Machine Learning Integrated Analytics of Electrode Microstructures

In the pursuit to develop safe and reliable lithium-ion batteries, it is imperative to understand all the variabilities that revolve around electrodes. Current cutting-edge physics-based simulations employ an image-based technique. This technique uses images of electrodes to extract effective properties that are used in these physics-based simulations or employ the simulation on the structure itself. Though the electrode images have spatial variability, various particle morphology, and aberrations that need to be accounted for. This work seeks out to help quantify these variabilities and pinpoint uncertainties that arise in image-based simulations by using machine learning and other data analytic techniques. First, we looked at eighteen graphite electrodes with various particle morphologies to gain a better understanding on how heterogeneity and anisotropy interplay with each other. Moreover, we wanted to see if higher anisotropic particles led to greater heterogeneity, and a higher propensity for changes in effective properties. Multiple image-based algorithms were used to extract tortuosity, conductivity, and elucidate particle shape without the need for segmentation of individual particles. What was found is highly anisotropic particles induces greater heterogeneity in the electrode images, but also tightly packed isotropic particles can do the same. These results arise from porous pathways becoming bottlenecked, resulting in greater likelihood to change values with minimal changes in particle arrangement. Next, a model was deployed to see how these anisotropies and heterogeneities impact electrochemical performance. The thought of whether particle morphology and directional dependencies would have impact on plating energy and heat generation, leading to poor electrochemical performance. By using a pseudo-2D model, we elucidated that the larger the tortuosity the greater the propensity to plate and generate heat. Throughout these studies, it became clear that the segmentation of the greyscale images became the origin for subjectiveness to appear in these studies. We sought to quantify this through machine learning techniques, which employed a Bayesian convolutional neural network. By doing so we aimed to see if image quality impacts uncertainties in our effective properties, and whether we might be able to predict this from image characteristics. Being able to predict effective property uncertainty through image quality did not prove possible, but the ability to predict physics properties based on geometric was able to be done. With the largest uncertain particles occurring at the phase boundaries, morphologies that have a large specific surface area presented with the highest structural uncertainty. Lastly, we wanted to see the impact carbon binder domain morphology uncertainty impacts our effective properties. By using a set of sixteen NMC electrodes, which specify the carbon binder domain weight percentage, we can see how uncertainties in morphology, segmentation, spatial variability, and manufacturing variability impact effective properties. We expected there to be an interplay on which uncertainty impacts various effective properties, and if manufacturing variability plays a large role in determining this. By using surrogate models and statistical methods, we show that there is an eb and flow in uncertainties and effective properties are dependent on which uncertainty is being changed.