Prediction of porous media fluid flow using physics informed neural networks
Prediction of porous media fluid flow using physics informed neural networks
复制标题
使用物理信息神经网络预测多孔介质流体流动
DOI：
10.1016/j.petrol.2021.109205
复制
发表时间：
2021-08-09
影响因子：
--
通讯作者：
Abu-Al-Saud, Moataz O.
中科院分区：
文献类型：
--
作者：
Almajid, Muhammad M.;Abu-Al-Saud, Moataz O.
由于数据数字时代的爆炸式增长，不同物理科学的深度学习应用已经获得了发展势头。在本文中，我们实现了一种物理信息神经网络（PINN）技术，该技术结合了来自流体流动物理学的信息以及观察到的数据来对Buckley-Leverett问题进行建模。经典问题的排水气体到一个充满水的多孔介质被用来验证我们的实现。测试了几种情况，表明观测数据和物理信息神经网络之间的耦合对于不同参数空间的重要性。我们的研究结果表明，PINN是能够捕捉的整体趋势的解决方案，即使没有观测数据，但分辨率和精度的解决方案大大提高了观测数据。仅当使用观测数据时，向PDE约束损失函数添加少量扩散才稍微改善了解决方案。此外，PINN被用来解决反问题，并推断出最佳的多相流参数。PINN的性能相比，人工神经网络（ANN）没有任何物理。我们表明，当用于训练人工神经网络的观察数据包括跨越早期和晚期行为的时间时，人工神经网络的性能与PINN相当。与PINN相反，当只提供早期饱和度曲线作为观测数据和外推时，ANN无法预测解决方案。
Due to the explosion of the digital age of data, deep learning applications for different physical sciences have gained momentum. In this paper, we implement a physics informed neural network (PINN) technique that incorporates information from the fluid flow physics as well as observed data to model the Buckley-Leverett problem. The classical problem of drainage of gas into a water-filled porous medium is used to validate our implementation. Several cases are tested that signify the importance of the coupling between observed data and physics-informed neural networks for different parameter space. Our results indicate that PINNs are capable of capturing the overall trend of the solution even without observed data but the resolution and accuracy of the solution are improved tremendously with observed data. Adding a small amount of diffusion to the PDE-constrained loss function improved the solution slightly only when observed data were used. Moreover, the PINN is used to solve the inverse problem and infer the most optimal multiphase flow parameters. The performance of the PINN is compared to that of an artificial neural network (ANN) without any physics. We show that the ANN performs comparably well to the PINN when the observed data used to train the ANN include times that span the early- and late-time behavior. As opposed to the PINN, the ANN is not able to predict the solution when only early-time saturation profiles are provided as observed data and extrapolation are needed.