Harnessing AI: Accurate Predictions of Battery Capacity Fade and Battery Cycle Life

Predicting battery degradation is pivotal for advancing material research and unlocking opportunities in battery design, utilization, testing, and recovery. This enables better battery management and ensures the maintenance of desired performance characteristics over an extended period including both up and downstream manufacturing processes such as material selection, design qualification, technology/cell selection (battery management), and warranty, etc. However, traditional models and machine learning techniques often fall short in capturing these characteristics. <br/><br/>Lithium-ion cells typically show a gradual capacity degradation up to a certain point, known as the knee-point, beyond which the degradation accelerates rapidly, leading to the cell's End-of-Life. SandboxAQ’s collaboration with NOVONIX has resulted in a robust method to predict long-term cell capacity fade from early-life cycling data. The approach begins with extracting early-life current, voltage, and capacity data collected on NOVONIX's world-class Ultra-High Precision Coulometry (UHPC) cycler systems. Electrochemically relevant metrics are then combined with advanced feature engineering and sophisticated data transformations are employed to construct features conducive for machine learning.<br/><br/>Proprietary machine learning and deep learning models were trained on several comprehensive, but limited datasets. Despite limited training data, these models demonstrate a remarkable ability to accurately classify the cells with and without catastrophic failure at a given point in time, predict the cycle number at which 80% of cell’s initial capacity remains, and estimate the remaining useful life of the cell, which provides key indicators of cell health and longevity. <br/><br/>One of the standout features of these models is their flexibility. They are designed to seamlessly incorporate new data as it becomes available, ensuring that the predictions remain relevant and accurate over time. This approach not only provides a more cost and time-efficient alternative to long-term experiments traditionally used to estimate battery life but also opens avenues for more dynamic and responsive battery management strategies.<br/><br/>In summary, our solution leverages AI to rapidly make accurate, physics-based, end-of-life predictions using early-life electrochemical data. We leveraged comprehensive manufacturing cell data to deliver custom machine learning and deep learning models that were able to reliably predict cycle life only from high fidelity early-life data. The models are designed to easily integrate any future data, offering a cheap and efficient alternative to lengthy benchtop experiments.