Different Perspectives on the Application of Deep Learning for Pore-Scale Two-Phase Flow

报告题目：Different Perspectives on the Application of Deep Learning for Pore-Scale Two-Phase Flow

报告人： 蒋泽云（Heriot-Watt大学，教授）

时间：2023年6月8日 16:30-18:30

地点：明理楼C302B

     摘要: The recent pore-scale literature is not short of compelling studies on deep learning applications. The prediction of two-phase flow fields, however, remains elusive. This is partly due to focusing on model architecture and data quality, rather than the quantity of data. This work presents an end-to-end and accurate deep learning workflow to predict phase distributions during steady-state two-phase drainage, directly from dry images and input features of pixel size, IFT, contact angle, and pressure. A highly diverse dataset is first constructed by subsampling CT scans of synthetic and real rocks. We then devise a new vision transformer (ViT) that drains pores solely based on their size, regardless of their spatial location, where the phase connectivity to inlet(s) is enforced as a post-processing step. With this setup, inference on images of any size with various pixel sizes can efficiently be made by patching input images and stitching results.

       报告人简介：蒋泽云教授从2004年以后主要从事孔隙介质（如岩石、土壤等）异质多尺度结构分析和流体渗流模型研究, 在Water Resource research, Transport in Porous Media, Fuel等期刊和国际会议上发表论文40多篇，在国内外参与并主导若干重大科研项目。主要从事微观空隙结构分析及网络渗流模型的研究, 独立开发软件系统PAT–Pore Analysis Tools（这一软件在学术和工业界得到广泛的使用）。擅长于岩石微观图像几何拓扑属性的分析,并建立多尺度孔隙模型和实施数值模拟,建立其微观结构（如孔隙度、孔尺寸、形状、连通性、孔壁粗糙度等）与宏观流体属性（如渗透率、毛细管压力、性、相对渗透率、电阻率等）间的理论或经验公式。

主办单位: 理学院､人工智能研究院､非线性动力系统研究所､数理力学研究中心 ､科学技术发展研究院