Date and Time:
January 24, 2018 - 12:30pm
Green Earth Sciences Building 365
*Refreshments at 12:15 pm
Using Convolutional Neural Network to Calculate Effective Properties of Porous Electrode
Speaker: Dr. Slava Korneev
Abstract: It was shown recently that the macro-scale transport of ions in Li-Ion batteries (LIBs) are highly affected by the inhomogeneity of the micro-scale structure of porous electrodes. The homogenization through multiple scale expansions provides the most accurate estimation of the macro-scale dispersion tensor for a highly heterogeneous domain. The dispersion tensor couples the diffusive dynamic across multiple scales through solution of the "closure problem" in the form of Poisson's equation with Neumann boundary condition. The homogenization requires an accurate imaging of the 3D pore-scale geometry. Microscope techniques are successfully used for imaging of the electrodes structure with a high resolution. Detailed reviews of modern imaging techniques in LIBs research exist. Since microscopes are inherently orthographic devices, reconstruction of a 3D digital porous structure where the "closure problem" can be solved requires identification of phases (e.g. void, solid) and their 3D depth. While a human brain perceives depth remarkably well given just one image, it is an extremely challenging problem for computer vision. Reconstruction of a semi 3D scene structure from a single 2D image was recently done using methods from supervised machine learning.
We train a convolutional neural network to estimate the dispersion tensor from a single 2D microscopic image. The network is designed to input the image of the pore-scale structure and output the macro-scale diffusion tensor. For the training set, we generate a large number of the 2D synthetic porous medium by randomly varying grains geometry and the total number of grains. Then, the dispersion tensor of the training sample is defined by solving numerically the "closure problem". Trained network shows good accuracy in estimation of the dispersion tensor.