When high accuracy misleads in literature-derived machine learning for deep eutectic solvent recommendation

Area of Science:

• Analytical Chemistry
• Computational Chemistry

Background:

• Deep eutectic solvents (DESs) are versatile for analytical sample preparation.
• Selecting optimal DES systems is challenging due to vast design space and empirical methods.
• Machine learning (ML) offers a data-driven approach, but its efficacy with literature data is uncertain.

Purpose of the Study:

• To evaluate the reliability of ML models for predicting DES performance using literature-derived data.
• To assess if improved predictive accuracy translates to practical recommendation capabilities.
• To identify data quality and structure limitations impacting ML model performance in DES selection.

Main Methods:

• Reconstructed a literature-derived dataset of 757 DES-based pesticide extraction experiments from 94 studies.
• Implemented a leakage-controlled, DOI-grouped validation framework.
• Developed and compared a hybrid ML model (classification and preference learning) against a baseline.

Main Results:

• The hybrid ML model improved record-level classification but did not yield stable group-level ranking benefits.
• Performance degraded in restricted comparable groups, with the hybrid model often underperforming.
• Baseline models captured significant structure, but the hybrid model offered no consistent additional advantage.

Conclusions:

• Predictive accuracy gains in ML models for DES extraction do not automatically ensure reliable chemical recommendations from literature data.
• Sparse comparative data, study-specific targets, and limited descriptors hinder ML model utility.
• Future ML applications require data-centric evaluation, standardized reporting of comparable candidates, and structured metadata.