Ground-motion generations using Multi-label Conditional Embedding–conditional Denoising Diffusion Probabilistic Model (ML–cDDPM)

Abstract

Current studies on ground motion prediction equations require the categorization of earthquake records by regions with different attenuation patterns. Conversely, machine learning based on entire datasets provides an opportunity to explore the factors that dominate ground motion at a site and their complex interactions. In this study, Multi-label Conditional Embedding is proposed to modify a conditional Denoising Diffusion Probabilistic Model (cDDPM) and achieve ground motion prediction of earthquake scenarios. A database was created for neural network training, consisting of 7154 horizontal ground motion records from 105 earthquakes selected from the Next Generation Attenuation (NGA)-West2 database of the Pacific Earthquake Engineering Research Center (PEER). Each record was labeled using four conditional parameters: V_S30, F, M_W, and R_rup. To embed multi-label and location information into the neural network, One-hot Encoding and Positional Encoding techniques were integrated. Hence, a Multi-label Conditional Embedding–conditional Denoising Diffusion Probabilistic Model (ML–cDDPM) was constructed for ground motion prediction. The model was used to simulate ground motions of past earthquakes and was compared with recorded motions. The model was also compared with two other neural-network-based prediction models. The comparisons demonstrate the reliability of ML–cDDPM in simulating ground motions for earthquake scenarios and its superiority over the other two models in representing the complexity of ground motion.