Force field-inspired transformer network assisted crystal density prediction for energetic materials

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Engineering

Background:

• Machine learning, particularly Graph Neural Networks (GNNs), shows promise for predicting chemical properties.
• Predicting properties of energetic materials using machine learning is nascent due to limited data.
• Traditional methods for molecular descriptor generation are time-consuming.

Purpose of the Study:

• To develop an improved machine learning model for predicting energetic materials properties.
• To enhance existing force field-inspired neural networks (FFiNets) with Transformer architecture.
• To address the challenge of insufficient data in energetic materials research.

Main Methods:

• Curated a dataset of 12,072 CHON compounds from the Cambridge Structural Database.
• Implemented a Transformer encoder into a force field-inspired neural network (FFiNet) to create FFiTrNet.
• Evaluated FFiTrNet's performance against other machine learning and GNN-based models.

Main Results:

• The developed force field-inspired Transformer network (FFiTrNet) demonstrated superior predictive performance.
• FFiTrNet showed strong predictive capabilities, especially for high-density energetic materials.
• The model successfully predicted the crystal density for a potential energetic materials dataset.

Conclusions:

• FFiTrNet offers powerful predictive capabilities for energetic materials properties.
• The model's accuracy is particularly notable for high-density materials.
• This approach can facilitate high-throughput screening of novel energetic materials.