Ageing-Aware Deep Reinforcement Learning for Adaptive Fast Charging of Li-ion Battery Considering Coupled Degradation mechanisms

Ben Shang, Yinglong He, Lei Wang, Zeyu Sun, Jie Shao, Constantina Lekakou, Jing Zhao, Youping Fan2025 · DOI

TL;DR

Researchers have developed an "ageing-aware" fast-charging strategy that uses Deep Reinforcement Learning (DRL) to balance charging speed with battery longevity. By accounting for how different degradation mechanisms—like SEI growth and lithium plating—interact over a battery's entire lifespan, the system can adaptively throttle charging currents to extend battery life without significantly sacrificing speed.

What Problem Does This Paper Address?

The "holy grail" of electric vehicle technology is fast charging that doesn't kill the battery. Traditionally, charging strategies are either passive (fixed rules like Constant Current-Constant Voltage, which ignore the battery's health) or adaptive (using models to adjust based on current state).

However, most existing adaptive models have a major blind spot: they usually focus on a single degradation mechanism in isolation. In reality, battery aging is a "coupled" process where one problem (like particle cracking) accelerates another (like SEI layer growth). Furthermore, most studies look at a single charging cycle rather than how the optimal strategy should evolve as the battery moves from "fresh" to "fully aged." This paper fills that gap by considering coupled degradation across the battery's entire lifecycle.

Methods

The authors combined advanced electrochemical modeling with machine learning:

  1. The Physics Model: They utilized the Doyle-Fuller-Newman (DFN) model, a gold standard in electrochemical modeling. They specifically integrated four coupled degradation mechanisms:
  2. SEI Layer Growth: The buildup of a film on the anode.
  3. Particle Cracking: Physical fatigue of the battery materials.
  4. Loss of Active Material (LAM): Reduction in the "stuff" that holds lithium.
  5. Lithium Plating: The dangerous formation of metallic lithium on the anode surface.
  6. Mars Charging Theory: Instead of arbitrary current steps, they used Mars theory to define the charging profile, which uses an initial current ($I_0$) and an attenuation coefficient ($K$) to describe how current should drop as the battery fills up.
  7. Deep Reinforcement Learning (DRL): They employed a Deep Deterministic Policy Gradient (DDPG) algorithm. The "agent" observes the battery's state (SOC, voltage, and degradation levels) and decides the optimal $I_0$ and $K$ to maximize a reward function that penalizes both slow charging and high degradation.
  8. Simulation Environment: The entire framework was implemented and tested using PyBAMM (Python Battery Mathematical Modelling).

Datasets

The study did not use a static external dataset. Instead, it used:

  • Cell Parameters: Physical parameters for an LG M50T cylindrical cell, sourced from existing literature (Chen et al.).
  • Synthetic Data: The DRL agent was trained and evaluated on data generated in real-time by the PyBAMM simulation environment, covering the battery's transition from 100% State of Health (SOH) down to 60% SOH.

Key Insights and Ideas

  • The Three Stages of Aging: The paper highlights that battery degradation isn't linear. It starts with rapid SEI formation, moves into a steady "moderately aged" decline, and ends with a sharp "capacity drop" caused by lithium plating and electrolyte depletion.
  • Coupled Feedback Loops: One of the most interesting "aha moments" is the description of feedback loops. For example, Loss of Lithium Inventory (LLI) actually provides negative feedback that inhibits lithium plating (because there's less lithium to plate), while Loss of Active Material (LAM) creates a positive feedback loop that accelerates particle fracture.
  • Adaptability is Key: The "best" charging current for a brand-new battery is often detrimental to an old one. The DRL agent learned to progressively slow down the charging speed as the battery aged to avoid triggering the accelerated degradation seen in the "fully aged" stage.

Key Results

  • Successful Modeling of Life Stages: The simulation accurately captured the transition from SEI-dominated aging in fresh cells to side-reaction-dominated aging in older cells.
  • Optimized Parameters: The DRL agent identified a "best combination" of initial current ($I_0$) and decay index ($K$). For instance, when the initial State of Charge (SOC) is high, the agent learned it must lower $I_0$ to achieve a sustainable $K$.
  • Aging Sensitivity: The results demonstrated that the maximum achievable charging speed ($K_{max}$) drops significantly as the battery ages. The DRL-based strategy showed "significant adaptability," automatically regulating the current more strictly as the battery reached its end-of-life stages to prevent a total capacity crash.

Limitations and Future Work

  • Thermal Omission: Due to limited data availability for the specific cell model used, the authors did not include a thermal model; the temperature was assumed to remain constant. In real-world scenarios, heat is a massive driver of degradation.
  • Simulation-Only: While PyBAMM is a highly respected tool, the results are currently limited to simulations. The authors note that experimental validation on physical battery cells is a necessary next step.
  • Computational Complexity: DFN models are computationally expensive. While the DRL agent can run quickly once trained, the initial training and the underlying physics model might require optimization for real-time deployment in a vehicle's Battery Management System (BMS).

Summary updated

BibTeX
@misc{shang2025ageing,
  author = {Ben Shang and Yinglong He and Lei Wang and Zeyu Sun and Jie Shao and Constantina Lekakou and Jing Zhao and Youping Fan},
  title = {{Ageing-Aware Deep Reinforcement Learning for Adaptive Fast Charging of Li-ion Battery Considering Coupled Degradation mechanisms}},
  year = {2025},
  doi = {10.46855/energy-proceedings-11765},
}
Figure