Abstract Wellbore instability is a major safety and environmental concern in both onshore and offshore drilling. Geological drilling risk assessment for wellbore collapse is critical for the optimization of well plans, and to reduce potential costs before drilling. In this work, we propose a physics-guided deep learning approach to predict wellbore instability using seismic attributes data. We first trained an auto-encoder to extract latent representation (five principle features) from 17 typical seismic attributes, and then we introduced a regularization term, based on geomechanics in the objective function to train a neural network. As long as there is no significant over-pressure in the formation, the physics-based regularization term indicating wellbore instability risk is a function of neutron porosity, and of the true vertical depth obtained from well logging data. In this way, we combined drill log data for five wells, as prior information, with latent seismic attribute representations, to train the neural network. After training, our approach needed only seismic data to predict wellbore instability risk in new locations, and our case study showed that the physics-based regularizer, with an appropriate weight, prevented overfitting to training data and enhanced the generalization accuracy of the neural network (by ~4%) in two new test wells. We argue that statistical correlations between seismic attributes and rock properties are algorithm dependent, and have to be treated cautiously in the absence of a base of petrophysical reasoning. The physics-guided deep learning method presented here has potential application for the quantification of geology-based wellbore instability risk before drilling.

中文翻译：

---

基于物理的深度学习利用地震属性数据预测井眼失稳地质钻井风险

摘要 井筒不稳定是陆上和海上钻井中的主要安全和环境问题。井眼坍塌的地质钻井风险评估对于优化井计划和降低钻井前的潜在成本至关重要。在这项工作中，我们提出了一种物理引导的深度学习方法，使用地震属性数据预测井眼不稳定性。我们首先训练了一个自动编码器来从 17 个典型的地震属性中提取潜在表示（五个主要特征），然后我们引入了一个正则化项，基于目标函数中的地质力学来训练神经网络。只要地层中没有明显的超压，表明井眼不稳定风险的基于物理的正则化项是中子孔隙度的函数，以及从测井数据获得的真实垂直深度。通过这种方式，我们将 5 口井的钻井测井数据作为先验信息与潜在的地震属性表示相结合，以训练神经网络。训练后，我们的方法只需要地震数据来预测新位置的井筒不稳定风险，我们的案例研究表明，具有适当权重的基于物理的正则化器可以防止过度拟合训练数据并提高神经网络的泛化精度（约 4%）在两个新的测试井中。我们认为地震属性和岩石特性之间的统计相关性取决于算法，并且在没有岩石物理推理基础的情况下必须谨慎对待。