物理信息神经网络(physics-informed neural networks,PINN)由于嵌入了物理先验知识,可以在少量训练数据的情况下获得自动满足物理约束的代理模型,受到了智能科学计算领域的广泛关注.但是,PINN的离散时间模型(PINN-RK)无法同时近似多个...物理信息神经网络(physics-informed neural networks,PINN)由于嵌入了物理先验知识,可以在少量训练数据的情况下获得自动满足物理约束的代理模型,受到了智能科学计算领域的广泛关注.但是,PINN的离散时间模型(PINN-RK)无法同时近似多个物理量相互耦合的偏微分方程系统,限制了其处理复杂多物理场的能力.为了打破这一限制,文章提出了一种基于龙格库塔法的多输出物理信息神经网络(multi-output physics-informed neural networks based on the Runge-Kutta method,MO-PINN-RK),MO-PINN-RK模型在离散时间模型的基础上采用了并行输出的神经网络结构,通过将神经网络划分为多个子网络,建立了多个神经网络输出层.采用不同输出层近似不同物理量的方式,MO-PINN-RK模型不仅可以同时表征多个物理量,而且还能够实现求解偏微分方程系统的目的.另外,MO-PINN-RK克服了PINN离散时间模型仅适用于一维空间的局限性,将其应用范围扩展到了更为普遍的多维空间.为了验证MO-PINN-RK的有效性,文章对圆柱绕流问题进行了流场预测和参数辨识研究.测试结果表明,与PINN相比,MO-PINN-RK在流场预测问题中的准确性获得了提升,其精度至少提高了2倍,而在参数辨识问题中,MO-PINN-RK的相对误差降低了一个数量级.这凸显了MO-PINN-RK在流体动力学领域的卓越能力,为解决复杂问题提供了更准确、更有效的解决方案.展开更多
For the case of carbonate reservoir water flooding development, the flow field identification method based on streamline modeling result was proposed. The Ocean for Petrel platform was used to build the plug-in that e...For the case of carbonate reservoir water flooding development, the flow field identification method based on streamline modeling result was proposed. The Ocean for Petrel platform was used to build the plug-in that exported the streamline data, and the subsequent data was processed and clustered through Python programming, to display the flow field with different water flooding efficiencies at different time in the reservoir. We used density peak clustering as primary streamline cluster algorithm, and Silhouette algorithm as the cluster validation algorithm to select reasonable cluster number, and the results of different clustering algorithms were compared. The results showed that the density peak clustering algorithm could provide better identified capacity and higher Silhouette coefficient than K-means, hierachical clustering and spectral clustering algorithms when clustering coefficients are the same. Based on the results of streamline clustering method, the reservoir engineers can easily identify the flow area with quantification treatment, the inefficient water injection channels and area with developing potential in reservoirs can be identified. Meanwhile, streamlines between the same injector and producer can be subdivided to describe driving capacity distribution in water phase, providing useful information for the decision making of water flooding optimization, well pattern adjustment and deep profile modification.展开更多
In this paper, on the basis of experimental data of two kinds of chemical explosions, the piston-pushing model of spherical blast-waves and the second-order Godunov-type scheme of finite difference methods with high i...In this paper, on the basis of experimental data of two kinds of chemical explosions, the piston-pushing model of spherical blast-waves and the second-order Godunov-type scheme of finite difference methods with high identification to discontinuity are used to the numerical reconstruction of part of an actual hemispherical blast-wave flow field by properly adjusting the moving bounary conditions of a piston. This method is simple and reliable. It is suitable to the evaluation of effects of the blast-wave flow field away from the explosion center.展开更多
文摘物理信息神经网络(physics-informed neural networks,PINN)由于嵌入了物理先验知识,可以在少量训练数据的情况下获得自动满足物理约束的代理模型,受到了智能科学计算领域的广泛关注.但是,PINN的离散时间模型(PINN-RK)无法同时近似多个物理量相互耦合的偏微分方程系统,限制了其处理复杂多物理场的能力.为了打破这一限制,文章提出了一种基于龙格库塔法的多输出物理信息神经网络(multi-output physics-informed neural networks based on the Runge-Kutta method,MO-PINN-RK),MO-PINN-RK模型在离散时间模型的基础上采用了并行输出的神经网络结构,通过将神经网络划分为多个子网络,建立了多个神经网络输出层.采用不同输出层近似不同物理量的方式,MO-PINN-RK模型不仅可以同时表征多个物理量,而且还能够实现求解偏微分方程系统的目的.另外,MO-PINN-RK克服了PINN离散时间模型仅适用于一维空间的局限性,将其应用范围扩展到了更为普遍的多维空间.为了验证MO-PINN-RK的有效性,文章对圆柱绕流问题进行了流场预测和参数辨识研究.测试结果表明,与PINN相比,MO-PINN-RK在流场预测问题中的准确性获得了提升,其精度至少提高了2倍,而在参数辨识问题中,MO-PINN-RK的相对误差降低了一个数量级.这凸显了MO-PINN-RK在流体动力学领域的卓越能力,为解决复杂问题提供了更准确、更有效的解决方案.
基金Supported by the the CNPC Science and Technology Innovation Fund Program(2017D-5007-0202)
文摘For the case of carbonate reservoir water flooding development, the flow field identification method based on streamline modeling result was proposed. The Ocean for Petrel platform was used to build the plug-in that exported the streamline data, and the subsequent data was processed and clustered through Python programming, to display the flow field with different water flooding efficiencies at different time in the reservoir. We used density peak clustering as primary streamline cluster algorithm, and Silhouette algorithm as the cluster validation algorithm to select reasonable cluster number, and the results of different clustering algorithms were compared. The results showed that the density peak clustering algorithm could provide better identified capacity and higher Silhouette coefficient than K-means, hierachical clustering and spectral clustering algorithms when clustering coefficients are the same. Based on the results of streamline clustering method, the reservoir engineers can easily identify the flow area with quantification treatment, the inefficient water injection channels and area with developing potential in reservoirs can be identified. Meanwhile, streamlines between the same injector and producer can be subdivided to describe driving capacity distribution in water phase, providing useful information for the decision making of water flooding optimization, well pattern adjustment and deep profile modification.
文摘In this paper, on the basis of experimental data of two kinds of chemical explosions, the piston-pushing model of spherical blast-waves and the second-order Godunov-type scheme of finite difference methods with high identification to discontinuity are used to the numerical reconstruction of part of an actual hemispherical blast-wave flow field by properly adjusting the moving bounary conditions of a piston. This method is simple and reliable. It is suitable to the evaluation of effects of the blast-wave flow field away from the explosion center.
