Training a continuous conditional diffusion-based model to simulate horizontal strike-slip ground motions for scenario seismic events

Parametric models of the time–frequency-dependent power spectral density (TFPSD) function of seismic ground motion records can be developed from historical records, though model selection involves some subjectivity. The TFPSD function varies complexly in time–frequency domain, with the duration and spectral content depending on the moment magnitude M , rupture distance R rup , and shear-wave velocity of the top 30 m of soil V S 30 at the site. These conditional TFPSD images serve as training data for generative neural networks which then generate new TFPSD images conditioned on scenario parameters. A new model is trained using the continuous conditional diffusion model (CCDM), and a framework is proposed to simulate the horizontal components of strike-slip ground motions conditioned on ( M , R rup , V S 30 ). It integrates three parts: (1) variables predicting equations for defining the marginal probability distribution function (PDF), duration, and total energy; (2) the trained CCDM-based model; (3) an algorithm to simulate nonstationary non-Gaussian time histories matching the targets. The proposed framework is validated by comparing pseudospectral acceleration from established ground motion models, actual recorded ground motions, and simulated records, as well as by comparing the Arias intensity, significant durations, peak ground velocity, cumulative absolute velocity and ductility demand calculated using actual and simulated records. The simulated ground motions reproduce several important real seismic features: TFPSD, duration, and nonstationary non-Gaussian behaviour.