18650 Battery Health Estimator

ISEF Category: Energy: Sustainable Materials and Design

Subcategory: Energy Storage  ·  Difficulty: Intermediate  ·  Setup: School Lab  ·  Time: 1 to 2 Months

The Hook

A battery can look fine and still be badly worn out. That matters if you build anything that depends on steady power, from a robot to a flashlight to a backup pack. With a used 18650 cell, the first few seconds of charging can give you clues about its health. Your job is to teach a model to spot those clues.

What Is It?

State of health, or SOH, tells you how much useful life a battery has left compared with a fresh one. A new 18650 cell and an old one may both charge, but they do not charge in the same way. The old cell may show a different curve at the start, like a runner who takes a different first few steps when tired.

A charging curve is a graph of how voltage, current, or another signal changes during charging. In this project, you use the first 30 seconds as a tiny fingerprint. A machine-learning model looks for patterns in that fingerprint and tries to predict battery health. Think of it like recognizing a person from the first few notes of their voice, not the whole speech.

Why This Is a Good Topic

This makes a strong science fair topic because the question is clear, measurable, and tied to a real problem. Battery grading matters for safety, reuse, and recycling, and early prediction can save time compared with full-cycle testing. You can learn data collection, feature engineering, model training, and validation without needing a university lab.

Research Questions

• How does the first 30 seconds of the charging curve differ between used 18650 cells with different state-of-health levels?
• What is the effect of using voltage-only features versus voltage, current, and slope features on model accuracy?
• Does a model trained on one batch of cells generalize to a separate batch from a different source?
• To what extent can early charging data predict battery capacity compared with full-charge measurements?
• Which machine-learning algorithm predicts 18650 state of health best from the early charging curve?
• How does adding temperature or internal resistance data change prediction quality?

Basic Materials

• Used 18650 cells from a single known source or matched set
• Battery charger with logging capability or a programmable bench power supply with data output
• Digital multimeter
• USB data logger or Arduino-compatible board
• Clamp meter or current sensor, if available
• Nonflammable battery holder or test fixture
• Fire-safe storage bag or metal container
• Laptop for data analysis
• Notebook for recording cell IDs and test conditions

Advanced Materials

• Battery cycler with logging software
• Environmental chamber or temperature-controlled test area
• Four-wire resistance measurement setup
• Precision balance
• Thermocouple or temperature probe
• Programmable electronic load
• Battery analyzer with capacity test mode
• Data acquisition system with time stamps
• Protective enclosure for cells under test

Software & Tools

• ImageJ: Convert plotted curve screenshots into digitized data if your logger exports poorly.

Experiment Steps

1. Define your target, choosing whether SOH means remaining capacity, internal resistance, or a class label such as good, fair, or poor.
2. Select the charging signal you will measure, then decide which parts of the first 30 seconds count as features.
3. Build a cell set with known labels, so your model learns from measured health values instead of guesses.
4. Plan a train-test split that keeps cells from the same battery group out of both sets, so your accuracy is honest.
5. Compare a simple baseline model with a more flexible model, then test whether early-curve features really help.
6. Design error checks for noisy cells, outliers, and measurements that may come from bad contacts or damaged cells.

Common Pitfalls

• Mixing cells with unknown histories, which makes your labels too noisy for a trustworthy model.
• Training and testing on data from the same battery group, which can inflate accuracy and hide weak generalization.
• Using raw charging curves without normalizing for current or starting voltage, which makes the model learn setup differences instead of battery health.
• Ignoring damaged or swollen cells, which can create safety risk and distort the pattern set.
• Collecting data with inconsistent contact pressure or loose leads, which adds fake spikes to the first 30 seconds.

What Makes This Competitive

A stronger project goes past a basic prediction demo. You can compare several label types, test whether your model still works on cells from a new source, and show which early features carry the most signal. Clear validation matters more than a flashy algorithm. A careful error analysis can make the project feel real and research driven.

Project Variations

• Use laptop battery cells instead of 18650 cells and compare whether the same early-charge features still predict health.
• Try a physics-based feature set, such as initial voltage rise and curve slope, instead of a larger machine-learning model.
• Compare room-temperature charging data with slightly warmer or cooler conditions to see how temperature changes prediction accuracy.

Learn More

• IEEE Xplore Abstracts: Search for recent papers on battery health estimation, then use the abstracts and references to find methods to compare.