Application of a graph-to-graph translation model to predict new melt-castable molecule

Developing new molecules under multiple constraints is a longstanding challenge, particularly when experimental data is limited and costly to obtain. The design of new melt-cast explosives exemplifies these challenges, and an improved melt-cast explosive would have an immediate impact on demolition, mining, and other applications. Melt-casting is desired because it is a consistent and inexpensive process, yet there are constraints on sensitivity, performance, volatility, and melting point. In this work we iteratively train and apply a graph-to-graph translation machine learning (ML) model to propose new melt-cast explosive molecules that minimize volatility, while also improving performance and constraining on synthesizability, as measured by SA score. To minimize the volatility over a range of temperatures, we maximize the degree of separation between melting point and boiling point, predicted by two new ML property models introduced here. Ultimately, we generate new molecules that have lower predicted volatility and higher oxygen balances than common melt-cast carriers, such as TNT and DNAN, while maintaining low SA scores. We also evaluate the impact of model architecture on melting and boiling point prediction performance and optimize the generative ML model's training epochs, SA score cutoff, starting input dataset, and pairing method. Given accurate property models and sufficient training data, a graph-to-graph translation ML model effectively generates new energetic molecules.