Machine-Learning-Based Prediction of the Glass Transition Temperature of Organic Compounds Using Experimental Data

Area of Science:

• Atmospheric chemistry
• Materials science
• Computational chemistry

Background:

• The glass transition temperature (Tg) of organic compounds in atmospheric aerosols is crucial for understanding particle viscosity, reaction rates, and phase behavior.
• Experimental Tg data is scarce for the vast diversity of organic molecules found in aerosols.
• Accurate Tg prediction is needed to model aerosol processes effectively.

Purpose of the Study:

• To develop a machine learning model for predicting the glass transition temperature of organic compounds relevant to atmospheric aerosols.
• To compare different machine learning approaches for Tg prediction using molecular descriptors.
• To provide accessible tools and data for researchers in atmospheric and material sciences.

Main Methods:

• Utilized the extremely randomized trees (extra trees) machine learning algorithm.
• Developed two prediction approaches: one based on functional groups and another using SMILES string-derived descriptors.
• Incorporated melting temperature as an additional input variable for improved accuracy.

Main Results:

• Both functional group and SMILES-based models achieved a mean absolute error of approximately 12-13 K.
• SMILES-based predictions showed slightly superior performance.
• The developed model outperformed previous parameterizations and existing machine learning models for Tg prediction.

Conclusions:

• The machine learning model demonstrates robust predictive power for the glass transition temperature of diverse organic aerosol compounds.
• The study provides a user-friendly web interface, Python code, and the Bielefeld Molecular Organic Glasses (BIMOG) database for wider application.
• This tool is valuable for advancing research in atmospheric aerosol science and material science.