Predicting Energetics Materials' Crystalline Density from Chemical Structure by Machine Learning.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Predicting molecular bulk properties from chemical structure is crucial for accelerating compound development.
• High explosives (HE) crystalline density prediction is a key challenge in energetic materials research.
• Featurization methods (handcrafted vs. learned representations) impact machine learning model performance.

Purpose of the Study:

• To develop machine learning models for predicting the crystalline density of high explosives (HE) directly from their chemical structures.
• To compare the effectiveness of handcrafted molecular features versus learned representations (e.g., from graph neural networks) for property prediction.
• To enhance the interpretability of complex machine learning models in chemistry.

Main Methods:

• Utilized message passing neural networks (MPNNs) with learned molecular representations.
• Evaluated models using both handcrafted and learned molecular features for predicting crystalline density.
• Curated a dataset of HE-like molecules from the Cambridge Structural Database.
• Performed comparative analysis between MPNNs and models with fixed feature representations.

Main Results:

• MPNN-based models with learned representations achieved state-of-the-art performance in predicting crystalline density.
• Models demonstrated robust performance even on out-of-distribution datasets.
• Analysis provided insights into the features learned by MPNNs for accurate density prediction.

Conclusions:

• Machine learning, particularly MPNNs with learned representations, can effectively predict crystalline densities of high explosives without requiring crystal structure information or quantum mechanical calculations.
• Learned molecular representations offer superior predictive power compared to handcrafted features for this task.
• Interpretable analysis of MPNNs is feasible and valuable for understanding structure-property relationships.