Building Chemical Property Models for Energetic Materials from Small Datasets Using a Transfer Learning Approach

Predicting the enthalpy of formation of energetic molecules via conventional machine learning and GNN

Machine learning (ML) provides a promising method for efficiently and accurately predicting molecular properties. Using ML models to predict the enthalpy of formation of energetic molecules helps in fast screening of potential high-energy molecules, thereby accelerating the design of energetic materials. A persistent challenge is to determine the optimal featurization methods for molecular representation and use an appropriate ML model. Thus, in our study, we evaluate various featurization methods (CDS, ECFP, SOAP, GNF) and ML models (RF, MLP, GCN, MPNN), dividing them into two groups: conventional ML models and GNN models, to predict the enthalpy of formation of potential high-energy molecules. Our results demonstrate that CDS and SOAP have advantages over the ECFP, while the GNFs in GCN and MPNN models perform better. Furthermore, the MPNN model performs best among all models with a root mean square error (RMSE) as low as 8.42 kcal mol-1, surpassing even the best performing CDS-MLP model in conventional ML models. Overall, this study provides a benchmark for ML in predicting enthalpy of formation and emphasizes the tremendous potential of GNN in property prediction.

Force field-inspired transformer network assisted crystal density prediction for energetic materials

Machine learning has great potential in predicting chemical information with greater precision than traditional methods. Graph neural networks (GNNs) have become increasingly popular in recent years, as they can automatically learn the features of the molecule from the graph, significantly reducing the time needed to find and build molecular descriptors. However, the application of machine learning to energetic materials property prediction is still in the initial stage due to insufficient data. In this work, we first curated a dataset of 12,072 compounds containing CHON elements, which are traditionally regarded as main composition elements of energetic materials, from the Cambridge Structural Database, then we implemented a refinement to our force field-inspired neural network (FFiNet), through the adoption of a Transformer encoder, resulting in force field-inspired Transformer network (FFiTrNet). After the improvement, our model outperforms other machine learning-based and GNNs-based models and shows its powerful predictive capabilities especially for high-density materials. Our model also shows its capability in predicting the crystal density of potential energetic materials dataset (i.e. Huang & Massa dataset), which will be helpful in practical high-throughput screening of energetic materials.

Which molecular properties determine the impact sensitivity of an explosive? A machine learning quantitative investigation of nitroaromatic explosives

We decomposed density functional theory charge densities of 53 nitroaromatic molecules into atom-centered electric multipoles using the distributed multipole analysis that provides a detailed picture of the molecular electronic structure. Three electric multipoles, (the charge of the nitro groups), (the total dipole, i.e., polarization, of the nitro groups), (the total electron delocalization of the C ring atoms), and the number of explosophore groups (#NO₂) were selected as features for a comprehensive machine learning (ML) investigation. The target property was the impact sensitivity h₅₀ (cm) values quantified by drop-weight measurements, with a large h₅₀ (e.g., 150 cm) indicating that an explosive is insensitive and vice versa. After a preliminary screening of 42 ML algorithms, four were selected based on the lowest root mean square errors: Extra Trees, Random Forests, Gradient Boosting, and AdaBoost. Compared to experimental data, the predicted h₅₀ values of molecules having very different sensitivities for the four algorithms have differences in the range 19-28%. The most important properties for predicting h₅₀ are the electron delocalization in the ring atoms and the polarization of the nitro groups with averaged weights of 39% and 35%, followed by the charge (16%) and number (10%) of nitro groups. A significant result is how the contribution of these properties to h₅₀ depends on their actual sensitivities: for the most sensitive explosives (h₅₀ up to ∼50 cm), the four properties contribute to reducing h₅₀, and for intermediate ones (∼50 cm ≲ h₅₀ ≲ 100 cm) #NO₂ and contribute to increasing it and the other two properties to reducing it. For highly insensitive explosives (h₅₀ ≳ 200 cm), all four properties essentially contribute to increasing it. These results furnish a consistent molecular basis of the sensitivities of known explosives that also can be used for developing safer new ones.

Prediction and Construction of Energetic Materials Based on Machine Learning Methods

Energetic materials (EMs) are the core materials of weapons and equipment. Achieving precise molecular design and efficient green synthesis of EMs has long been one of the primary concerns of researchers around the world. Traditionally, advanced materials were discovered through a trial-and-error processes, which required long research and development (R&D) cycles and high costs. In recent years, the machine learning (ML) method has matured into a tool that compliments and aids experimental studies for predicting and designing advanced EMs. This paper reviews the critical process of ML methods to discover and predict EMs, including data preparation, feature extraction, model construction, and model performance evaluation. The main ideas and basic steps of applying ML methods are analyzed and outlined. The state-of-the-art research about ML applications in property prediction and inverse material design of EMs is further summarized. Finally, the existing challenges and the strategies for coping with challenges in the further applications of the ML methods are proposed.