Electrochemical Batteries find wide and diverse applications, ranging from wearable devices to electric cars. While there exist several relatively accurate models of the processes underpinning their charge and discharge phases, the ageing effect affecting such devices after repeated cycles is much less understood. This results in significant discrepancies between physics-based models and real observations of voltage discharge curves, i.e. when the underlying ageing state is unknown. In this talk, I will present a solution to this problem based on deep learning. In particular, we propose a Transformer-based architecture which, given only the initial part of the discharge curve and the input current, is able to simultaneously infer the ageing state and predict the full voltage curve until discharge with high precision. Akin to Vision Transformers, in order to allow our model to handle possibly very long time series, we split the original sequence into smaller sub-chunks, that we treat as individual tokens. Besides providing excellent performance on simulated data, we show that a minimal amount of fine-tuning allows the model to fill the sim2real gap between simulations and data gathered from a set of real batteries.

Download the slides for this talk.Download ( PDF, 723.26 MB)