A deep learning framework for ground motion simulation with time–frequency feature integration toward response-consistent seismograms

Artificial ground motions (GMs) that preserve the effects of time-frequency characteristics (TFCs) of recorded motions on structural responses are important for nonlinear time-history analysis (NLTHA). In seismic design practice, NLTHA is often conducted using a limited number of GMs, such as 7 or 11 records, and their mean response is used for design or performance evaluation. However, conventional spectrum-compatible GMs mainly satisfy target spectral requirements. They do not fully constrain the time characteristics, such as temporal evolution of frequency content, which strongly affect nonlinear responses. Thus, the mean response estimated from a small GM set may deviate from that obtained from a larger set, causing selection-induced bias. This study proposes a deep learning framework for generating response-consistent artificial GMs by incorporating response-relevant TFCs learned from recorded motions. Response diagrams in the time domain (RDTDs), which describe elastic response histories of single-degree-of-freedom systems over a range of periods, are used as compact descriptors of TFC effects. A generative adversarial model, GenRDTD, is developed to synthesize RDTD features associated with target mean structural responses. The synthesized RDTDs are then mapped to acceleration time histories through a Graph-to-GM neural model, producing artificial GMs with controlled energy, timing, and frequency content. An iterative spectrum-alignment procedure is further used to satisfy the target response spectrum while preserving response-relevant TFCs. The generated GMs are evaluated through NLTHA and compared with state-of-the-art methods. Results show that a limited set of generated GMs can reproduce mean structural responses close to those obtained from a larger recorded GM set, while maintaining reasonable response distributions. The proposed method improves the reliability of mean-response estimation using limited GM inputs and reduces the computational cost of NLTHA.