Not from Scratch: Predicting Thermophysical Properties through Model-Based Transfer Learning Using Graph...

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Predicting thermophysical properties is crucial but often limited by experimental data availability and uncertainty.
• Traditional methods like group-contribution and machine learning struggle with data scarcity and generalization.

Purpose of the Study:

• To explore a framework for predicting thermophysical properties using transfer learning from existing estimation models.
• To improve model performance and generalization across diverse chemical spaces, including data-scarce areas.

Main Methods:

• Utilizing a pretraining scheme to leverage knowledge from existing estimation methods.
• Fine-tuning deep learning networks with accurate experimental data.
• Employing graph-based deep learning models for flexible molecular feature generation.

Main Results:

• Improved prediction accuracy for critical properties of compounds, including common organic structures.
• Enhanced model generalization to less explored chemical spaces like inorganics and heavier organic compounds.
• Development of molecular features that can distinguish isomers and demonstrate robustness to data outliers.

Conclusions:

• Transfer learning offers a robust framework for enhancing thermophysical property predictions.
• This approach expands the applicability of predictive models to broader chemical spaces with limited experimental data.
• The method provides insights into structure-property relationships and improves model reliability.