Understanding and simulating the underlying microscopic physics of the rock matrix is very useful for determining macroscopic physical properties such as permeability. Matrix diffusion is an important transport parame...Understanding and simulating the underlying microscopic physics of the rock matrix is very useful for determining macroscopic physical properties such as permeability. Matrix diffusion is an important transport parameter controlling the late-time behaviour of breakthrough curves (BTCs). We compute the memory function, implemented in the sink/source term of Mobile-immobile mass transfer by solving the matrix diffusion using a time diffusion random-walk approach. The diffusion is controlled by different parameters like the porosity, tortuosity, mobile-immobile interface and immobile domain cluster shapes. All these properties are investigated by X-ray microtomography that captures the main characteristics of matrix diffusion at three dimensions. We compare the memory function deduced from the field-scale tracer tests well with the computed memory function. Simulation results of the memory function appeared to be coherent with that measured from the tracer test for a large tortuosity value. Probably, the diffusion paths are longer, and they are controlled by the properties mentioned above. From a representative elementary volume of natural reservoirs studied here, we conclude that, microscale diffusion process in the immobile domain play a crucial role to better understand the non-Fickian dispersion measured from the tracer test.展开更多
链路预测是通过已知网络节点或者网络拓扑结构预测未产生链接的两个节点间产生链接的可能性.传统方法大多从原始图中提取转移矩阵,导致获取的信息稀疏.鉴于此,设计了一种基于图神经网络和随机游走的链路预测框架(link prediction-graph ...链路预测是通过已知网络节点或者网络拓扑结构预测未产生链接的两个节点间产生链接的可能性.传统方法大多从原始图中提取转移矩阵,导致获取的信息稀疏.鉴于此,设计了一种基于图神经网络和随机游走的链路预测框架(link prediction-graph neural network and random walk,LP-GNRW).首先,通过基于注意力机制的图神经网络Bert学习节点的多种嵌入表示;然后,结合随机游走,获取图的高阶结构信息;最后,将链路预测转换成二分类问题,通过图神经网络对获得的高阶结构信息进行二分类实现链路预测.实验表明LPGNRW能更有效地学习图结构特征,与基于步行的启发式方法相比,获得了更好的AUC指标,提高了链路预测的性能.展开更多
在现有的图聚类方法中,大多数聚类方法只关注图的拓扑结构或节点属性而忽略另一方面.为解决这一问题,相关文献中提出了基于图的结构与属性的图聚类方法.但这些聚类方法存在建立的图模型不准确、聚类效果不理想、算法执行效率低等缺点....在现有的图聚类方法中,大多数聚类方法只关注图的拓扑结构或节点属性而忽略另一方面.为解决这一问题,相关文献中提出了基于图的结构与属性的图聚类方法.但这些聚类方法存在建立的图模型不准确、聚类效果不理想、算法执行效率低等缺点.针对上述图聚类方法中存在的问题,提出了一种基于结构-属性的时空对象图聚类方法(spatio-temporal object graph clustering algorithm based on structure and attribute,STSA).首先提出了属性加权图模型,在此基础上建立了结构-属性的统一度量方法,并采用随机游走模型技术将节点间结构与属性关系转换为相应的相似度矩阵,结合图结构-属性关系及相似度矩阵,采用信息传递算法对图进行聚类,解决了现有图聚类方法中所存在的问题,最后通过实验验证了提出的STSA方法的正确性和有效性.展开更多
文摘Understanding and simulating the underlying microscopic physics of the rock matrix is very useful for determining macroscopic physical properties such as permeability. Matrix diffusion is an important transport parameter controlling the late-time behaviour of breakthrough curves (BTCs). We compute the memory function, implemented in the sink/source term of Mobile-immobile mass transfer by solving the matrix diffusion using a time diffusion random-walk approach. The diffusion is controlled by different parameters like the porosity, tortuosity, mobile-immobile interface and immobile domain cluster shapes. All these properties are investigated by X-ray microtomography that captures the main characteristics of matrix diffusion at three dimensions. We compare the memory function deduced from the field-scale tracer tests well with the computed memory function. Simulation results of the memory function appeared to be coherent with that measured from the tracer test for a large tortuosity value. Probably, the diffusion paths are longer, and they are controlled by the properties mentioned above. From a representative elementary volume of natural reservoirs studied here, we conclude that, microscale diffusion process in the immobile domain play a crucial role to better understand the non-Fickian dispersion measured from the tracer test.
文摘链路预测是通过已知网络节点或者网络拓扑结构预测未产生链接的两个节点间产生链接的可能性.传统方法大多从原始图中提取转移矩阵,导致获取的信息稀疏.鉴于此,设计了一种基于图神经网络和随机游走的链路预测框架(link prediction-graph neural network and random walk,LP-GNRW).首先,通过基于注意力机制的图神经网络Bert学习节点的多种嵌入表示;然后,结合随机游走,获取图的高阶结构信息;最后,将链路预测转换成二分类问题,通过图神经网络对获得的高阶结构信息进行二分类实现链路预测.实验表明LPGNRW能更有效地学习图结构特征,与基于步行的启发式方法相比,获得了更好的AUC指标,提高了链路预测的性能.
文摘在现有的图聚类方法中,大多数聚类方法只关注图的拓扑结构或节点属性而忽略另一方面.为解决这一问题,相关文献中提出了基于图的结构与属性的图聚类方法.但这些聚类方法存在建立的图模型不准确、聚类效果不理想、算法执行效率低等缺点.针对上述图聚类方法中存在的问题,提出了一种基于结构-属性的时空对象图聚类方法(spatio-temporal object graph clustering algorithm based on structure and attribute,STSA).首先提出了属性加权图模型,在此基础上建立了结构-属性的统一度量方法,并采用随机游走模型技术将节点间结构与属性关系转换为相应的相似度矩阵,结合图结构-属性关系及相似度矩阵,采用信息传递算法对图进行聚类,解决了现有图聚类方法中所存在的问题,最后通过实验验证了提出的STSA方法的正确性和有效性.
