We present a novel physics‐guided neural network to estimate shear‐tensile focal mechanisms for microearthquakes using displacement amplitudes of direct P waves. Compared with conventional data‐driven fully connected (FC) neural networks, our physics‐guided neural network is implemented in an unsupervised fashion and avoids the use of training data, which may be incomplete or unavailable. We incorporate three FC layers and a scaling and shifting layer to estimate shear‐tensile focal mechanisms for multiple events. Then, a forward‐modeling layer, which generates synthetic amplitude data based on the source mechanisms emerging from the previous layer, is added. The neural network weights are iteratively updated to minimize the mean squared error between observed and modeled normalized P‐wave amplitudes. We apply this machine‐learning approach to a set of 530 induced events recorded during hydraulic‐fracture simulation of Duvernay Shale west of Fox Creek, Alberta, yielding results that are consistent with previously reported source mechanisms for the same dataset. A distinct cluster characterized by more complex mechanisms exhibits relatively large Kagan angles (5°–25°) compared with the previously reported best double‐couple solutions, mainly due to model simplification of the shear‐tensile focal mechanism. Uncertainty tests demonstrate the robustness of the inversion results and high tolerance of our neural network to errors in event locations, the velocity model, and P‐wave amplitudes. Compared with a single‐event grid‐search algorithm to estimate shear‐tensile focal mechanisms, the proposed neural network approach exhibits significantly higher computational efficiency.

中文翻译：

---

使用无监督物理引导神经网络反演剪切拉伸焦距机制

我们提出了一种新的物理引导的神经网络，以使用直接 P 波的位移幅度来估计微地震的剪切-拉伸震源机制。与传统的数据驱动的全连接 (FC) 神经网络相比，我们的物理引导神经网络以无监督的方式实现，避免使用可能不完整或不可用的训练数据。我们结合了三个 FC 层和一个缩放和移动层来估计多个事件的剪切-拉伸聚焦机制。然后，添加一个前向建模层，该层基于从前一层出现的源机制生成合成幅度数据。神经网络权重被迭代更新，以最小化观察到的和建模的归一化 P 波振幅之间的均方误差。我们将这种机器学习方法应用于在阿尔伯塔省 Fox Creek 以西的 Duvernay 页岩水力压裂模拟过程中记录的一组 530 个诱发事件，产生的结果与先前报告的同一数据集的源机制一致。与之前报道的最佳双偶解相比，以更复杂机制为特征的独特集群表现出相对较大的卡根角（5°-25°），这主要是由于剪切-拉伸聚焦机制的模型简化。不确定性测试证明了反演结果的稳健性和我们的神经网络对事件位置、速度模型和 P 波幅度误差的高容忍度。与估计剪切-拉伸聚焦机制的单事件网格搜索算法相比，