Distance-dependent spatial correlation modeling of within-event ground-motion residuals using a graph-based generative approach

Tariq Anwar Aquib, & Paul M. Mai

Submitted September 7, 2025, SCEC Contribution #14494, 2025 SCEC Annual Meeting Poster #TBD

Accurate simulation of spatially distributed ground motions is essential for regional seismic risk assessment of spatially distributed infrastructure. While site specific ground-motion models predict shaking intensity at individual locations, they do not capture the spatial correlation structure of within-event residuals, which is critical for simulating realistic ground motions from a single earthquake. Previous studies have quantified spatial correlations, primarily of spectral accelerations, using semi-variograms. These studies consistently report decreasing correlation with increasing interstation separation and a decay rate dependent on spectral period. Reported correlation ranges vary between events, and some models include dependencies on source to site azimuth and site conditions.

However, existing correlation models do not account for variations with source to site distance, likely due to limited near field recordings and the desire for simple distance independent models. Because near-field sites are more strongly influenced by the rupture process, their spatial correlation characteristics may differ from far field sites. To investigate this, we analyze recordings from 62 Japanese earthquakes (Mw 5.5–9.0) with finite-fault inversions, yielding 94 rupture models. We compute within-event residuals for each event and fit exponential semi-variograms to data binned by Joyner–Boore distance (Rjb). For Rjb < 25 km, the fitted correlation range is ~15 km, increasing to ~40-50 km beyond 50 km, highlighting the need for a distance dependent model.

We address this using a machine learning based generative graph neural network (GNN). Stations are represented as nodes, with edges encoding spatial connectivity. Node features include relative epicentral position, Vs30, source-to-site azimuth, and finite-fault distance metrics. The model is trained on the same 62-event dataset, augmented by creating thousands of sub-networks through random station selection. Given a station network and exponentially correlated noise with a random range, the GNN generates spatially correlated within-event residuals. We evaluate the learned correlation structure across distance bins and azimuths and demonstrate its impact on hazard estimates. Results show that incorporating distance-dependent spatial correlations significantly alters near-field hazard estimates compared to traditional models, providing a more physically consistent basis for regional seismic risk assessment.

Key Words
Ground-motion variability, Machine learning, spatial correlation

Citation
Aquib, T., & Mai, P. M. (2025, 09). Distance-dependent spatial correlation modeling of within-event ground-motion residuals using a graph-based generative approach. Poster Presentation at 2025 SCEC Annual Meeting.

Related Projects & Working Groups
Ground Motions (GM)