image: Based on using peridynamics to describe the physical processes of regional land subsidence, deep learning methods, including neural networks and Gaussian Process Regression, are employed to construct various boundary conditions that adapt to temporal developments and changes. This approach enables the optimization of boundary conditions within the peridynamics-based land subsidence model.
Credit: Beijing Zhongke Journal Publising Co. Ltd.
Land subsidence is a geological disaster that occurs when natural factors or human engineering activities cause the consolidation and compression of loose underground rock formations, resulting in a drop in ground elevation over a certain area. Currently, the problem of land subsidence is getting more serious. Its complexity, harm, and uncertainty are on the rise, affecting over 150 countries and regions globally. Modeling, simulating, and predicting land subsidence is of utmost importance for the sustainable development of the regional resource - environment and social economy.
As a new physical model method in the research of land subsidence, peridynamics builds models based on the idea of non-local interaction and can be used to study the deformation, damage, and fracture of homogeneous and inhomogeneous targets. Existing research has shown that peridynamics has good applicability in land subsidence modeling. Considering that physical models and deep learning can complement each other very well, their integration is a promising approach to further improve the simulation and prediction of land subsidence.
Recently, a study in the Journal of Geo-information Science brought good news. A research team from Capital Normal University, led by Professor Huili Gong and Professor Xiaojuan Li, proposed an innovative method. By combining peridynamics and deep learning, they aimed to enhance the accuracy of ground subsidence simulation. The research focused on Tongzhou District in Beijing, which has long been troubled by land subsidence. The team analyzed data from September 2021 to May 2023, dividing it into training and test sets. Results indicated that for the combined model, the Root Mean Square Error (RMSE) was 6.25 mm for the training set and 7.71 mm for the test set. Compared with traditional methods, the errors were reduced by 72.37% and 65.92% respectively.
This research highlights the complementary strengths of physical models and deep learning. Peridynamics provides a solid framework for understanding land subsidence's physical processes, while deep learning improves the model's ability to handle complex and dynamic boundary conditions. Their combination offers a more effective way to simulate land subsidence in areas with complex geological structures. The study's results are significant for urban planning, disaster prevention, and mitigation, especially in areas prone to land subsidence, and can strongly support the safe development of cities.
For more details, please refer to the original article:
Cheng S Y, Guan Z B, Gong H L, Li X J*, et al. Land subsidence modeling and simulation methods using peridynamics and deep learning[J]. Journal of Geo-information Science, 2025, 27(1): 181-192. https://www.dqxxkx.cn/CN/10.12082/dqxxkx.2025.240506
https://www.sciengine.com/JGIS/doi/10.12082/dqxxkx.2025.240506（If you want to read the English version of the full text, please click on the “iFLYTEK Translation” in the article pages.）
Article Title
Land Subsidence Modeling and Simulation Methods Using Peridynamics and Deep Learning
Article Publication Date
25-Jan-2025