Prediction of Concentration Fields in Porous Media Using a Multiscale Neural Network

Pore scale modeling is expensive, in terms of time, and computing resources. Some applications require to run many CFD simulations in similar geometries which is very time consuming, especially since the information from previous simulations can't be used in new runs. On the other hand, machine learning models are an alternative to learn from existing simulation data to be able to perform fast inference in new geometries. Nevertheless, it is a well-known fact that these models do not work out of the box, and require expert knowledge of the specific application.

Our target endeavor requires to obtain locally-accurate concentration fields from a wide variety of computationally-large sphere pack arrangements under different operating conditions (Péclet and Reynolds numbers). To train our model, we generated a large dataset of CFD simulations of transport and deposition in porous media. In order to fit the porous media domains in memory, we employed a multi scale convolutional neural network (the MS-Net). However, we found out that it is relevant to provide the network with an appropriate description of the system's operation conditions in order to obtain a model that generalizes the problem accurately. We carried-out research on how different input descriptors affect the performance capabilities of our model.

Our trained model can provide instantaneous predictions, compared to around twenty hours for a single sample from the CFD workflow, and the average error on new porous media geometries and transport conditions is negligible. Our train model could be easily integrated in optimization and multiscale workflows where fast response is needed, but if accuracy is the target, it could also be used as a start point for CFD simulations, which would decrease run times significantly. The present multi scale neural network and the study of the input features can be easily transferred to any study of complex physical systems.