Generative machine learning methods for the design and simulation of lithium-ion battery electrodes
File(s)
Author(s)
Kench, Steven
Type
Thesis
Abstract
The design of lithium-ion batteries is a challenging task, with myriad manufacturing processes to consider and a range of performance requirements. Consequently, new cell designs commonly require long and expensive experimental projects. This thesis develops a modular optimisation tool to aid this process, leveraging machine learning methods, parallelised microstructural metric extraction, and electro-chemical simulations.
First, SliceGAN is introduced as a novel machine learning method that uses 2D cross sectional images to generate 3D volumes. The resulting microstructural datasets can have arbitrarily large dimensions, with resolution and phase contrast that is challenging to obtain through conventional 3D imaging. Building on this approach, the algorithm is adjusted to capture relationships between manufacturing conditions and microstructural features. This allows the generation of different cathodes across the manufacturing parameter space. Crucially, our method is over two orders of magnitude faster than equivalent state-of-the-art physics based manufacturing simulations. Furthermore, it does not use geometric assumptions such as particle sphericity, and instead leverages real training data to capture the complex features of cathode microstructures.
Cell optimisation requires performance evaluation of the microstructural volumes. In this work, the pseudo 2D model is employed, with microstructurally informed homogenised transport metrics. TauFactor 2, a GPU-parallelised transport solver, is introduced as a tool to extract these metrics from microstructural volumes. When compared to five open-source solvers, TauFactor achieves a ten-fold acceleration compared to the next fastest software.
These tools are integrated into a Bayesian optimization framework, creating an end-to-end optimisation workflow. Case studies show the optimal pathways through the parameter space for different applied powers and currents. The results give interpretable trends in the relative importance of common manufacturing parameters. Crucially, this methodology quantifies these design intuitions, allowing for a systematic and data-driven approach to cell design.
First, SliceGAN is introduced as a novel machine learning method that uses 2D cross sectional images to generate 3D volumes. The resulting microstructural datasets can have arbitrarily large dimensions, with resolution and phase contrast that is challenging to obtain through conventional 3D imaging. Building on this approach, the algorithm is adjusted to capture relationships between manufacturing conditions and microstructural features. This allows the generation of different cathodes across the manufacturing parameter space. Crucially, our method is over two orders of magnitude faster than equivalent state-of-the-art physics based manufacturing simulations. Furthermore, it does not use geometric assumptions such as particle sphericity, and instead leverages real training data to capture the complex features of cathode microstructures.
Cell optimisation requires performance evaluation of the microstructural volumes. In this work, the pseudo 2D model is employed, with microstructurally informed homogenised transport metrics. TauFactor 2, a GPU-parallelised transport solver, is introduced as a tool to extract these metrics from microstructural volumes. When compared to five open-source solvers, TauFactor achieves a ten-fold acceleration compared to the next fastest software.
These tools are integrated into a Bayesian optimization framework, creating an end-to-end optimisation workflow. Case studies show the optimal pathways through the parameter space for different applied powers and currents. The results give interpretable trends in the relative importance of common manufacturing parameters. Crucially, this methodology quantifies these design intuitions, allowing for a systematic and data-driven approach to cell design.
Version
Open Access
Date Issued
2023-09-27
Date Awarded
2024-03-01
Copyright Statement
Creative Commons Attribution NonCommercial NoDerivatives Licence
Advisor
Samuel, J Cooper
Sponsor
The Faraday Institution
Grant Number
FIRG003
Publisher Department
Dyson School of Design Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)