Session: 16-01-01: NSF-funded Research (Grad & Undergrad)

Paper Number: 77215

Start Time: Wednesday, 02:25 PM

77215 - Analysis of Transport Characteristics in Lithium-Ion Battery Porous Electrodes Based on Machine Learning 

Porous composite electrodes comprise of several constituent material phases. For e.g. lithium-ion battery (LIB) electrodes comprise of active material, conductive additive, binder and pore phase (electrolyte). An electrochemically coupled set of transport processes underlie these electrodes based on the relative microstructural arrangement of its constituent phases. Lithium-ion batteries show robust performance in a narrow temperature window, but real-world applications such as vehicular electrification expose LIBs to operational extremes. These can range from wide temperature variations to charging C-rates as high as 6C, in accordance with the US Department of Energy (DOE) goal for charge duration of ~10 minutes, at par with internal combustion engine vehicle refueling times. Such conditions can lead to several challenges associated with Li plating and thermal runaway which can severely degrade battery performance and ultimately cause battery shutdown. The composition and microstructure of electrodes dictates their kinetic and transport limitations and thereby the electrochemical performance of the battery in terms of delivered capacity, temperature rise and plating propensity. Thus, in order to design electrodes for optimal performance over a wide range of, including off-nominal operating conditions, accurate estimation of effective transport properties of battery electrodes is of primary importance.

However, porous electrode microstructure is intrinsically stochastic, anisotropic and heterogeneous, which makes the characterization of their effective transport properties very complex. Conventionally, finite volume Direct Numerical Simulation (DNS) has been used to extract electrode transport properties, but this is an extremely time-consuming and computationally expensive process. In this work, we examine the use of machine learning, specifically image-based neural networks, as a viable route for fast and accurate estimation of effective electrode properties. We use three-dimensional X-ray tomography (XCT) images of lithium-ion battery electrodes and calculate effective transport properties, namely tortuosity and electronic conductivity of Representative Volume Elements (RVEs) using finite-volume DNS. This dataset comprising of XCT images of RVEs and their ground truth (calculated) transport property values is used to train a Convolutional Neural Network (CNN). The CNN trained on in-plane property data, when given an electrode RVE image input, is able to predict the tortuosity and conductivity values in the in-plane directions with reasonable accuracy. We also note that different CNN models are required for estimation of in-plane and through-plane transport properties, since the through-plane properties are significantly different from in-plane due to calendaring in the electrode manufacturing process. The proposed computational framework is not just restricted to battery electrodes but can be extended to characterize various modes of transport resistances in porous media in general.    

Presenting Author: Debanjali Chatterjee Purdue University

Authors:

Debanjali Chatterjee Purdue University
Bairav S. Vishnugopi Purdue University
Partha P. Mukherjee Purdue University