docs/research_notes/01_eda_findings.md

EDA Research Notes — Notebook 01

Date: 2025-07-15
Researcher: AI Battery Lifecycle Research Pipeline
Dataset: NASA PCoE Li-ion Battery Dataset (cleaned)

---

1. Dataset Overview

Metric	Value
Total batteries (after exclusions)	30
Excluded batteries	B0049–B0052 (corrupt/incomplete)
Total cycle records	7,317
Discharge cycles	2,678
Charge cycles	2,715
Impedance measurements	1,908
Nominal capacity	2.0 Ah (18650 Li-ion)
Temperature groups	5: {4, 22, 24, 43, 44} °C
Mean cycles per battery	89
Max cycles (any battery)	196
Batteries reaching EOL	22 / 30 (73.3%)

2. Critical Discovery: Temperature Groups

Original documentation claimed 3 temperature groups (4°C, 24°C, 43°C).
Actual data contains 5 distinct groups:

Temperature (°C)	# Batteries	Category
4	12	Cold
22	3	Near-ambient
24	14	Room temperature
43	4	Elevated
44	3	Elevated

Note: Some batteries appear in multiple temperature groups (tested at different conditions).
The 22°C and 44°C groups have never been separately analyzed in published literature using this dataset — this represents a potential novel contribution.

3. Capacity Fade Analysis

3.1 Overall

• Capacity range: [0.0441, 2.4441] Ah
• SOH range: [2.2%, 122.2%]
• Values >100% SOH indicate initial measured capacity above the 2.0 Ah rated nominal — common in fresh Li-ion cells
• Clear exponential-like decay visible across all batteries, with some showing abrupt drops (regeneration artifacts)

3.2 Temperature Effects (Key Finding)

Temperature	Mean Capacity (Ah)	Std Dev	Observations
4°C (Cold)	0.91	0.46	54.5% capacity reduction vs nominal — severe cold degradation
22°C	~1.50	~0.25	Near-room performance
24°C (Room)	1.54	0.26	Widest distribution (most batteries)
43°C (Elevated)	1.72	0.07	Narrowest distribution — rapid degradation with fewer cycles
44°C	~1.65	~0.15	Similar to 43°C

Research Implication: Cold-temperature operation (4°C) is more damaging to Li-ion cycle life than elevated temperature (43°C) in terms of absolute capacity. However, elevated-temperature batteries have fewer cycles before EOL.

3.3 SOH Distribution by Temperature

• 4°C: Extremely wide, flat KDE centered at ~55% — batteries spend most of their life in degraded state
• 22°C: Tight peak at ~82%
• 24°C: Bimodal — some batteries healthy (90%), others significantly degraded (65%)
• 43°C: Very tight KDE peak at ~85% — consistent but short-lived
• 44°C: Peak at ~90%, narrow

4. Impedance Analysis

• 1,908 impedance measurement records available
• Electrolyte resistance (Re): Clear upward trend with cycle number — SEI layer growth
• Charge transfer resistance (Rct): More dramatic increase, with outlier spikes (notably B0034)
• Phase Space (Re vs Rct): Distinct battery-group clusters visible, with drift from lower-left (healthy) to upper-right (degraded) confirming dual-resistance aging signature

5. Voltage Surface Analysis

• 3D discharge voltage surface for B0005 shows:
  • Progressive "sinking" of voltage plateau with aging
  • Voltage knee region shifts earlier (less capacity delivered)
  • End-of-discharge voltage drops become sharper in later cycles
  • Surface is smooth except for noise near cutoff voltage region

6. Anomalies Detected

1. B0034 impedance spikes — Rct shows anomalous jumps, possibly measurement artifact or internal micro-short
2. SOH > 100% — 122.2% max indicates initial capacity above nominal; may need capping at 100% for model inputs
3. Very low capacity (0.044 Ah) — Some cycles with near-zero capacity, likely incomplete discharge or measurement error; should be filtered in preprocessing
4. Battery overlap across temperature groups — Same batteries tested at multiple temperatures complicates group-independent analysis

7. Implications for Modeling

1. Feature engineering should include temperature as a critical feature — capacity degradation mechanisms differ fundamentally between cold and hot operation
2. SOH capping at 100% should be considered for target labels
3. Outlier filtering for capacity < 0.1 Ah recommended (likely measurement artifacts)
4. Impedance features (Re, Rct) are strong predictors — clear monotonic relationship with SOH
5. Cross-temperature generalization is the most challenging task — models should be evaluated for transfer across temperature domains
6. Sequence length varies significantly (89 mean, 196 max) — variable-length handling or padding strategies needed

---

Generated figures saved to artifacts/figures/:

• capacity_fade_all.png
• capacity_fade_by_temp.png
• capacity_violin_box.png
• impedance_evolution.png
• re_vs_rct_phase.png
• soh_distribution.png
• voltage_surface_3d.png
• voltage_surface_3d_interactive.html
• capacity_fade_interactive.html