A new study published in Engineering presents an innovative computational method for predicting the distribution of blast loads on complex 3D buildings using graph neural networks. The research, led by a team from the School of Mechanical Engineering at Nanjing University of Science and Technology, introduces a data-driven approach that significantly enhances the accuracy and efficiency of predicting blast loads, which is crucial for structural design, civil defense, and safety assessments.

Traditional methods for predicting blast loads, such as experimental research, theoretical models, and numerical simulations, face several limitations. Experimental setups are costly and difficult to implement, while theoretical models are often restricted to simple geometries and lack precision. Numerical simulations, though accurate, require substantial computational resources and time, making them impractical for real-time applications. To address these challenges, the researchers developed an encoder–decoder graph neural network (GNN) model named BlastGraphNet.

BlastGraphNet leverages a message-passing mechanism to predict overpressure and impulse load distributions on buildings with both conventional and complex geometries. The model is trained using a dataset generated from numerical simulations of explosions in a computational domain containing various building configurations. By focusing on key parameters such as peak overpressure, peak impulse, and wavefront arrival time, BlastGraphNet simplifies the prediction process while maintaining high accuracy.

The results of the study demonstrate that BlastGraphNet achieves a prediction error of less than 2% for conventional building tests, with an inference speed that is 3 to 4 orders of magnitude faster than state-of-the-art numerical methods. In more complex scenarios, such as buildings with intricate geometries and building clusters, the model exhibits high accuracy and excellent generalizability. The study also highlights the model’s potential for downstream applications, including structural damage assessment and virtual city explosion simulations.

The researchers constructed a comprehensive dataset by generating 4,000 samples of diverse building configurations for training. To evaluate the model’s geometric adaptability, three types of test datasets were prepared, featuring rotational geometries, multibody geometries, and complex geometries. The model’s performance was assessed using metrics such as the relative mean squared error (RMSE) and the coefficient of determination (R²). The results showed that BlastGraphNet effectively captures the fine-grained features of shock wave propagation and maintains consistent predictive capabilities across different attributes.

In addition to its high accuracy and efficiency, BlastGraphNet demonstrates strong robustness and generalizability. The model was tested on various complex building geometries and multibody scenarios, achieving satisfactory prediction performance. The study also explored the model’s potential for engineering applications, such as structural damage assessment using pressure–impulse (P–I) curves and virtual city explosion simulations. The results indicate that BlastGraphNet can provide critical support for building design, urban planning, and emergency management by enabling rapid and accurate assessments of explosion risks.

The research highlights the potential of graph neural networks for handling unstructured data and complex physical phenomena, offering a new paradigm for predicting blast loads and assessing damage in protective engineering. Future work may focus on incorporating more diverse explosion scenarios and optimizing the model for even larger-scale applications, further enhancing its capabilities for urban safety assessments.

The paper “BlastGraphNet: An Intelligent Computational Method for the Precise and Rapid Prediction of Blast Loads on Complex 3D Buildings Using Graph Neural Networks,” is authored by Zhiqiao Wang, Jiangzhou Peng, Jie Hu, Mingchuan Wang, Xiaoli Rong, Leixiang Bian, Mingyang Wang, Yong He, Weitao Wu. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.03.007. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.