Using interpretable machine learning techniques to analyze the thermal decomposition behavior of high-energy compounds for dataset partitioning

In recent years, machine learning has shown great potential in the field of molecular research. However, for energy-containing molecules with small datasets, the difficulty of model prediction is high, and improving model prediction performance is a challenge. This paper aims to analyze the thermal decomposition behavior of energetic compounds using interpretable machine learning techniques, build a dataset containing nearly 900 energetic compound data, and employ interpretable machine learning methods to construct a set of molecular descriptors. After reasonable partitioning of the dataset, directed message passing neural network (D-MPNN) is used for training and performance evaluation. The method in this paper can improve model training effectiveness, with a decrease of approximately 15% in MAE and RMSE, and an increase of about 8% in R-2. This has important value for model training on small-scale datasets.