Advanced characterisation of lithium-ion batteries using machine learning and physics-based simulation
File(s)
Author(s)
Squires, Isaac
Type
Thesis
Abstract
Lithium-ion batteries are essential for addressing climate change. Developing cheaper, higher-performing, and longer-lasting batteries remains crucial for supporting an electrified future. Characterisation of electrode materials and electrochemical modelling are vital for understanding the relationship between manufacturing processes, structure, and battery function. This thesis applies novel machine learning approaches to overcome limitations in existing characterisation techniques, explores some of the assumptions of reduced-order modelling, and presents a pathway to accelerate 3D modelling.
This work introduces an innovative tool that uses artificial intelligence to inpaint defects in micrographs, enabling the use of larger domains for simulation. Applied to a large dataset of images, this technique improved data quality by inpainting scale-bars and increased the dataset by 20%, enhancing the accuracy of subsequent machine learning models for battery research.
Reduced-order electrochemical models are computationally efficient for simulating battery behaviour. The widely used Doyle-Fuller-Newman model assumes a separation of scales between micro- and macroscale features. However, in many real battery electrodes, this separation is limited. This thesis compares 3D and 1D simulations across varying scales, revealing conditions where traditional modelling may fail and highlighting the need for nuanced approaches in scenarios with minimal scale separation or high operational stress.
Reduced-order models may not fully capture 3D battery dynamics. This thesis explores an approach to accelerating 3D modelling of voxel-based image data using GPU architectures and machine learning libraries for rapid matrix multiplication. The development and validation of this framework demonstrate its potential to outperform existing numerical approaches in capturing complex electrode behaviour, particularly for intricate geometries. This initial framework lays the groundwork for future advancements, which could significantly speed up 3D simulations of Li-ion battery electrochemistry.
This work introduces an innovative tool that uses artificial intelligence to inpaint defects in micrographs, enabling the use of larger domains for simulation. Applied to a large dataset of images, this technique improved data quality by inpainting scale-bars and increased the dataset by 20%, enhancing the accuracy of subsequent machine learning models for battery research.
Reduced-order electrochemical models are computationally efficient for simulating battery behaviour. The widely used Doyle-Fuller-Newman model assumes a separation of scales between micro- and macroscale features. However, in many real battery electrodes, this separation is limited. This thesis compares 3D and 1D simulations across varying scales, revealing conditions where traditional modelling may fail and highlighting the need for nuanced approaches in scenarios with minimal scale separation or high operational stress.
Reduced-order models may not fully capture 3D battery dynamics. This thesis explores an approach to accelerating 3D modelling of voxel-based image data using GPU architectures and machine learning libraries for rapid matrix multiplication. The development and validation of this framework demonstrate its potential to outperform existing numerical approaches in capturing complex electrode behaviour, particularly for intricate geometries. This initial framework lays the groundwork for future advancements, which could significantly speed up 3D simulations of Li-ion battery electrochemistry.
Version
Open Access
Date Issued
2024-05
Date Awarded
2024-09
Copyright Statement
Creative Commons Attribution NonCommercial NoDerivatives Licence
Advisor
Cooper, Samuel
Publisher Department
Dyson School of Design Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)