⚡ electrical engineering

Machine Learning Detection of Lithium Plating in Lithium-ion Cells: A Gaussian Process Approach

This paper proposes a Gaussian Process framework that analytically models the charge-voltage relationship to robustly detect lithium plating in lithium-ion cells by inferring noise-aware differential capacity features with quantified uncertainty, offering a scalable solution for real-time battery management systems.

Original authors: Ayush Patnaik, Jackson Fogelquist, Adam B Zufall, Yiwei Ji, Stephen K Robinson, Peng Bai, Xinfan Lin

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ayush Patnaik, Jackson Fogelquist, Adam B Zufall, Yiwei Ji, Stephen K Robinson, Peng Bai, Xinfan Lin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Silent Killer" in Your Battery

Imagine your electric car battery is a busy highway. When you charge it, lithium ions (the cars) drive onto the road and park in their designated spots (the battery's internal structure). This is normal and healthy.

However, if you charge the battery too fast or if it's too cold, the traffic gets jammed. Some of the lithium "cars" can't find a parking spot. Instead of parking, they pile up on the side of the road, forming a metallic puddle. This is called Lithium Plating.

This is dangerous because:

It steals energy (your battery dies faster).
It can cause short circuits, leading to fires or explosions.

The big challenge is that by the time you notice the battery is dying, it's often too late. We need a way to spot this "traffic jam" while it's happening, before the cars pile up.

The Old Way: Trying to Hear a Whisper in a Storm

Scientists have tried to detect plating by looking at the battery's voltage (its electrical pressure). They use a math trick called Incremental Capacity Analysis (looking at how much charge enters for every tiny bit of voltage change).

Think of this like trying to hear a specific bird chirp (the plating signal) while standing next to a roaring jet engine (sensor noise).

The Old Method: To hear the bird, they tried to "smooth out" the noise using filters. But this is like putting on heavy earmuffs; you might block the jet engine, but you also muffle the bird. Plus, the math they used to calculate the signal (finite differencing) is like trying to measure speed by looking at a blurry photo; it creates a lot of "static" and errors, making it hard to know if the signal is real or just noise.

The New Solution: The "Crystal Ball" (Gaussian Processes)

The authors of this paper propose a new method using Gaussian Processes (GP).

Imagine you are trying to draw a smooth, perfect curve through a bunch of scattered, messy dots on a piece of paper.

The Old Way: You connect the dots with a jagged, shaky line. If you try to measure the slope of that shaky line, your ruler slips, and your measurement is wrong.
The GP Way: You use a "smart rubber band" (the Gaussian Process). This rubber band knows that the underlying reality (the battery's chemistry) is smooth and continuous. It stretches across the messy dots, ignoring the tiny wiggles caused by sensor errors, and draws the most likely smooth curve.

Why is this better?

It knows the noise: The rubber band learns how "jittery" the sensors are and automatically ignores the tiny, meaningless wiggles.
It calculates the slope perfectly: Because the rubber band is a smooth mathematical function, we can calculate its slope (the derivative) perfectly without the math breaking down.
It gives a confidence score: Most importantly, the rubber band doesn't just give you a line; it gives you a shaded zone around the line. It says, "I am 95% sure the curve is inside this shaded area." This tells us if a signal is a real "bird chirp" or just random static.

The Discovery: The "Secondary Peak"

When the researchers used this "smart rubber band" to look at the battery's voltage curve, they found a specific signature of lithium plating: A secondary peak (a small bump) on the graph that appears above 4.0 Volts.

No Plating: The graph looks like a smooth hill.
Plating: A new, distinct little hill (a shoulder) appears on the side of the main hill.

The Proof: Did it Work?

The team tested this on real batteries under different conditions:

Cold and Fast: (The danger zone). The GP method saw the "secondary peak" and correctly said, "Plating is happening!"
Warm and Slow: The GP method saw no peak and correctly said, "All clear."
The Tricky Cases:
- One battery was charged at a moderate speed in the cold. It lost a lot of capacity, but the GP method saw no peak. When they opened the battery later, they found no metallic lithium, just chemical gunk from the electrolyte. The GP method was right: it wasn't plating; it was something else.
- Another battery was charged very slowly in the cold. The GP method saw a tiny peak. Even though the battery wasn't dead yet, the method correctly flagged it as the very beginning of plating.

The Takeaway

This paper introduces a "smart, noise-canceling microscope" for battery data. Instead of guessing or getting confused by sensor static, this method uses probability to draw a clean picture of what's happening inside the battery.

It can spot the first tiny signs of lithium plating (the "traffic jam") before the battery is damaged, giving us a chance to stop charging and save the battery. This is a huge step toward making electric vehicles and satellites safer and longer-lasting.

1. Problem Statement

Lithium plating is a critical degradation mechanism in lithium-ion batteries, particularly during fast charging and low-temperature operation. It occurs when the anode potential drops below 0 V vs. Li/Li⁺, causing metallic lithium to deposit on the graphite surface rather than intercalating. This leads to rapid capacity fade and poses severe safety risks (e.g., thermal runaway).

Current Challenges:

Detection Limitations: Traditional detection methods (post-mortem analysis, NMR, XRD) are destructive, intrusive, or require expensive lab equipment, making them unsuitable for real-time Battery Management Systems (BMS).
Voltage-Based Diagnostics: While incremental capacity analysis ($dQ/dV$) can reveal plating via a secondary high-voltage peak (>4.0 V), calculating this derivative from noisy sensor data is problematic.
Noise Amplification: Conventional methods rely on finite differencing combined with ad-hoc filtering (e.g., Savitzky-Golay). This approach amplifies sensor noise, introduces bias in feature location, and requires manual parameter tuning, making robust real-time detection difficult.

2. Methodology: Gaussian Process (GP) Framework

The authors propose a probabilistic framework that models the charge-voltage relationship $Q(V)$ directly as a Gaussian Process (GP) rather than using numerical differentiation.

Core Principles:

Stochastic Modeling: The accumulated charge $Q$ is modeled as a function of voltage $V$ ( $y = f(x) + \epsilon$ ), where $f(x)$ is a GP and $\epsilon$ represents measurement noise.
Analytical Derivatives: A key property of GPs is that their derivatives are also GPs and are jointly distributed with the original process. This allows for the analytical and probabilistic inference of $dQ/dV$ directly from the posterior distribution, bypassing unstable numerical differencing.
Kernel Selection: The authors use a squared-exponential (RBF) kernel, which ensures the function is mean-square differentiable ( $C^2$ regularity), guaranteeing the existence of the derivative process.

Implementation Steps:

Training: Hyperparameters (length scale $\ell$ , signal variance $\sigma_f$ , noise variance $\sigma_n$ ) are learned via maximum likelihood estimation on a single reference charge cycle.
Inference: For subsequent cycles, the GP is conditioned on the observed $(V, Q)$ data pairs. The posterior mean provides the smooth $dQ/dV$ curve, while the posterior variance provides credible intervals (uncertainty quantification).
Feature Detection: The algorithm searches for a distinct secondary peak above 4.0 V in the inferred $dQ/dV$ curve. If the peak is present and the uncertainty is low, plating is detected.

3. Key Contributions

Noise-Robust Differentiation: The method eliminates the need for manual filtering and finite differencing, providing a principled way to handle sensor noise.
Uncertainty Quantification: Unlike point-estimate methods, the GP framework provides credible intervals for the $dQ/dV$ curve. This allows the system to distinguish between a genuine plating peak and noise-induced fluctuations with high confidence.
Scalability: The framework is designed to be scalable to online variants (e.g., sparse or state-space GPs), making it suitable for embedded BMS in EVs and aerospace applications.
Specificity: The method successfully distinguishes lithium plating from other degradation mechanisms (like electrolyte decomposition) that cause capacity fade but do not produce the specific high-voltage secondary peak.

4. Experimental Results

The method was validated on EEMB LIR-2032 coin cells (Graphite||LiCoO₂) across various C-rates (0.2C–1C) and temperatures (0°C–40°C).

Key Findings:

Plating Detection: The GP-based method reliably identified a secondary peak above 4.0 V in conditions known to induce plating (e.g., 0.4C at 10°C, 0.6C at 0°C).
False Negative Avoidance: It correctly identified no secondary peak in non-plating scenarios (e.g., 1C at 40°C), even when capacity fade was present due to other mechanisms.
Early Warning Capability: The method detected plating onset in "mild" conditions (e.g., 0.2C at 0°C and 1C at 25°C) where bulk degradation metrics (capacity loss) were still negligible. This demonstrates superior sensitivity compared to traditional throughput-based monitoring.
Validation against Ground Truth:
- Post-mortem Microscopy: Cells identified as "plating-positive" showed metallic lithium deposits; "plating-negative" cells showed no metallic lithium (only electrolyte decomposition products in some cases).
- Capacity Fade: Plating-positive cases generally correlated with higher rates of charge throughput loss and capacity fade, though the GP method detected the onset before significant bulk degradation occurred.
Confidence Levels: The credible intervals around detected peaks were exceptionally tight (within ~1.9% of peak magnitude), providing high confidence in the detection.

5. Significance and Impact

Real-Time Safety: This approach enables real-time detection of lithium plating during the charging process, allowing BMS to terminate charging immediately upon detecting the onset of plating, thereby preventing catastrophic failure.
Operational Flexibility: It challenges the simplistic "rate vs. temperature" heuristics for plating. The study shows that plating depends on the specific electrochemical state (anode potential) rather than just C-rate and temperature, and the GP method captures these complex interactions.
Practical Deployment: By avoiding complex hardware and destructive testing, and by providing a mathematically rigorous, noise-tolerant algorithm, this work establishes a practical pathway for integrating advanced plating detection into commercial and aerospace battery management systems.

In conclusion, the paper demonstrates that Gaussian Processes offer a superior alternative to traditional signal processing for battery diagnostics, providing a robust, uncertainty-aware, and early-warning system for one of the most critical failure modes in lithium-ion batteries.