Large Model-based Settlement Prediction for Shield Tunneling Constrained by Random Field

In the prediction of ground settlement induced by shield tunneling, the stratum is generally assumed to be homogeneous, which fails to account for its spatial variability and consequently leads to insufficient accuracy and robustness. To address this limitation, a large model for settlement prediction constrained by random fields (RF-LLM) is proposed in this article. This method utilizes random field theory based on the Karhunen-Loève (K-L) expansion to generate spatially variable soil parameters. Combined with FLAC3D simulation, a large number of settlement response samples are obtained. These samples, containing high-dimensional uncertainty information, are then converted into semantic descriptions for fine-tuning a large language model (LLM). The LLM is trained using a hybrid supervision mechanism, enhancing the model's ability to transfer and process uncertainty information. The results indicate that the proposed RF-LLM has distinct advantages over traditional machine learning models in both accuracy and robustness. Especially, the RF-LLM shows greater stability in complex water-rich geological conditions and samples with high variability. Furthermore, this model can output risk interpretations and control recommendations in the form of natural language, achieving a dual-channel output that combines numerical prediction with semantic decision support. This study provides a novel technical approach for disturbance control and intelligent decision-making in shield tunneling construction.