In WP1, we have developed a framework to learn the missing physics and unknown parameters within the models. Given that the performance of today's leading energy storage technology, lithium-ion batteries, is fundamentally limited by the properties and behaviors of graphite, we validated our models using current-voltage data of graphite electrodes. The primary focus is addressing a critical gap in understanding the internal intercalation dynamics of graphite electrodes in lithium-ion batteries — an area that remains poorly understood. Although the methodology is generalizable to other energy materials, we have specifically developed realistic chemical potential and free energy models for graphite, which is a key component for multi-phase battery models. The developed models are shown to accurately describe the key electrochemical multi-phase transition characteristics of graphite electrodes while of reasonable complexity for model development in Objectives 2.
Journal publication: [J2] referenced in the periodic report
In WP2, we have developed a multiphase model coupled with plating dynamics. The model adopted the chemical potential and free energy models as an outcome of Objective 1. We show that under battery fast charging and discharging conditions, voltage behaviour can be accurately captured. In addition, we validate the model using local variables, such as local state of charge (SOC) on different slices of graphite electrodes, obtained experimentally using operando X-ray diffraction tomography. Furthermore, the XRD data can also obtain quantitatively the local plating patterns throughout the thickness of graphite electrode. This was then used to validate models plating dynamics. The coupled electrochemical-plating model developed contains coupled partial differential-algebraic equations (PDAEs), which are difficult to solve in control applications. This makes it necessary to reduce the order of the model in Objective 3.
Journal publication: [J3] referenced in the periodic report
In WP3, we developed a framework called model-integrated neural networks (MINNs). While existing model order reduction techniques often offer speed, they tend to sacrifice accuracy, and more critically, they lack the generalizability, adaptability, and interpretability of the original model, which are essential for modeling safety-related local phenomena. MINN represents a novel architecture of physics-based learning that is capable of learning the physics-based dynamics of systems consisting of PDAEs with a control input. The developed architecture offers a systematic way to obtain optimally simplified models that are physically insightful, numerically accurate, and computationally tractable simultaneously, which can be used as a powerful tool to for integrating more advanced electrochemical models and safety-aware algorithmns in BMS.
Journal publication: [J1] referenced in the periodic report
