17:15 〜 19:15
[SCG49-P07] Joint Simulation of Ground and Building Motion Reconstruction based on Diffusion Model
キーワード：diffusion models, data reconstruction, multi-physics
The simulation of underground earthquake wave propagation and building motion is significant for disaster prediction and urban disaster prevention. Traditional physics-based modeling methods are often computationally expensive and require expertise from multiple disciplines. These methods, which use partial differential equations to predict ground and building motion caused by disasters like earthquakes, tornadoes, and landslides, can be complex to implement. Traditional methods typically simulate ground and building motion separately to reduce computational costs. This separation limits their ability to account for ground and building motion interactions and restricts the ability to make comprehensive disaster scenario predictions.
To address this issue, we propose an innovative method based on Diffusion Models, aiming to simultaneously simulate both the ground and building motion patterns. The method takes multiple physical models as inputs, including the location and strength of the earthquake source, the building structural model, and the medium distribution. Using different neural encoders, we integrate these physical systems and utilize Latent Diffusion Models to generate the latent space representation of the target physical processes, finally outputting the simulated ground and building dynamic responses through neural decoders.
We selected the Beijing Urban Cluster as the test model and trained the model using datasets obtained from numerical simulation programs, verifying its effectiveness. Compared to traditional separate calculation methods, the joint simulation based on the Diffusion Model improves the simulation's speed and accuracy. It provides new technical tools for digital twin disaster mitigation and emergency management, enabling more comprehensive and efficient disaster prediction.
To address this issue, we propose an innovative method based on Diffusion Models, aiming to simultaneously simulate both the ground and building motion patterns. The method takes multiple physical models as inputs, including the location and strength of the earthquake source, the building structural model, and the medium distribution. Using different neural encoders, we integrate these physical systems and utilize Latent Diffusion Models to generate the latent space representation of the target physical processes, finally outputting the simulated ground and building dynamic responses through neural decoders.
We selected the Beijing Urban Cluster as the test model and trained the model using datasets obtained from numerical simulation programs, verifying its effectiveness. Compared to traditional separate calculation methods, the joint simulation based on the Diffusion Model improves the simulation's speed and accuracy. It provides new technical tools for digital twin disaster mitigation and emergency management, enabling more comprehensive and efficient disaster prediction.