Multiphase Flow Prediction And Uncertainty Quantification In Porous Media Based On Deep Learning

Porous media widely exist in nature and many industrial and agricultural production scenarios.There are complex flow characteristics in the immiscible multiphase fluids in porous media.Modeling and solving the flow field has important practical significance for improving the production efficiency of the oil recovery industry,the protection of groundwater resources,and the geological storage of greenhouse gases,etc.Traditional research methods,such as experimental research and numerical simulation,require expensive experimental equipment or a large number of computing resources and time costs.In recent years,with the continuous development and maturity of deep learning algorithms in the computer field,neural networks as a surrogate model have been widely used in various physical problems,which have the advantages of fast calculation speed and high accuracy,etc.In the thesis,a convolutional neural network is utilized to learn and predict the flow fields of the immiscible two-phase fluids in heterogeneous porous media,to achieve rapid prediction of the flow field.Also,the uncertainty quantification of the prediction is obtained to make the results more reliable and stable.The main work is briefly described as follows:1.In the first part of the work,the prediction of the flow field is achieved when the two-phase displacement in the porous medium reached the outflow boundary.The prediction of saturation fields of immiscible two-phase fluid is transformed into an image segmentation task.Using the semantic segmentation neural network model in the field of computer vision,and taking the two-dimensional random geometric structure of porous media as input,to learn and predict the saturation field and pressure field.Aiming at the characteristics of porous media with different sparsity and heterogeneity in nature,the prediction accuracy and generalization ability of the model are verified in porous media with different characteristics.To better understand the working principle of the model,the prediction process inside this black box system is explained through the visualization of the internal feature map of the neural network.2.In the second part of the work,the model is improved to realize the prediction of the dynamic two-phase displacement flow field in porous media at any time.The two methods of embedding the recurrent module and taking the time feature map as input are introduced,so that the model can predict the dynamic flow field of multiple frames at the same time.In the training process,the transfer learning method is used to load the parameters of the pre-trained model,which reduces the number of training set samples required by the current model and improves accuracy.The model only needs to be trained on the observation data at several moments and can predict the dynamic flow field at any moment.3.In the third part of the work,further improvements are made to the model and algorithm.A new Bayesian neural network model is constructed,and the uncertainty quantitative analysis of the prediction results of the flow field is obtained.Based on the Bayesian principle,the Bernoulli prior distribution of the model parameters is given through the Dropout layer.Models with posterior distribution parameters are obtained using the Monte Carlo sampling method.The uncertainty quantification of the predicted flow field is analyzed,so that the prediction results have a measure of confidence even when there is no numerical solution as a reference.