Abstract
Lithium plating in lithium ion batteries is a critical safety concern, particularly irreversible plating that leads to dendrite formation and internal short circuits. Existing monitoring methods predominantly focus on reversible lithium plating and face challenges due to insufficient real-world lithium plating data for machine learning model training. To address these limitations, we propose an unsupervised domain adaptation (UDA)-based transfer learning framework for detecting irreversible lithium plating. Our approach combines electrochemical-thermal-aging simulations and low-temperature lithium plating experiments to generate source and target domain datasets. Features extracted from discharge curves are used to train a multilayer perceptron (MLP) model under the UDA framework, enabling robust cross-domain adaptation. Results demonstrate >99% accuracy in identifying irreversible plating in simulated data and consistent qualitative alignment with experimental observations. This method provides a novel solution for online monitoring of irreversible lithium plating in energy storage systems.
1. Introduction
Lithium ion batteries dominate energy storage applications due to their high energy density and longevity. However, lithium plating—especially irreversible plating—remains a persistent safety hazard. During fast charging or low-temperature operation, lithium ions deposit on the anode surface instead of intercalating into the graphite layers. While reversible plating dissolves during discharge, irreversible plating accumulates as metallic lithium, forming dendrites that pierce separators and trigger thermal runaway. Traditional detection methods (e.g., microscopy, X-ray imaging) are offline and impractical for real-time monitoring. Machine learning (ML)-based approaches offer promise but struggle with limited labeled data and domain shifts between simulated and experimental conditions.
This work introduces a transfer learning strategy to bridge the simulation-experimentation gap. By leveraging electrochemical simulations and low-temperature aging tests, we develop a UDA-MLP model that detects irreversible lithium plating with high accuracy.
2. Methodology
2.1 Electrochemical-Thermal-Aging Model
A pseudo-two-dimensional (P2D) model coupled with thermal and aging dynamics simulates lithium plating under diverse operating conditions. Key equations include:
SEI Growth Reaction (Cathodic Tafel Equation):
iSEI=−i0,SEI⋅e−RTαcFηSEI
where i0,SEI is the exchange current density, αc is the charge transfer coefficient, and ηSEI is the overpotential.
Lithium Plating/Stripping Kinetics (Butler-Volmer Equations):
- Plating (ηpl≤0):
ipl=i0,pl(eRTαaFηpl−e−RTαcFηpl) - Stripping (ηst>0):
ist=i0,st(eRTαaFηst−e−RTαcFηst)⋅f(qcorg(cycle)qpl−qst)
Here, g(cycle) modulates reversible lithium dissolution, and f(⋅) ensures stripping does not exceed plating capacity.
2.2 Simulation Conditions
Simulations covered temperatures (−10∘C to 25∘C), charge/discharge rates (0.05C–2C), and >200 cycles per condition (Table 1).
Table 1: Simulation Parameters
| Temperature (°C) | Charge Rate (C) | Discharge Rate (C) |
|---|---|---|
| -10, -5, 0, 10, 25 | 2C | 1C |
| -10, -5, 0, 10, 25 | 1C | 0.05C |
| -10, -5, 0, 10, 25 | 2C | 0.5C |
Irreversible lithium accumulation increased exponentially below 0∘C (Figure 1).
3. Low-Temperature Lithium Plating Experiments
12 LiFePO4/graphite pouch cells (12 Ah) underwent cycling under four conditions (Table 2).
Table 2: Experimental Groups
| Group | Temperature (°C) | Charge Rate (C) | Discharge Rate (C) |
|---|---|---|---|
| 1–3 | 25 | 2C | 1C |
| 4–6 | 35 | 2C | 1C |
| 7–9 | -11 | 0.2C | 0.1C |
| 10–12 | -16 | 0.1C | 0.05C |
Post-aging capacity retention at 25∘C revealed severe degradation in low-temperature groups (Table 3).
Table 3: Capacity Retention Results
| Cell | Cycles | Capacity Retention (%) | Irreversible Plating? |
|---|---|---|---|
| 1–3 | 1000 | 72.99–85.12 | No |
| 7–9 | 83–182 | 33.22–53.41 | Yes |
4. Feature Extraction and Model Design
4.1 Discharge Curve Features
Thirteen features were extracted from discharge curves (Figure 2):
- Total discharge time (tmax)
- Discharge current rate (I)
- Voltage drop in first 10% time (ΔU)
- Midpoint voltage (U50)
- Average temperature (Tmean)
6–7. Absolute/relative voltage at point A (Ua1, Ur1)
8–9. Absolute/relative voltage at point B (Ua2, Ur2)
10–12. Relative times at A/B (tr1, tr2, Δt) - DC resistance (R)
4.2 UDA-MLP Architecture
The model comprises:
- Feature Extractor: MLP with two hidden layers (100 nodes each, ReLU activation).
- Task Classifier: Sigmoid output for binary classification (plating/no plating).
- Domain Classifier: Adversarial component to align source (simulation) and target (experiment) distributions via gradient reversal.
Loss Functions:
- Task loss (Ly): Binary cross-entropy.
- Domain loss (Ld): Binary cross-entropy.
- Total loss: L=Ly+Ld.
Gradient Reversal Layer (GRL):
∂h∂Ld→−∂h∂Ld
This forces the feature extractor to learn domain-invariant representations.
5. Results
5.1 Simulation Data Performance
The MLP achieved 99.28% accuracy, outperforming SVM (98.31%) and KNN (99.16%) (Table 4).
Table 4: Model Comparison on Simulated Data
| Model | Accuracy (%) | Precision (%) | Recall (%) |
|---|---|---|---|
| SVM | 98.31 | 96.98 | 96.98 |
| KNN | 99.16 | 99.13 | 97.84 |
| MLP | 99.28 | 98.29 | 99.14 |
5.2 Experimental Validation
The UDA-MLP correctly identified irreversible plating in low-temperature cells (Groups 7–12) with no false positives in control groups (1–6). Early-stage plating detection (<10 cycles) aligned with capacity fade trends (Figure 3).
6. Conclusion
We developed a UDA-based MLP framework for online monitoring of irreversible lithium plating in lithium ion batteries. By integrating electrochemical simulations and adversarial domain adaptation, the model achieved high accuracy (>99%) in simulated data and reliable experimental validation. Key features like ΔU, U50, and DC resistance provided robust indicators of plating severity. This approach addresses the critical need for real-time, non-destructive lithium plating detection in energy storage systems, enhancing safety and longevity.