A property-oriented adaptive design framework for rapid discovery of energetic molecules based on small-scale labeled datasets.

It remains an important challenge to apply machine learning in material discovery with limited-scale datasets available, in particular for the energetic materials. Motivated by the challenge, we developed a Property-oriented Adaptive Design Framework (PADF) to quickly design new energetic compounds with desired properties. The PADF consists of a search space, machine learning model, optimization algorithm and an evaluator based on quantum mechanical calculations. The effectiveness and generality of the PADF were assessed by two case studies on the heat of formation and heat of explosion as the target properties. 88 compounds were selected as the initial training dataset from the search space containing 84 083 compounds generated. SVR.lin/Trade-off coupled with E-state + SOB and KRR/KG coupled with CDS + E-state + SOB were determined to be the best combination pairs for the heat of formation and the heat of explosion, respectively. Most of the ten compounds selected from the first ten iterations exhibit better properties than the optimal sample in the initial dataset. Besides, the heat of explosion as the target property outperforms the heat of formation in designing energetic compounds with high detonation performance. In particular, a new compound selected at the 3rd iteration exhibits high potential as an explosive. Our strategy could be extended to other domains limited by small-scale datasets labeled.