Abstract: Accurate prediction of building explosion power field is very important for damage assessment and anti-explosion design of engineering structures. The prediction accuracy of traditional empirical formulas is often limited due to the failure to fully consider the complexity of environmental factors. Numerical simulation can provide more accurate overpressure load parameters, but it is inefficient in dealing with large-scale urban scenes and difficult to meet the needs of rapid damage assessment. In order to solve this problem, this paper innovatively constructs a prediction model of explosion power field based on graph neural network ( GNN ), which aims to directly use the geometric characteristics of buildings to achieve rapid and accurate prediction of three-dimensional physical fields such as explosion peak overpressure, peak impulse and shock wave arrival time on their surfaces. By comparing with the numerical simulation results, the model shows excellent prediction performance : the mean square error of the prediction of the surface overpressure parameters of the single building with different geometric structures is0.97 % ; The average prediction error of the surface overpressure parameters of complex geometric buildings and building communities is 3.17 %. When applied to actual urban areas, the average prediction error is 1.29 % ; the single prediction of the physical field takes no more than 0.6 seconds, which is 3 to 4 orders of magnitude faster than the numerical simulation. The high-precision prediction based on the model can not only reconstruct the overpressure time history curve at any position on the building surface, but also accurately evaluate the damage degree of the structure. The GNN model proposed in this paper provides a new method for rapid and accurate prediction of the explosion power field of urban buildings in explosion scenarios, which can greatly improve the explosion damage assessment and anti-explosion design capabilities of ultra-large-scale complex scene engineering buildings, and has great engineering value.