Machine learning-enhanced morphology and size control of 1,3,5-trinitro-1,3,5-triazinane (RDX) in Nanofiltration-assisted cooling crystallization

The introduction of organic solvent nanofiltration (OSN) technology into the crystallization process of energetic materials (EMs) can enhance solvent utilization and enable precise control of supersaturation during crystallization. However, incorporating OSN technology increases the number of tunable variables in the crystallization process, complicating the optimization of process parameters. Consequently, there is a pressing need to develop an efficient method for optimizing crystallization parameters based on experimental data. Herein, 1,3,5-trinitro-1,3,5-triazinane (RDX) crystals were selected as the research subject, and a method for optimizing experimental operating conditions for nanofiltration-assisted cooling crystallization (NF-CCr) was established using machine learning (ML) techniques. This method includes feature engineering, model prediction interpretation, and verification. In the feature engineering, 85 sets of high-quality experimental datasets were collected, and dimensionality reduction was performed using Principal Component Analysis (PCA) and correlation analysis to enhance the model’s trainability while preserving data variability. Four ML models were trained and optimized: Random Forest (RF), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Back Propagation Neural Network (BPNN). The prediction yield R2 of the BPNN regression model reached 0.89, while the RF binary classification model achieved an accuracy of 91.66% for particle size and 84.72% for sphericity. The SHapley Additive exPlanation (SHAP) method was employed to elucidate the intrinsic correlations between experimental operating conditions and product characteristics. Finally, under the experimental conditions predicted by the RF classification model, RDX crystals with high sphericity (>85%), large particle size (>100 μm), and high purity (>99%) were successfully obtained, achieving an experimental verification accuracy of 80%. This study provides a data-driven optimization solution for the development of complex crystallization processes for EMs.