Abstract The responses of hydrate reservoir during gas production are complex due to the spatially and temporally evolving thermo-hydro-mechanical properties. Accurate modeling of the behavior, therefore, requires a coupled multiphysics simulator with a large number of parameters, leading to substantial computational demands. This makes it challenging to efficiently predict long-term reservoir responses. In this study, by utilizing an artificial neural network (ANN) algorithm, a meta-model is proposed to deep learn the relationship between the material properties and reservoir responses, including borehole displacement and fluid production. As such, a set of 950 coupled thermo-hydro-mechanical simulations of a one-layer sediment axisymmetric model is carried out for six-day gas production via depressurization. Eighteen input parameters are considered in each simulation covering four physical aspects, namely hydrate dissociation, thermal flow, fluid flow, and mechanical response. With this comprehensive dataset of the responses, a meta-model is established based on the trained neural network, resulting in an efficient prediction of the responses with significantly reduced computational demand. The model is then further utilized to predict the future reservoir responses, and it is found that the results are in a good agreement with those from the fully-coupled simulator.

中文翻译：

---

水合物储层热-水-力耦合行为元建模

摘要 由于热-水-力学性质的时空演化，水合物储层在产气过程中的响应是复杂的。因此，行为的准确建模需要具有大量参数的耦合多物理场模拟器，从而导致大量的计算需求。这使得有效预测长期储层响应变得具有挑战性。在这项研究中，通过利用人工神经网络 (ANN) 算法，提出了一种元模型来深入学习材料特性与储层响应之间的关系，包括井眼位移和流体生产。因此，对单层沉积物轴对称模型进行了一组 950 次耦合的热-水-力学模拟，用于通过减压进行六天的天然气生产。在每个模拟中考虑了 18 个输入参数，涵盖四个物理方面，即水合物分解、热流、流体流动和机械响应。有了这个综合的响应数据集，基于经过训练的神经网络建立元模型，从而在显着减少计算需求的情况下有效预测响应。然后进一步利用该模型来预测未来的储层响应，发现结果与全耦合模拟器的结果非常吻合。从而以显着减少的计算需求对响应进行有效预测。然后进一步利用该模型来预测未来的储层响应，发现结果与全耦合模拟器的结果非常吻合。从而以显着减少的计算需求对响应进行有效预测。然后进一步利用该模型来预测未来的储层响应，发现结果与全耦合模拟器的结果非常吻合。