Bayesian inverse analysis of unsaturated slope parameters using fine-tuning deep operator network model[J]. Chinese Journal of Geotechnical Engineering. DOI: 10.11779/CJGE20250067
Citation:
Bayesian inverse analysis of unsaturated slope parameters using fine-tuning deep operator network model[J]. Chinese Journal of Geotechnical Engineering. DOI: 10.11779/CJGE20250067
Bayesian inverse analysis of unsaturated slope parameters using fine-tuning deep operator network model[J]. Chinese Journal of Geotechnical Engineering. DOI: 10.11779/CJGE20250067
Citation:
Bayesian inverse analysis of unsaturated slope parameters using fine-tuning deep operator network model[J]. Chinese Journal of Geotechnical Engineering. DOI: 10.11779/CJGE20250067
The Bayesian method is an effective tool for inferring the posterior distribution of parameters by combining prior information with monitoring data. This process typically requires thousands of computationally expensive numerical model evaluations, leading to substantial computational costs. To alleviate this, surrogate models are often used as substitutes for time-consuming numerical computations. However, the current Bayesian inverse analysis methods fail to account for the spatio-temporal evolution characteristics of the output response. For monitoring data with spatio-temporal variations, surrogate models must be constructed separately at different temporal and spatial points. Additionally, the integration of a large number of time-series monitoring data requires multiple Bayesian updates. Traditional methods generally rely on surrogate models based solely on prior information for Bayesian inverse analysis, which leads to poor computational accuracy in the posterior inference. To address these challenges, this paper proposes a Bayesian inverse analysis method based on fine-tuning deep operator network (DeepONet) model by combining Bayesian updating methods with DeepONet. This method incorporates spatiotemporal characteristics into the Bayesian inverse analysis by replacing the numerical model with DeepONet model that reflects the spatiotemporal evolution of the output response. Additionally, extra samples are selected in each subset simulation layers to fine-tune the surrogate model, ensuring the accuracy of the posterior distribution inference. The proposed method is validated using a case study of a slope in Hong Kong. The results demonstrate that the proposed method effectively addresses the challenge of inferring the posterior distribution of slope parameters from a large number of time-series monitoring data by constructing a surrogate model that reflects the spatiotemporal evolution characteristics of output responses and performing real-time fine-tuning. Furthermore, this method lays the foundation for studying the evolution of unsaturated slope’s stability under rainfall infiltration.