A deep learning based model for aluminum agglomeration in solid propellant

• A GCN-attention model for aluminum agglomeration prediction is firstly proposed • The model greatly improves the computational efficiency by inputing local information • The model can adapt to different agglomeration criteria of aluminum through GCN • The error of size distribution between prediction and experimental data is less than 5% This study develops a GCN-attention neural network to predict agglomerate distribution in solid propellants. The prediction engine was trained on 25,000 datasets generated from Random Packing Algorithms and Random Packing Agglomeration Model. It uses local particle information and adjacency matrices to learn agglomeration patterns under different agglomeration criteria. The prediction engine demonstrates high accuracy under appropriate parameters, with errors in key metrics—volume mean diameter ( D 4,3 ), median particle diameter ( D 50 ), and diameter at 90% cumulative volume ( D 90 )—all below 5%. It effectively learns from large datasets, mitigating the impact of noisy data while capturing underlying patterns. The GCN component enables the prediction engine to adapt to different agglomeration criteria by processing particle connectivity via adjacency matrices, without requiring retraining for new judgment rules. Furthermore, the deep learning model significantly outperforms the Random Packing Agglomeration Model in computational efficiency. Using only 523 aluminum particles as input, it achieves higher accuracy than the traditional model with 18,000 particles, while reducing computation time by a factor of 485.6 and requiring substantially less memory. This work highlights the potential of GCN-attention networks in enhancing propellant performance prediction and optimizing formulation design.