Graph convolutional neural network applied to the prediction of normal boiling point.

Area of Science:

• Computational chemistry
• Machine learning in chemical research

Background:

• Accurate prediction of physical properties like normal boiling points is crucial for chemical engineering and research.
• Existing methods for boiling point prediction can be limited in scope or require extensive input data.

Purpose of the Study:

• To develop and validate a machine learning model for predicting normal boiling points of organic compounds.
• To leverage graph neural networks for property prediction from molecular structures.

Main Methods:

• Utilized a graph neural network (GNN) architecture for the machine learning model.
• Extracted molecular features directly from 2D chemical structure sketches.
• Trained and validated the model using experimental data from the NIST TRC SOURCE Data Archival System.

Main Results:

• The final GNN model achieved high accuracy, predicting normal boiling points within 6 K (1.32% MAE).
• The model demonstrated a low sample standard deviation of less than 8 K.
• The model effectively identified erroneous data points within the training dataset.

Conclusions:

• Graph neural networks offer a powerful approach for predicting physical properties of organic compounds.
• The developed model provides a rapid and reliable method for normal boiling point prediction.
• The model's ability to detect data errors highlights its utility in data curation and validation efforts.