Assisted Energetic Material Property Prediction through Advanced Transfer Learning with Graph Neural Networks

In this study, we explore the use of transfer learning
to predict
the properties of energetic materials using a force-field-inspired
transformer graph neural network (FFiTrNet). We began by pretraining
the model on a large data set of CHNOF compounds and then fine-tuning
it on a smaller data set of experimental enthalpy of formation data
for energetic materials. Our results show that transfer learning significantly
enhances the accuracy of predicting the enthalpy of formation, with
a reduction in mean absolute error and root-mean-square error compared
to direct training on the smaller data set. Furthermore, we demonstrate
the effectiveness of transfer learning in predicting other properties
of energetic materials, highlighting its potential to improve the
predictive capabilities of machine learning models for a range of
energetic materials properties. The result is the most accurate among
the state-of-the-art models for predicting energetic material properties.
The data set used in the fine-tuning enriches the database of energetic
materials’ properties, making this valuable data publicly available
for future research.