DA-STGAT: A dynamic spatiotemporal graph learning framework featuring a graph sampling mechanism for plutonium aerosol dispersion forecasting in explosion scenarios

Escalating geopolitical tensions have increased risks of chemical explosions and radioactive dispersion, making accurate aerosol prediction critical for emergency response. To address this challenge, this study proposes a Dynamic Adjacency Spatio-Temporal Graph ATtention network (DA-STGAT) model. The model introduces a key innovation: a physics-informed dynamic spatiotemporal graph sampling algorithm that embeds physical priors determined by wind direction factors and inter-station distances into the network structure, enabling adaptive reconstruction of graph topologies throughout the dispersion process. DA-STGAT achieved superior predictive accuracy (R 2 ), outperforming benchmark models (STGAT, STGCN, GAT, GCN, LSTM) by 7.21% to 33.10%, while maintaining low errors across all spatiotemporal scales. In real explosion cases, it achieved exceptional accuracy (R 2 = 0.9954) and was over 7 times faster than a traditional physics-based model. Notably, the model retained R 2 values above 0.83 even under conditions with high proportions of missing data or strong noise interference, demonstrating remarkable robustness. To support this research, we developed the SPADCE simulation program and constructed a radioactive aerosol dispersion database, providing a reliable foundation for data-driven modeling. Experimental results confirm that DA-STGAT exhibits high accuracy, strong robustness, and good interpretability within a physics-informed deep learning framework, offering significant value for emergency decision-making in complex explosive incidents.