The effect of descriptor choice in machine learning models for ionic liquid melting point prediction

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Predicting ionic liquid properties from structure is a key computational chemistry challenge.
• Existing methods like group contribution have limitations in scope.
• Experimental data on ionic liquid melting points is extensive.

Purpose of the Study:

• To develop accurate machine learning models for predicting ionic liquid melting points.
• To investigate the impact of incorporating first-principles quantum mechanical data into structural descriptors.
• To create a universally applicable model for diverse ionic liquid types.

Main Methods:

• Kernel ridge regression models were trained on a dataset of 2212 ionic liquid melting points.
• Structural descriptors (ECFP4 fingerprints, Coulomb matrix) were augmented with quantum mechanical data (molecular orbital energy, charge density, interaction energy).
• Model performance was evaluated based on mean absolute error (MAE).

Main Results:

• Augmenting structural descriptors with quantum mechanical data improved melting point prediction accuracy.
• The best model, using ECFP4 fingerprints and molecular orbital energies, achieved an MAE of 29 K.
• This model demonstrated applicability across diverse ionic liquid ion types, outperforming traditional methods.

Conclusions:

• Combining structural descriptors with first-principles quantum mechanical data is effective for predicting ionic liquid melting points.
• The developed ECFP4-based model offers a highly accurate and broadly applicable approach.
• This method advances computational chemistry's ability to characterize ionic liquid properties.