Conductivity Prediction In The Porous Media Based On CNN Model

Porous medium is made of a combination of skeleton and pores,including rocks,bones and many man-made materials,etc.It is widely distributed in nature and industrial production scenarios.Among them,the modeling of fluid flow processes in the pore space inside the medium is important for the study of the mechanism of oil and gas field extraction,groundwater protection and other technologies,and even the mechanism of nutrient transport inside living organisms.The process and state of fluid movement in the pore medium are mainly determined by the shape of the pore cross-section.Therefore,the study of the conductivity of pores with different pore cross-sections in porous media has a certain contribution to the study of the conductivity of unconventional and complex pore structures.With the continuous development of deep learning algorithms,among which the convolutional neural network model is widely used in the study of digital images and physical properties because of its characteristics of high accuracy,adequate feature extraction,fast computation and pervasiveness.In this paper,based on previous work,the relationship between pore shape and pore conductivity was investigated by analyzing the pore cross-sectional shape through convolutional neural network,extracting the pore cross-sectional shape features,and predicting the values of conductivity of pores with different cross-sectional shapes.Finally,the mapping relationship from pore cross-sectional shape to the conductivity performance was obtained.The results obtained by the LBM method are used as the "gold standard".The average error of 6% is obtained by comparing the results of this paper and the LBM results,which proves that the proposed method can quickly and accurately predict the conductivity of the pore shape.The specific work is as follows:(1)A sample selection method based on confidence interval estimation is proposed and used according to the sample set characteristics,and the obtained sample set is screened to remove the noisy samples and enhance the significance level of the sample set characteristics.(2)The convolutional neural network model is improved by using the feature pyramid technique,and the mapping relationship from the cross-sectional shape to the corresponding conduction capacity is obtained by scale information fusion and feature extraction.Meanwhile,the model performance is verified by two types of samples,regular shape and real shape.(3)The attention mechanism is used to improve the convolutional neural network model,which improves the model’s ability to extract shape features and has some interpretability.The performance of this method is then compared with the model proposed in this paper by reproducing the previous work to demonstrate the superiority of this method in shape feature extraction and analysis.