The discovery and design of energetic materials is traditionally a rigorous trial-and-error process that requires a detailed characterization of various properties. This presents an extraordinary experimental challenge because of the destructive nature of many of these tests, which drives up costs and time. Modern applications in materials science have implemented data-driven techniques, such as machine learning, to develop predictive tools capable of providing insights prior to synthesis and testing. These efforts seek to accelerate the materials screening process, allowing researchers to spend more time synthesizing and testing promising candidates.

 

In this work, we present a comprehensive framework for the predictive modeling of energetic materials that bridges interpretability and accuracy in data-driven approaches while addressing the challenges imposed by sparse and multimodal datasets. The work is motivated by the need for efficient and reliable performance predictions for energetics, including key properties such as specific impulse, detonation velocity, and detonation pressure, as well as other fundamental properties such as sensitivity and thermal stability. Lastly, we demonstrate the ability to leverage these models within a generative approach to explore property spaces to discover novel materials with optimal properties.