A New Machine-Learning Tool for Fast Estimation of Liquid Viscosity. Application to Cosmetic Oils

Area of Science:

• Physical Chemistry
• Computational Chemistry
• Materials Science

Background:

• Accurate prediction of liquid viscosity is crucial for chemical process design and formulation development.
• Traditional methods for viscosity estimation often rely on empirical correlations or experimental measurements, which can be time-consuming and resource-intensive.
• Advancements in machine learning offer new avenues for developing predictive models based on molecular structure.

Purpose of the Study:

• To estimate the viscosities of pure liquids at 25 °C using three distinct modeling approaches.
• To compare the performance of group contribution, COSMO-RS σ-moment-based neural networks, and graph machine methods.
• To assess the accuracy and applicability of these methods across a diverse range of chemical structures.

Main Methods:

• Employed group contributions, COSMO-RS σ-moment-based neural networks, and graph machines for viscosity prediction.
• Trained machine learning models using a database of 300 molecules at 25 °C.
• Utilized 2D molecular structures (SMILES codes) for group contribution and graph machines, and COSMO-RS σ-moments for neural networks.
• Implemented virtual leave-one-out for efficient graph machine selection.

Main Results:

• Graph machine models, using only 2D structures, provided the most accurate viscosity predictions for an independent set of 22 cosmetic oils.
• The performance of COSMO-RS σ-moment-based neural networks and graph machines was validated.
• A demonstration tool using Docker technology was developed for easy duplication of results and prediction of viscosities for new molecules.

Conclusions:

• Graph machine learning models represent a powerful and accurate approach for predicting liquid viscosities from molecular structures.
• The developed demonstration tool enables accessible viscosity predictions for a wide range of moderate-sized molecules containing C, H, O, or Si.
• This work facilitates the design and optimization of chemical processes and formulations by providing reliable viscosity estimation tools.