Research On Thermal Conductivity Of Porous Media Materials Based On Machine Learning

With the rapid development of machine learning algorithms,researchers began to intersect research in multiple fields,and materials informatics was born.This makes it easy to solve the problem of material properties that are originally data-centric but cannot be measured or calculated by traditional methods due to efficiency issues such as calculation time and experimental costs.Porous media has a wide range of engineering applications.As important physical properties of porous media,thermal conductivity and microstructure have always been the focus of research by researchers.Therefore,we conduct a cross-research on the thermal conductivity and microstructure of porous media by using machine learning algorithms.The data set of porous media is generated by four-parameter random generation algorithm and Lattice Boltzmann method.In order to consider the influence of microstructure on porous media,we select five factors that affect the effective thermal conductivity of porous media as features,namely porosity,aspect ratio,heat flow path length,average pore size,and anisotropy.Through the control variable method,use ten machine learning algorithms to train ten models respectively and compare the prediction results and the scores on the three evaluation indicators of MAE,RMSE,and R2,and select the optimal machine learning model—the random forest regression model.In addition,the improved autoencoder is used to compress the grayscale image of porous media to obtain hidden vectors.The hidden vector is used to train the fully connected neural network model,and then the trained decoder is spliced with the fully connected network to realize the construction of the porous media reverse design network.Using the convolutional neural network to verify the generated image,it is found that the generated image is basically consistent with the preset input.At the same time,we use more powerful classification conditions to generate an adversarial network model,use porosity,thermal conductivity,and anisotropy as the labels of the grayscale image of porous media,and realize the reverse generation of porous media structure by optimizing the model and adding local loss functions picture.This paper proposes a new framework for intersecting research using machine learning algorithms and porous media.On the one hand,machine learning is used to predict the thermal conductivity of porous media.On the other hand,the self-built reverse design network and classification conditional confrontation network are used to realize the reverse generation of porous media structure through the microstructural parameters of porous media.