Applying machine learning techniques to predict the properties of energetic materials.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Traditional screening of energetic materials relies on computationally expensive quantum simulations and thermochemical codes.
• Developing accurate predictive models for energetic materials is crucial for efficient discovery.

Purpose of the Study:

• To demonstrate the feasibility of using machine learning (ML) to predict properties of CNOHF energetic molecules based on their structures.
• To compare various ML models and molecular featurization techniques for this prediction task.

Main Methods:

• A diverse dataset of 109 CNOHF energetic molecules was utilized.
• Evaluated featurization methods: sum over bonds, custom descriptors, Coulomb matrices, Bag of Bonds, and fingerprints.
• Compared ML models including kernel ridge regression.

Main Results:

• The 'sum over bonds' (bond counting) featurization and kernel ridge regression model yielded the best performance.
• Achieved acceptable prediction errors and correlations for detonation pressure, velocity, energy, heat of formation, and density.
• Increasing training data size reduced errors, though convergence was slow.

Conclusions:

• Machine learning provides a viable and efficient method for predicting energetic molecule properties.
• This work supports automated lead generation and gaining chemical insights in energetic materials research.