Assessing Graph-based Deep Learning Models for Predicting Flash Point.

Area of Science:

• Computational Chemistry
• Machine Learning
• Chemical Engineering

Background:

• Flash points are critical for assessing flammability hazards of organic molecules.
• Existing measured flash point data is limited, necessitating predictive methods for unmeasured compounds.
• Quantitative structure-property relationship (QSPR) models are commonly used for flash point prediction.

Purpose of the Study:

• To introduce and evaluate graph-based deep learning (GBDL) models for predicting the flash points of organic molecules.
• To compare the performance of GBDL models against traditional QSPR methods.
• To establish the largest dataset to date for flash point prediction using GBDL.

Main Methods:

• Implementation and assessment of two GBDL models: message-passing neural network (MPNN) and graph convolutional neural network (GCNN).
• Comparison of GBDL model performance against 12 established QSPR studies.
• Collection and utilization of the largest available flash point dataset (10,575 molecules).

Main Results:

• The MPNN model demonstrated superior performance over the GCNN model.
• MPNN achieved comparable, though slightly different, performance metrics (R-squared and Mean Absolute Error) compared to previous QSPR studies.
• On the complete dataset, the optimized MPNN achieved an R-squared of 0.803 and MAE of 17.8 K.

Conclusions:

• GBDL, particularly MPNN, represents a viable and powerful alternative to traditional QSPR for flash point prediction.
• The developed models show potential for accurately predicting flash points across diverse molecular classes.
• This work expands the predictive capabilities for chemical flammability hazards through advanced machine learning techniques.