Predictive and Generative AI Approaches to the Design of Deep Eutectic Solvents

This is the repository associated to the review by Luna Liviero, Philipp Leclercq, Erwan Privat, Sergio Rampino and Antonino Polimeno under review in the Sustainability and Circularity NOW journal.

The motivation and context for this work is described in the aforementioned article. We focus here only on the illustrative example results of accuracy improvement by combining the datasets from Odegova et al.¹ and Luu et al.² Available code here is based on the first¹ team code. Repository links are in the bibliography.

Datasets

More information on the datasets can be found on the main review text. To summarize, the 2024 publication by Odegova et al.¹ compiled a dataset of DES properties from experimental studies dating back to 2003, resulting in 2,303 data entries for melting temperature, 4,369 for density, and 4,216 for viscosity. In 2023, Luu et al.² collected a dataset with 402 entries from literature, reporting the melting temperature as well as the component percentages for each solvent, and successfully developed a prompt-based transformer model for both predictive and generative tasks, using SMILES to represent the molecules. They trained the model jointly on DES tasks and tasks related to the well-established and much larger QM9³ dataset.

In this repository

• data/DES_Melting.csv: Odegova et al.¹ dataset,
• data/DES1.csv: Luu et al.² dataset,
• data/DES_TMELT.csv: our combined dataset,
• data/DES_TMELT.ipynb: code used for the merging.

Our results

For illustrative purposes, and building on the data available from recent works on DESs, computational analysis was conducted leading to insightful exploration in regard to two aspects: the impact of dataset size and the contribution of feature selection to model prediction performance. The datasets compiled by Luu et al.² and Odegova et al.¹, along with the latter’s codebase, served as the foundation for further investigation. The first step involved compiling a unified dataset by merging the existing data frames. Luu et al.² provides a file with 401 entries, while Odegova at al.¹ contributed 2,259 entries. The data is homogenized through unit conversions where necessary and by retaining only the common columns. Eight redundant entries were identified and handled differentially. For seven of these, both data pairs were retained, as they showed slight variations in component melting temperatures—considered acceptable within a ±10 K margin. These duplicates were grouped by averaging the melting temperature values to preserve as much information as possible with some curating. After merging, the resulting dataset contained 2,651 entries. Each DES entry includes, where available: the SMILES formula of both components, the melting temperature of each component, molar ratios, the melting temperature of the resulting DES, the full names of both components, the DES type, and the reference/DOI of the source study. From this dataset, the final feature set and label were extracted for model training and testing. Using the codebase from Odegova et al.¹, molecular descriptors were generated from SMILES formulas using the RDKit library.⁴ Feature selection followed the methodology in the reference article, where over 200 descriptors were filtered using correlation matrices to exclude highly correlated variables. The final features included singular melting temperature, molar fractions, molecular weight, hydrogen bond donor count, various functional group counts, and toxicity. The target label was the melting temperature of the DES. The models are trained on 80 % of entries and tested on 20 % of them, using 5-fold cross-validation. The “mixtures-out”⁵ method was applied to prevent data leakage by grouping entries containing the same molecule into the same subset. Hyperparameter optimization was performed for each model. Evaluation metrics included the coefficient of determination ($R^2$) and root mean square error (RMSE), calculated for both training and testing sets. The evaluation encompasses training and testing of all models on three different variation of the datasets, including the original two datasets. The addition of more input data by curating and combining datasets show an overall improvement for all models. We can note that gradient-boost-like models results were always improved by adding more data to the base dataset. We also note that extending the dataset also improved results while not showing sign of overfitting for most of the tested models. This can be explained in part by the imbalance of the data, since nearly half of the entries belong to Type III DES and only 0.47 % to Type II. This discrepancy is related to industrial relevance of Type III DES and available experimental data.⁶

The results indicate that the relevance of the "type of DES" feature varies across models. The kernel based models, SVM, KNN and MLP, were the most sensitive to this feature removal with a notable drop in performance. In contrast, having more samples always leads to improved performance across all models. Also in this case, the kernel models are deeply affected in their performance. This highlights the importance of quality, structured, abundant data in the field of machine learning to enhance model accuracy. One limitation of this analysis lies in the uneven distribution of DES classification. Although the additional entries lacked class information, it is assumed that their inclusion did not attenuate the imbalance, leaving type III and V disproportionately represented. It cannot be excluded  that the seemingly improved model performance is driven primarily by the overrepresented classes, while the model may struggle to capture relationships within underrepresented DES types. Despite these concerns, the model's performance remains significant, with three models exceeding an $R^2$ of 0.80. The top performing model remains Cat Boosting Regression, even after removal of a feature and with an bigger dataset. In all conditions, it yielded the highest average $R^2$ on the test subset and as this value rose sharply, the corresponding training increased mildly. This suggests that the model performance improved on unseen data points and has therefore good generalization capabilities. SVM and KNN also demonstrated notable potential, and may outperform CBR in future applications if the DES type feature is reintroduced into the augmented dataset.

In the table below is a recap of the cross-validation metrics $R^2$ and RMSE for original and extended dataset for several models. These results are from targeting melting temperature and RMSEs are in Kelvin.

Model	Original $R^2$	Extended $R^2$	Original RMSE	Extended RMSE
DTR	0.583633	0.652694	47.3243	47.7102
RFR	0.713469	0.785704	39.3110	37.2780
GBR	0.733936	0.788667	37.9431	36.6524
CBR	0.757018	0.823381	36.2145	33.8333
XGB	0.733484	0.803071	38.0284	35.6826
SVM	0.629013	0.812928	43.7242	34.8591
KNN	0.642345	0.805520	43.0502	35.3635
MLP	0.469012	0.771632	52.9626	38.4747

Regression models:

• DTR: decision tree
• RFR: random forest
• GBR: gradient boosting
• CBR: cat boosting
• XGB: extreme boosting
• SVM: support vector machine
• KNN: K-nearest neighbors
• MLP: multilayer perceptron

In this repository

The source can be found in the src directory. Please note that most of the code is an adaptation of the already existing one made by Odegova at al.¹, link to repository in bibliography.

Footnotes

1. Odegova, V.; Lavrinenko, A.; Rakhmanov, T.; Sysuev, G.; Dmitrenko, A.; Vinogradov, V. Green Chem., 2024, 26, 3958. https://github.com/lamm-mit/MoleculeDiffusionTransformer ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

2. Luu, R.K.; Wysokowski, M.; Buehler, M.J. Appl. Phys. Lett. 2023, 122, 234103. https://github.com/lamm-mit/MoleculeDiffusionTransformer ↩ ↩² ↩³ ↩⁴ ↩⁵

3. Junde Li, Swaroop Ghosh (2024). Dataset: QM9. https://doi.org/10.57702/exp0m45r ↩

4. RDKit: Open-source cheminformatics. https://www.rdkit.org ↩

5. Muratov, E. N., Varlamova, E. V., Artemenko, A. G., Polishchuk, P. G., & Kuz'min, V. E. (2012). Existing and developing approaches for QSAR analysis of mixtures. Molecular informatics, 31(3‐4), 202-221. ↩

6. Petteri Vainikka, Sebastian Thallmair, Paulo Cesar Telles Souza, and Siewert J. Marrink, ACS Sustainable Chemistry & Engineering 2021 9 (51), 17338-17350, DOI: 10.1021/acssuschemeng.1c06521 ↩