(375o) Modelling Local Steady-State and Time-Dependent Reactive Dynamics in Porous Media By Multiscale Neural Networks

Porous media systems are relevant in many research fields of chemical engineering: packed bed catalytic reactors, filters, and batteries. The microscale modelling, i.e. pore-scale, of a representative volume of the porous structure is a state-of-the-art methodology to obtain the accurate evaluation of transport related properties, such as reaction rates or filtration efficiencies. Computational fluid dynamics (CFD) has been widely employed to this end. Nevertheless, these microscale simulations are computationally expensive. Thus, these models can hardly be integrated in multiscale modelling, or optimization workflows, where fast response models are needed. Machine learning, and in particular deep learning techniques, can be employed to train data driven models as surrogates of CFD simulations. To this end, convolutional neural networks are appropriate models for porous media applications since they allow as inputs images of the porous structure, in this way the choice of most relevant geometrical features is performed by the network itself.

In this work we train multiscale convolutional neural networks (MSNet) [1] as surrogate models for the local prediction of fields in porous media, two case studies are presented: the steady state prediction of concentration fields in filters, and the transient prediction of the discharge dynamics in the cathode side of lithium-ion batteries. The workflow we propose requires the creation of a dataset of CFD based simulations, which is then employed to train the neural networks and to test their generalization capabilities.

For the filtration application, the CFD simulations have been solved by the finite volume method implemented in OpenFOAM, and a wide range of geometries (as sphere packings) and operating conditions (Reynolds and Péclet numbers) have been explored. The different geometries and operating conditions are the input features for MSNet which is trained to predict the concentration fields. As displayed in Fig. 1, the trained MSNet is able to accurately predict the fields with errors on the average concentration lower than 5% [2].

For the lithium-ion batteries case study, the discharge simulations of the half cell (cathode side) have been solved by using the finite elements method implemented in COMSOL, even in this case different cathode geometries and different discharge conditions have been explored for the creation of the dataset [3]. In order to deal with the transient nature of the dataset, MSNet has been modified with an autoregressive approach. The resulting model is able to predict the lithium concentration and the potential dynamics in the solid phase, thus the discharge curves of the cells, Fig. 2.

[1] Santos, Javier E., et al. Transport in porous media 140.1 (2021): 241-272.

[2] Marcato, Agnese, et al. Chemical Engineering Journal (2022): 140367.

[3] Liu, Chaoyue, et al. Energy Storage Materials 54 (2023): 156-163.