Machine learning-based regional ground seismic motion simulation：A case study of the M_S6.8 Luding earthquake in 2022

The seismic time history plays a crucial role in pre-earthquake risk assessment and post-earthquake damage evaluation. However, due to the uncertainty in simulated ground seismic motion parameters, there are notable differences between simulation results and actual records. Fortunately, advancements in data mining techniques and the availability of high-quality strong ground seismic motion observation records have made it possible to simulate earthquakes by extracting seismic characteristic records.

On September 5, 2022, an M_S6.8 earthquake struck Luding County, Garze Tibetan Autonomous Prefecture, Sichuan Province, China, with epicenter location （29.59°N, 102.08°E）, and a focal depth of 16 km. This event occurred in the southeastern segment of the Xianshuihe fault zone, near the Moxi fault and Jinpingshan fault. According to the intensity survey results released by the China Earthquake Administration, the major axis of the isoseismal lines of the Luding earthquake trended northwest, with dimensions of 195 km by 112 km, and the intensity in the most affected area reached level Ⅸ. The area with intensity Ⅵ or above covered 19 089 square kilometers. The earthquake caused significant damage to infrastructure, including roads, bridges, buildings, and lifeline projects, resulting in casualties and triggering secondary hazards such as landslides. Following the Luding M_S6.8 main shock, several aftershocks occurred, the most notable of which was the M_S5.0 event on October 22, 2022.

This study selected seismic records from the M_S5.0 earthquake on October 22, 2022. According to the Global Centroid Moment Tensor （GCMT） results, the focal mechanism of the M_S6.8 main shock indicates a strike of 164°, dip of 78°, and slip of 7°, and the M_S5.0 aftershock focal mechanism shows a strike of 167°, dip of 48°, and slip of −53°. The main shock is characterized as a strike-slip earthquake, while the aftershock involves both thrust and strike-slip components. Fifteen stations with three-component records from the aftershocks were selected for feature principal component time history extraction, and then baseline correction, and filtering were applied to the seismic records. Due to the complex influence of rupture mechanisms, propagation paths, and site conditions, seismic records at different stations capture different seismic features. This study employs principal component analysis （PCA）, a machine learning-based data dimensionality reduction method, to extract seismic features. The PCA process involves treating each seismic record as a row vector, concatenating all seismic records to form a matrix, and applying linear transformations to map the high-dimensional vectors into a lower-dimensional space. The resulting matrix, composed of 45 seismic records from 15 stations, contains the primary features of the Luding earthquake. The number of feature principal component time histories is determined by selecting those with cumulative contribution rates exceeding 95%.

The simulation of ground motions for the main shock at target stations requires the calculation of coefficients for the 15 feature principal component time histories obtained through PCA. Particle swarm optimization （PSO） is employed to find the combination coefficients that minimize the error between the simulated ground motions and the actual records, with peak acceleration and acceleration response spectrum as constraints. This algorithm is an optimization computing technique that takes into account both the individual cognition of particles and the social influence of the population. Twelve target stations near the epicenter of the Luding main shock were selected for simulation, and PSO was used to calculate the combination coefficients for the feature principal component time histories. The simulated ground motions were then compared with the actual records.

The simulation results for most stations are satisfactory, with small errors between simulated and actual values in both high-frequency and low-frequency ranges. However, there is some discrepancy in the simulation results for the EW-direction seismic records between stations 51MNT and 51HYQ. To better assess the contribution of data from the stations 51HYQ and 51MNT to the feature principal component time histories, principal component analysis was conducted after removing data from these two stations. It was found that the number of feature principal component time histories with a cumulative contribution rate of 95% decreased from 15 to 12. The simulation results of seismic time history and response spectra remained consistent with the results before removal, indicating that the contribution of data from these two stations to the ground seismic motion simulation in this study is relatively small. To improve the simulation for these distant stations, the study replaced the four feature principal component time histories with lower contribution rates, leading to better results. The comparison of actual and simulated ground seismic motions, as well as acceleration response spectra, demonstrates the effectiveness of the proposed approach.

This study leverages machine learning techniques, specifically principal component analysis and particle swarm optimization, to simulate ground motions for the Luding earthquake sequence. The analysis demonstrates that it is feasible to extract seismic feature principal component time histories from aftershock data and to use them for main shock simulation. The combination of PSO optimization, peak acceleration, and acceleration response spectrum as constraints yields accurate simulations. The study suggests that utilizing data-driven approaches, especially in the regions with frequent seismic activity, provides valuable insights into the potential impact of strong earthquakes.