Fractures are widely distributed in aquifers, and their geometric characteristics and hydraulic properties play a crucial role in groundwater flow and the migration of contaminants. Therefore, accurately characterizing fractures within aquifers is of great importance. In this study, we developed a novel method to invert the geometric features and aperture of fractures in fractured aquifers using deep learning algorithms. This method utilizes hydraulic head and multi-phase solute concentration data from synthetic experiments to invert information about the fractured aquifer. Based on the SegNet network, the developed model can establish a one-to-one correspondence between input and output, rather than an uncertain probabilistic output. Furthermore, the multi-task network used in this study, through weight sharing and skip connections, improves the accuracy of the network by 5.59% compared to single network. The application of this method in synthetic experiments demonstrates that, compared to traditional methods, it offers higher efficiency and accuracy, benefiting from the special mechanisms of convolutional networks and exhibiting better generalization capabilities.

中文翻译：

---

多任务神经网络结合多源数据对离散裂隙网络孔径进行反演：Aperture-XNET

裂缝在含水层中广泛分布，其几何特征和水力特性在地下水流和污染物迁移中起着至关重要的作用。因此，准确描述含水层内的裂缝非常重要。在这项研究中，我们开发了一种使用深度学习算法反转裂缝含水层中裂缝的几何特征和孔径的新方法。该方法利用来自合成实验的水力水头和多相溶质浓度数据来反转有关裂缝含水层的信息。基于 SegNet 网络，开发的模型可以在输入和输出之间建立一一对应的关系，而不是不确定的概率输出。此外，本研究中使用的多任务网络通过权重共享和跳过连接，与单个网络相比，网络的准确性提高了 5.59%。该方法在合成实验中的应用表明，与传统方法相比，该方法具有更高的效率和准确性，受益于卷积网络的特殊机制，表现出更好的泛化能力。