Two‐dimensional analysis of tunnel design based on the convergence–confinement method, although commonly used in tunnel design, may not always be applied. For example, in squeezing grounds, if the support is installed very close to the tunnel face, three‐dimensional numerical modeling is required but is computationally expensive. Therefore, it is usually performed before or after tunnel excavation. A machine learning approach is presented here as an alternative to costly computations. Two surrogate models are developed based on synthetic data. The first model aims to assess the support pressure and the radial displacement at equilibrium in the lining and the radial displacement occurring close to the face at the installation distance of the support. The second model is intended to compute the extrusion of the core considering an unlined gallery. It is assumed a circular tunnel excavated in a Mohr–Coulomb elastoplastic perfectly plastic ground under an initial isotropic stress state. In particular, the bagging method is applied to neural networks to enhance the generalization capability of the models. A good performance is obtained using relatively scarce datasets. The modeling of the surrogate models is explained from the creation of the synthetic datasets to the evaluation of their performance. Their limitations are discussed. In practice, these two machine learning tools should be helpful in the field during the excavation phase.

中文翻译：

---

数据驱动工具，用于评估在弹塑性地基中开挖的隧道的支座压力、径向位移和工作面挤压


基于收敛-约束法的隧道设计二维分析虽然在隧道设计中常用，但可能并不总是适用。例如，在挤压地基中，如果支座安装在非常靠近隧道掌子面的地方，则需要进行三维数值建模，但计算成本很高。因此，通常在隧道开挖之前或之后进行。这里介绍了一种机器学习方法，作为昂贵的计算的替代方案。基于合成数据开发了两个代理模型。第一个模型旨在评估支座压力和衬砌平衡时的径向位移，以及在支座安装距离处靠近工作面发生的径向位移。第二个模型旨在计算考虑无衬线廊的型芯的拉伸。假设在初始各向同性应力状态下，在 Mohr-Coulomb 弹塑性完美塑性地基中开挖一条圆形隧道。特别是，将 bagging 方法应用于神经网络，以增强模型的泛化能力。使用相对稀缺的数据集可以获得良好的性能。从合成数据集的创建到其性能的评估，都解释了代理模型的建模。讨论了它们的局限性。在实践中，这两个机器学习工具应该在挖掘阶段的野外提供帮助。