmentally demanding. To address these limitations, we developed the Deep Green Optimization (DeepGO) framework, a hybrid CNN-LSTM deep learning model trained on more than five thousand experimental data points that includes sustainability indicators such as solvent properties, energy demand, and toxicity.

Results

The DeepGO framework predicts multiple extraction outcomes including extraction yields, phenolic, flavonoid, and tannin contents, as well as antioxidant and metal chelating activities. Decision-making algorithms revealed that moderately toxic and polar solvents, particularly n-butanol and ethyl acetate, generated the highest extraction efficiencies (R² reached 0.87). We demonstrate that hexane and water, as the two solvents of extreme polarity, showed limited compatibility with these conventional methods (0.44<R²<0.65).

Significance

Based on environmental impact and energy consumption, the DeepGO framework provides a valuable predictive tool to enhance plant-based extractions. This approach substantially reduces solvent consumption, lowers energy demand, and minimizes waste generation, which are key aspects of sustainable chemical practices.