Accelerating the Design of New Natural Deep Eutectic Solvents with Supervised Variational Autoencoders
- 1Unité de Catalyse et de Chimie du Solide (UCCS
Abstract
The rational design of natural deep eutectic solvents (NaDESs) is hindered by the vast combinatorial space arising from the possible combinations of hydrogen bond donors and acceptors. This enormous design space limits the systematic experimental exploration of their physicochemical properties. In this work, we propose a supervised variational autoencoder (SVAE) framework to efficiently explore this chemical space by coupling generative modeling with property prediction. The model is trained on a curated dataset of deep eutectic solvents to learn simultaneously the latent structural representations and correlations with target properties, including melting point, density, and viscosity. The SVAE generates virtual NaDES candidates characterized by molecular descriptors, component ratios, and target property values. To associate these virtual candidates with real molecules, a descriptor matching strategy is applied using the COCONUT natural product database. The resulting NaDES candidates are subsequently validated using predictive models trained on experimental DES data. The results show strong agreement between the SVAE-generated target property values and the corresponding predictions of the regression models (R2 = 0.89, RMSE=21.7 K for melting point; R2 = 0.91, RMSE = 0.04 g·cm-3 for density; and R2 = 0.89, RMSE=0.47 cP for ln(viscosity)). This workflow demonstrates how guided generative models can be combined with machine learning models to accelerate the discovery of sustainable solvents. Beyond NaDESs, the proposed strategy provides a general and transferable framework for the rational design of other chemical systems, where combinatorial complexity limits conventional experimental approaches.
Related
- Predicting fluorinated gas solubility in deep eutectic solvents using machine learning and deep learning approaches
- O2-tolerant CO2-to-CO reduction by air-sensitive carbon monoxide dehydrogenases in hydrated deep eutectic solvents
- Sensing Viscosity Decoupling and Dynamic Heterogeneity in Tetrabutylammonium Bromide:Polyol Deep Eutectic Solvents.
- Thermodynamics-Informed Learning Captures Hierarchical Hydrogen-Bond Networks in Deep Eutectic Solvents
- Unveiling the Molecular Driving Forces of Pollutant Extraction by Hydrophobic Eutectic Solvents.
BibTeX
@misc{laref2026accelerating,
title = {{Accelerating the Design of New Natural Deep Eutectic Solvents with Supervised Variational Autoencoders}},
author = {Rachid Laref and Mehdi Oubahmane and Nicolas Rolland and Véronique Nardello-Rataj and Adlane Sayede},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008532/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008532/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Accelerating the Design of New Natural Deep Eutectic Solvents with Supervised Variational Autoencoders
AU - Rachid Laref
AU - Mehdi Oubahmane
AU - Nicolas Rolland
AU - Véronique Nardello-Rataj
AU - Adlane Sayede
PY - 2026
DA - 2026/09/09
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008532/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008532/v1
AB - The rational design of natural deep eutectic solvents (NaDESs) is hindered by the vast combinatorial space arising from the possible combinations of hydrogen bond donors and acceptors. This enormous design space limits the systematic experimental exploration of their physicochemical properties. In this work, we propose a supervised variational autoencoder (SVAE) framework to efficiently explore this chemical space by coupling generative modeling with property prediction. The model is trained on a curated dataset of deep eutectic solvents to learn simultaneously the latent structural representations and correlations with target properties, including melting point, density, and viscosity. The SVAE generates virtual NaDES candidates characterized by molecular descriptors, component ratios, and target property values. To associate these virtual candidates with real molecules, a descriptor matching strategy is applied using the COCONUT natural product database. The resulting NaDES candidates are subsequently validated using predictive models trained on experimental DES data. The results show strong agreement between the SVAE-generated target property values and the corresponding predictions of the regression models (R2 = 0.89, RMSE=21.7 K for melting point; R2 = 0.91, RMSE = 0.04 g·cm-3 for density; and R2 = 0.89, RMSE=0.47 cP for ln(viscosity)). This workflow demonstrates how guided generative models can be combined with machine learning models to accelerate the discovery of sustainable solvents. Beyond NaDESs, the proposed strategy provides a general and transferable framework for the rational design of other chemical systems, where combinatorial complexity limits conventional experimental approaches.
ER -