Research On Permeability Prediction Of Porous Media Based On Deep Learning

As an inherent characteristic parameter of porous media,permeability is often used to represent the seepage capacity of porous media,and is one of the important hydraulic parameters of porous media.For the calculation of the permeability of porous media,the existing methods include hydrodynamic numerical method,pore network model method,lattice Boltzmann method and so on.However,such methods are prone to the limitations of difficult convergence and high computational cost when calculating complex porous media.Deep learning is a hot research method used in the field of computer vision in recent years,and its image modeling technology has the advantage of being fast and accurate.Considering that the permeability of porous media is only related to the physical structure of pores,and the physical structure of porous media can be represented by image information,this thesis proposes a deep learning method to model the porous media image and the permeability value,and then calculate the permeability.predict.The 2D and 3D convolutional neural networks were compared,and the 3D convolutional model with better prediction performance was adjusted.Finally,the optimized 3D convolutional network model was obtained.Through testing and analysis of unknown porous media data,it was verified that the superior performance of the predictive model.The main research work of this thesis is as follows.(1)Obtain a database of images of porous media structures.In this thesis,high-resolution X-ray three-dimensional micro-CT is used to scan and image the physical porous medium.After three-dimensional reconstruction and image post-processing operations,high-quality three-dimensional physical porous medium structure images are obtained.The study of deep learning methods requires a representative database.In order to obtain a rich and diverse sample of porous media,this thesis randomly generates batches of 3D synthetic porous media image data through a computer simulation synthesis algorithm.(2)Obtain porous media structure image permeability labels.In this thesis,the Multirelaxation time Boltzmann method(Multi-relaxation time LBM,MRT-LBM)is used to simulate the porous media structure database.In order to use computer resources to obtain faster computing speed,this thesis uses multi-GPU parallel acceleration technology,the corresponding permeability label of each porous medium structure is obtained by Matlab programming calculation.At the same time,in order to verify the effectiveness of the method,we use the experimental method for verification analysis.Due to the limitations of the experimental method,such as certain errors and inevitable conditions,We introduced the three-dimensional Poiseuille flow simulation into the MRT-LBM model and compared the calculation results of the Single relaxation time Boltzmann method(Single-relaxation-time LBM,SRT-LBM).The results further verify the effectiveness of the method of calculating the permeability label in this thesis.(3)Obtain a deep learning predictive model.First,the basic 2D and 3D convolutional networks are modeled through the deep learning framework Tensorlfow,and the appropriate regression model performance evaluation indicators are selected to evaluate and analyze the performance of the corresponding convolutional model on the porous media structure test data set,and then compare the comprehensive performance.The optimal 3D convolution model is used for further optimization operations such as structural adjustment and hyperparameter debugging,and finally a 3D convolution network model that can quickly and accurately predict the permeability of porous media is obtained.The end-to-end deep learning model constructed in this thesis from the input of porous media to the output of permeability has good prediction performance.The optimal value of the evaluation index of the prediction results on the test set is MAPE of 2.17%,R^2 of 0.9886,and RMSE of 0.2847,which shows that the prediction of the deep learning model in this paper has high accuracy.In addition,the prediction time of the permeability of each porous medium is 10~20ms,which is much faster than the lattice Boltzmann method.The experimental results show that the deep learning technology has certain research significance and development potential in the prediction of porous media permeability.