Computational machine learning analysis and validation for estimation of viscosity of ionic liquids versus...

Area of Science:

• Physical Chemistry
• Materials Science
• Computational Chemistry

Background:

• Ionic liquids (ILs) are versatile solvents with tunable properties.
• Accurate viscosity prediction is crucial for designing and optimizing IL-based processes.
• Existing models may struggle with the complexity of IL structures and diverse operating conditions.

Purpose of the Study:

• To develop and evaluate advanced regression models for predicting the viscosity of ionic liquid systems.
• To compare the performance of Spline Regression (SPR), Twin Support Vector Regression (TSVR), Adaptive Lasso (ALASSO), and Neural Oblivious Decision Ensembles (NODE).
• To identify the most accurate and reliable model for IL viscosity prediction based on a large dataset.

Main Methods:

• A dataset of 8,500 ionic liquid entries was used, including categorical (Cation, Anion) and numerical (Temperature, xIL) features.
• Data preprocessing involved Leave-One-Out encoding, Isolation Forest for outlier removal, and Min-Max normalization.
• Four regression models (SPR, TSVR, ALASSO, NODE) were implemented and hyperparameters optimized using the Firefly Algorithm.

Main Results:

• The NODE model demonstrated superior performance with a cross-validation R² of 0.99536, outperforming SPR (0.96940), TSVR (0.85577), and ALASSO (0.78169).
• NODE achieved the highest test R² (0.99721) and lowest test RMSE (0.0031499) and MAE (0.0022219).
• SPR showed competitive results, followed by TSVR and ALASSO, indicating a clear hierarchy of model effectiveness.

Conclusions:

• Robust data preprocessing is essential for accurate viscosity prediction in ionic liquid systems.
• The Neural Oblivious Decision Ensembles (NODE) model is identified as the most accurate and reliable tool for this task.
• This study provides a validated approach for predicting IL viscosity, aiding in material design and process optimization.