An End-to-End GNN for Particle Size Analysis Using Residual Enhancement and Attention-Based Segmentation

Particle size analysis is critical in fields like materials science, pharmaceuticals, and mining, but traditional methods often suffer from inefficiency and reliance on labeled data. This paper introduces a deep learning framework that integrates self-supervised learning, physics-based modeling, and graph neural networks (GNNs) to improve automation, accuracy, and generalization. Self-supervised learning reduces dependence on labeled data, while physics-based constraints ensure the model's predictions align with physical phenomena. GNNs capture interactions between particles, enhancing precision, particularly for overlapping or densely packed particles. Experimental results demonstrate that this framework outperforms traditional methods, providing higher accuracy and robustness across datasets. Specifically, the proposed framework achieves a mean absolute error (MAE) of 18.7 m, a root mean square error (RMSE) of 29.1 m, and an R2 value of 0.92. An ablation study further shows the importance of self-supervised learning and physics-based constraints in improving the model's performance.