Interpretable Performance Models for Energetic Materials using Parsimonious Neural Networks.

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Predictive models are crucial for explosives and propellant performance, design, and safety.
• Current models lack interpretability, hindering insights into material physics and chemistry.
• First-principles calculations provide fundamental data but are computationally intensive.

Purpose of the Study:

• To develop interpretable models for predicting propellant specific impulse and explosive detonation velocity and pressure.
• To balance model accuracy with simplicity for efficient material screening.
• To gain insights into the underlying physics and chemistry of energetic materials.

Main Methods:

• Utilized Parsimonious Neural Networks (PNNs) to discover interpretable models.
• Employed evolutionary optimization combined with custom neural networks.
• Trained models using objective functions that balance accuracy and complexity.
• Sourced data from the open literature.

Main Results:

• Successfully discovered interpretable models for specific impulse, detonation velocity, and pressure.
• Identified Pareto optimal models balancing accuracy and simplicity for all three properties.
• Demonstrated the effectiveness of PNNs in creating understandable predictive models.

Conclusions:

• Parsimonious Neural Networks offer a viable approach to developing interpretable predictive models for energetic materials.
• The developed models provide a balance between predictive power and interpretability, aiding material design and discovery.
• This work facilitates faster screening and deeper understanding of explosives and propellants.