Active Learning for Generalizable Detonation Performance Prediction of Energetic Materials

Area of Science:

• Computational chemistry
• Materials science
• Chemical engineering

Background:

• Developing new energetic materials is vital for technological advancement but faces challenges in experimental speed and cost.
• Computational methods for predicting material properties are often hindered by the high cost of obtaining accurate input data.
• Predicting detonation performance across a wide chemical space requires efficient and scalable computational strategies.

Purpose of the Study:

• To develop an efficient computational workflow for predicting the detonation performance of energetic materials.
• To create the largest publicly available database of potential CHNO explosives.
• To identify key molecular features that govern the detonation performance of energetic materials.

Main Methods:

• Integration of density functional theory (DFT) calculations, thermochemical modeling, message-passing neural networks, and Bayesian optimization.
• Implementation of an active learning strategy for iterative data set expansion and targeted molecule selection.
• High-throughput screening of over 70 billion candidate molecules.

Main Results:

• Generation of the largest publicly available database of potential CHNO explosives.
• Development of a generalizable surrogate model with high accuracy (R² > 0.98) for predicting detonation performance.
• Identification of oxygen balance as the primary driver of detonation performance, supported by electronic structure, density, and functional groups.

Conclusions:

• The active learning workflow significantly accelerates the discovery of energetic materials.
• The developed surrogate model and database provide a robust foundation for future high-throughput screening and targeted synthesis.
• Cheminformatics analysis reveals clustering of materials with similar performance, guiding future research directions.