Generative digital twin for deformation field reconstruction of high rockfill dams based on diffusion models

Deformation perception of rockfill dams is crucial for formulating and implementing safety control measures. However, due to the sparsity of deformation monitoring points and the failure of instruments, existing deformation monitoring techniques cannot fully meet the perception requirements. Meanwhile, the fusion of monitoring and simulation data for rockfill dams remains insufficient, limiting further improvement in both accuracy and real-time performance. Thus, this study develops a generative digital twin (DT) model for the reconstruction of rockfill dam deformation fields during the operation period based on Denoising Diffusion Probabilistic Model (DDPM). In model training, an active sampling strategy is introduced to improve training efficiency and generation diversity, thereby enhancing the expressive capability. During sampling, a monitoring-consistency gradient guidance strategy is proposed, in which sparse monitoring data are embedded into the reverse diffusion process to dynamically guide the sampling process. The proposed model is applied and validated in the tallest constructed rockfill dam in the world, Lianghekou Dam (303 m). The model inference time for generating a single deformation field is only nearly 30 s, and the overall point-wise relative reconstruction accuracy with respect to monitoring data reaches 90.7%, indicating its potential to satisfy the accuracy and timeliness requirements of deformation perception for rockfill dams. Codes and experimental data can be found at https://github.com/aizhitao123-art/Generative-deformation-field-reconstruction