Multi-task deep transfer learning for complicated seismic dynamic response prediction in slope systems

Slope engineering is an uncertain, dynamic, and complex nonlinear spatiotemporal system with time delays. High-fidelity prediction of slope seismic stability has long been a formidable challenge due to the inherent randomness and uncertainty associated with ground motion, geo-material properties, complex topography, etc. Traditional numerical modelling always takes a simplified model by forcedly ignoring those uncertainties, thus failing to replicate precisely the intricate nonlinear interactions between factors that affect slope instability. Notably, the newly emerging deep learning methods have the capability of handling multiple factors with uncertainties. However, these methods heavily rely on extensive and comprehensive sensor data, while arranging sensors at certain important positions is sometimes unachievable. Therefore, we propose a multi-task deep transfer learning (MT-DTL) framework in this study to enhance the prediction accuracy of slope seismic response especially in data-limited conditions. The dynamic response at the locations without sufficient accessible sensor data can be effectively predicted with a newly developed algorithm. To collect the necessary sensor data, we conduct a series of physics experiments with the world's largest multifunctional shaking table equipment. We demonstrate the efficacy and accuracy of our approach on the shaking-table datasets through comparisons with traditional machine learning (ML) methods. Our findings reveal that the MT-DTL framework can improve the confidence level of prediction results (within 5%) from the highest 86.4% by the optimal traditional ML methods to 92.7%, achieving comparable results with two-thirds fewer data. Additionally, a single response example showed that the trained deep transfer learning model has significantly improved the computational efficiency (0.018 – 0.019 s) compared to the dynamic finite element calculation with GeoStudio (10 min). This highlights its potential for integration into geo-hazards digital twin systems, facilitating rapid risk analysis based on real-time monitoring data. © 2025 China University of Geosciences (Beijing) and Peking University