Advancing material property prediction: using physics-informed machine learning models for viscosity

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning in Materials

Background:

• Physics-based models struggle with accurate computation of material properties like viscosity.
• Data-driven machine learning (ML) models face challenges in materials science due to limited data availability.
• Accurate prediction of viscosity is crucial for understanding liquid systems and material behavior.

Purpose of the Study:

• To enhance the accuracy and interpretability of ML models for predicting material properties.
• To integrate physics-informed descriptors from molecular dynamics (MD) simulations into quantitative structure-property relationship (QSPR) models.
• To accurately predict temperature-dependent viscosities of small organic molecules using QSPR models.

Main Methods:

• Curated a dataset of over 4000 small organic molecule viscosities from scientific literature and databases.
• Developed descriptor-based and graph neural network QSPR models incorporating MD simulation descriptors.
• Utilized feature importance tools to identify key predictive descriptors.

Main Results:

• Incorporating MD descriptors significantly improved viscosity prediction accuracy, particularly for datasets with fewer than 1000 data points.
• Intermolecular interactions, captured by MD descriptors, were identified as the most critical features for viscosity prediction.
• The developed QSPR models accurately predicted the inverse relationship between viscosity and temperature for six battery-relevant solvents, including unseen ones.

Conclusions:

• Integrating MD descriptors into QSPR models is an effective strategy for improving prediction accuracy of challenging material properties.
• This hybrid approach overcomes limitations of physics-based models and data scarcity in machine learning for materials science.
• The study demonstrates the utility of MD-enhanced QSPR for predicting temperature-dependent viscosity in liquid systems.