Learning the P2D Model for Lithium-Ion Batteries with SOH Detection
Maricela Best McKay, Bhushan Gopaluni, Brian Wetton · 2025
Read Learning the P2D Model for Lithium-Ion Batteries with SOH Detection by Maricela Best McKay, Bhushan Gopaluni, Brian Wetton (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-14147.txt.
Introduction
Maricela Best McKay, Bhushan Gopaluni, and Brian Wetton replace a heavy battery model with a fast learned stand-in, in a 2025 paper. Lithium-ion batteries need management systems that predict how they will perform and age. The trusted physics model, called P2D, tracks chemistry in fine detail but runs slowly at high currents.
The team trained a convolutional neural network on data simulated from the physics model across random driving cycles. The network learned to reproduce lithium concentration profiles across the battery with high accuracy. They also show the stand-in can be adjusted to match a battery's state of health as it wears. A fast, faithful copy of the physics model could let control systems plan charging in real time. The approach keeps physics accuracy while shedding physics cost.
Catalog record
| CREATOR | Maricela Best McKay, Bhushan Gopaluni, Brian Wetton |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 26 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-14147
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Learning the P2D Model for Lithium-Ion Batteries with SOH Detection Maricela Best McKay, Bhushan Gopaluni, Brian Wetton Abstract Lithium ion batteries are widely used in many applications. Battery management systems control their optimal use and charging and predict when the battery will cease to deliver the required output on a planned duty or driving cycle. Such systems use a simulation of a mathematical model of battery performance. These models can be electrochemical or data-driven. Electrochemical models for batteries running at high currents are mathematically and computationally complex. In this work, we show that a well-regarded electrochemical model, the Pseudo Two Dimensional (P2D) model, can be replaced by a computationally efficient Convolutional Neural Network (CNN) surrogate model fit to accurately simulated data from a class of random driving cycles. We demonstrate that a CNN is an ideal choice for accurately capturing Lithium ion concentration profiles. Additionally, we show how the neural network model can be adjusted to correspond to battery changes in State of Health (SOH). 1 Introduction Rechargeable Lithium Ion Batteries (LIB) are ubiquitous in portable electronics, electric cars, and small scale power storage [ 24 , 15 , 4 , 3 ] . Battery Management Systems [ 20 , 1 , 16 ] (BMS) control the charge and discharge of LIB systems, and estimate the State of Charge (SOC) of a system. Sophisticated BMS also adjust to the battery State of Health (SOH) [ 21 , 7 , 27 , 2 , 19 ] , how battery performance changes with time and use. All BMS rely on models of battery performance. These models can be simple ones, based on equivalent circuit approximations of LIB operation [ 10 ] . These are suitable for low current (low C-rate) applications. For the reader unfamiliar with battery units, a “C” is the current which will discharge a fully charged battery in an hour. It is well known that LIB operating at high currents need more complicated electrochemical models to accurately predict performance [ 11 ] . Here, the intercallation rate of Lithium ions in and out of electrode particles can be current limiting. Ionic currents in the electrolyte can also be rate limiting. If both these effects are important, then the well-known Pseudo Two Dimensional (P2D) electrochemical model is needed to predict performance. The P2D model is also known as the Doyle-Fuller-Newman (DFN) model after the researchers that first proposed it [ 5 ] . The models above can be classified as physical models, with parameters that can be fit from limited experiments and then extrapolated to other operating conditions. Another option for LIB models is modern Deep Learning (DL) [ 26 , 22 , 6 ] , which relies on large quantities of observed operational data. To our knowledge, no BMS currently in operation uses the P2D model for real time prediction and control. This model, which involves several coupled nonlinear Partial Differential Equations (PDE), is just too complex to be solved in real time [ 25 , 8 ] . In this article, we show that a (DL) surrogate model can give accurate predictions to a P2D model when trained appropriately on synthetic data from a computational simulation. A key aspect of this work is the use of a CNN model, whose structural features are particularly well suited to capturing Li-ion battery concentration data, enabling the model to achieve high accuracy. This work gives insight into the type of training data necessary to accurately capture high current operation and gives a computationally efficient simulation tool for a P2D model (fit to a particular LIB type) that can be used in a BMS. Because we employ a deep learning framework to replace the P2D model, operational data can be easily added to improve the accuracy of the P2D model fit. This is showcased in 5 where we demonstrate how to incorporate a SOH parameter into the surrogate modelling framework. This work is a continuation of an earlier project described in [ 17 ] . In this paper, we extend the earlier work from the simple Single Particle Dynamics (SPD) electrochemical model to the full P2D model and employ a CNN rather than a simple, fully connected Neural Network. We show simulations in our driving cycle framework where the P2D model is significantly more accurate. We also show that a mechanism to identify SOH parameters can be incorporated into the surrogate modelling framework. 2 below describes the P2D model and its simulation that gives the synthetic data for training the surrogate model. We use the PyBaMM [ 23 ] library developed at Oxford for the simulations. We also show the driving cycle we simulate and train to, and show the electrochemical effects that must be considered when operating at high current. We model discharge only, but a similar procedure can be used for charging. In Section 3 we describe the DL framework we use and details about the training with results in 4 . It is shown here that constant discharge data is not sufficient to adequately train the neural network for our high current, variable current driving cycle. The extensions to SOH identification is done in Section 5 . We summarize our findings and discuss some promising ideas for future research in a final section. 2 P2D Model and Driving Cycle We train a Deep Neural Network (DNN) on data generated by a P2D model simulation, implemented with the PyBaMM [ 23 ] library v22.12 with default parameters. A detailed visual and mathematical description of the P2D model can be found in [ 8 ] . The ensemble of conditions to be modelled involves a battery charged to full capacity, then discharged through a random driving cycle of prescribed currents until it fails to operate at the desired current. The DNN described in the next section is trained on data sampled from the simulation every 100 seconds. The model is trained with predicting the battery voltage and the intercallated Lithium concentrations in electrode particles on a grid in x (scaled position through the electrodes) and r (scaled electrode particle radius). The particle concentrations are internal quantities that are not directly measurable in operation and can only be obtained through the use of electrochemical models. They are necessary to understand and predict the battery behaviour in time dependent high current operation. These model outputs can be combined, due to the deep learning framework we employ, with real operational voltage data to retain the insights from the approximation of these effects from the model as well as correcting for the mismatch between the model and the actual battery. In the current work, we train to the simulation results only. The driving cycle consists of choosing random discharge currents every 100 seconds uniformly between zero and 6C. A piecewise linear current profile I(t) is constructed from these values. Batteries are built quite differently for different applications, but generally 1C operation can be modelled well with simple (equivalent circuit) models. It is at higher currents, such as those we consider here, that electrochemical models are needed for accurate voltage predictions. For reference, the maximum constant current that can be drawn over 100s from a fully charged cell modelled by the simulation is 7C (the cell