Mesoscale Machine Learning Analytics for Electrode Property Estimation

Abstract

The development of next-generation batteries with high areal and volumetric energy density requires the use of high active material mass loading electrodes. This typically reduces the power density, but the push for rapid charging has propelled innovation in microstructure design for improved transport and electrochemical conversion efficiency. This requires accurate effective electrode property estimation, such as tortuosity, electronic conductivity, and interfacial area. Obtaining this information solely from experiments and 3D mesoscale simulations is time-consuming while empirical relations are limited to simplified microstructure geometry. In this work, we propose an alternate route for rapid characterization of electrode microstructural effective properties using machine learning (ML). Using the Li-ion battery graphite anode electrode as an exemplar system, we generate a comprehensive data set of ∼17 000 electrode microstructures. These consist of various shapes, sizes, orientations, and chemical compositions, and characterize their effective properties using 3D mesoscale simulations. A low dimensional representation of each microstructure is achieved by calculating a set of comprehensive physical descriptors and eliminating redundant features. The mesoscale ML analytics based on porous electrode microstructural characteristics achieves prediction accuracy of more than 90% for effective property estimation.

Band Gaps and Optical Properties of RENiO3 upon Strain: Combining First-Principles Calculations and Machine Learning

“…Machine learning was adopted to build the models that predict the band gaps (E g ), optical absorption coefficient (α(ω)), and reflectivity (R(ω)) [24][25][26][27]. For the modeling of the band gaps, many intrinsic parameters, including planar strain (S), relative atomic mass (M), electronegativity (X), volume of supercells (V), nickel ion radius (R N ), oxygen ion radius (R O ), rare earth element ion radius (R R ), and tolerance factor (t) [28] were taken as initial feature descriptors.…”