Physics-based electrochemical models are useful tools for optimizing battery cell and material design, managing battery use, and understanding physical phenomena, all of which are key in enabling adoption of batteries to electrify transportation, grid storage, and other high carbon emission industries. Fitting these models to experiments can be a useful approach to determine missing parameters that may be difficult to identify experimentally. In this dissertation, two use cases of this approach — model parameterization and degradation tracking — are explored.
An introduction to the need for batteries and an overview of challenges in the field is presented in Chapter 1. Of these challenges, those that can be addressed by battery modeling solutions are discussed in further detail. An overview of continuum level physics-based electrochemical models is provided, and the case is made for the utility of parameter estimation. In Chapter 2, an extension of a published model for lithium trivandate cathodes for lithiumion batteries is outlined. While the original model described (de)lithiation and phase change in the cathode, the new model describes simultaneous lithiation of the original phase, lithiation of the newly formed phase, and phase change. Parameters associated with the thermodynamics and kinetics of charge transfer and lithium transport in the second phase are estimated directly from experimental data. This study serves as an example of using the model fitting approach to determine model parameters that would be difficult to isolate and measure experimentally.
Chapter 3 explores a similar concept of model parameterization, this time focusing on the electrode tortuosity. Tortuosity is a hard to quantify parameter that describes how tortuous of a path lithium ions must travel through an electrode or separator. Because there are several experimental measurement techniques suggested in the literature that do not always provide consistent results, an approach to fit the tortuosity to a standard rate capability experiment is introduced. The Bayesian approach returns uncertainties in tortuosity estimates, which can be used to predict a range of outcomes for high-rate performance. Covariance between parameters in the model and their impact on uncertainties in tortuosity is also discussed.
Beyond model parameterization, parameter estimation can also be useful in the context of tracking degradation by fitting a physics-based model over the course of cycling and interpreting the evolution of the parameter estimates. In Chapter 4, this idea is explored by fitting the model developed in Chapter 2 to cycling of an LVO cell. Parameter estimates are interpreted in conjunction with traditional tear down and electrochemical analysis to identify root causes of degradation for this cell.
Depending on the number of parameters being simultaneously estimated, it can become an onerous task to fit model parameters, especially if the physics-based model cannot easily be enclosed in an efficient optimization algorithm. To this end, machine learning (ML) can be useful. If a ML model is trained offline on synthetic data generated by a battery model to map the observable electrochemical data to parameters in the battery model, the ML model can be deployed to estimate parameters from experiment. These models can be referred to as inverse ML models, since they perform the inverse task of a "forward" physics based model.
The procedure described above is implemented in Chapter 5. Interpretable ML models are trained on published synthetic data generated by equivalent circuit models. Pseudo-OCV (slow charge, C/25) full cell voltage curves are passed into the inverse ML models to estimate degradation modes in lithium ion batteries and classify which electrode limits cell capacity. These models are useful in diagnosing the state of the battery at any given time. Accuracies of the inverse ML models are evaluated on independent test sets also composed of synthetic data and are published to benchmark future diagnostic studies. The insights derived from the trained ML models in terms of which features in the full cell voltage curves are predictive of the degradation modes are compared to expert insights.
In chapter 6, the robustness of the inverse ML approach towards model-experiment disagreement is probed. If the experiment does not directly map onto the protocol used to generate the synthetic training data for the ML model, or if the model itself is inherently a poor descriptor of experiment, the inverse ML model will inevitably return inaccurate estimates. In this chapter, a feed forward neural network (NN) is employed as the inverse ML model. In two case studies of model-experiment disagreement, the NN returns biased parameter estimates. A simple data augmentation procedure is introduced to mitigate these biases.
Chapter 7 ties together the understanding developed in the previous chapters by applying more robust neural networks to estimate parameters for LVO cells cycled at different rates. This study demonstrates how to interpret parameter estimates in conjunction with cycling data to gain mechanistic insight into degradation. A complex map of coupled degradation hypotheses is reduced to a smaller subset of possible mechanisms for two exemplary LVO cells, and parameter estimates for a larger set of LVO cells are discussed. The framework presented in this study synergistically combines experiment, physics-based modeling, and machine learning to better understand degradation phenomena.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/nz6d-z243 |
| Date | January 2022 |
| Creators | Mayilvahanan, Karthik |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0029 seconds