# Understanding PEM Electrolyzer Data

```{objectives}
- Understand what PEM electrolyzers are and the SWVF technology
- Learn about the three NORCE experimental datasets and their characteristics
- Know the input features and their physical meaning
- Understand data filtering (strict vs minimal) and why it matters
- Understand the OOD evaluation protocol and keep-out validation
```

```{admonition} Why This Matters
:class: tip

**The Scenario:** A quality assurance engineer at a fuel-cell testing facility receives three batches of polarization curve data from different test campaigns. Some measurements look suspiciously noisy — maybe a sensor was drifting, or a test was interrupted. Before feeding this data into a machine learning model, she needs to know: *which data points are trustworthy, and which should be excluded?*

**The Research Question:** How do you systematically detect and exclude contaminated experimental data, and how do you design a validation protocol that proves your model generalizes — rather than just memorizing the training conditions?

**What This Episode Gives You:** The data pipeline — from raw NORCE measurements to clean, filtered datasets with a rigorous keep-out validation split.
```

## PEM Electrolysis

PEM (Proton Exchange Membrane) electrolyzers produce hydrogen by splitting water using electricity:

- **Anode**: 2H₂O → O₂ + 4H⁺ + 4e⁻
- **Cathode**: 4H⁺ + 4e⁻ → 2H₂

The cell voltage depends on operating conditions: current, pressure, and temperature.

## SWVF Technology

The electrolyzer uses **Static Water Vapour Feed (SWVF)** technology (TRL4), which is fundamentally different from conventional liquid-fed PEM:
- Water vapor diffuses through a **Water Feed Barrier** membrane to the anode
- This eliminates liquid water management but adds ionic resistance
- The SWVF barrier explains why the learned ohmic resistance (~1.0 Ohm*cm2) is higher than typical liquid-fed PEM values (0.05-0.20 Ohm*cm2)

This is important context when interpreting the learned physics parameters — the model correctly captures this technology-specific resistance.

## NORCE Experimental Data

The data comes from a Proton OnSite M400 electrolyzer stack at the **NORCE test facility in Bergen, Norway**, consisting of 4 cells with a total active area of **50 cm²** (4 x 12.5 cm2).

### Available Datasets

| Dataset | Date | Purpose | Samples | Duration | Description |
|---------|------|---------|---------|----------|-------------|
| Test4 | 2024-10-16 | **Training** | ~171,838 | ~50 hrs | Long-term stability, discrete steps, bimodal 15/35 bar |
| Test2 | 2024-09-27 | **OOD Evaluation** | ~10,603 | ~9 hrs | Variable current sweep, 0-30 bar range |
| Test3 | 2024-09-30 | **OOD Evaluation** | ~25,749 | ~10 hrs | Constant 6A, pressure swap 0-40 bar |

The three test campaigns were conducted under **different operating protocols**, ensuring the model is evaluated on truly different conditions — not just held-out samples from the same experiment.

```{figure} ../images/data_overview.png
:alt: Overview of NORCE experimental datasets
:width: 100%

Overview of the NORCE experimental data across the three test campaigns, showing the range of operating conditions covered.
```

### Input Features

| Column | Variable | Unit | Description |
|--------|----------|------|-------------|
| PS-I-MON | Current | A | Stack current (4 cells in series) |
| H-P1 | H₂ Pressure | bar | Hydrogen back pressure |
| O-P1 | O₂ Pressure | bar | Oxygen back pressure |
| T-ELY-CH1 | Temperature | °C | Stack temperature |
| CV-mean | Voltage | V | Mean cell voltage (target) |

```{figure} ../images/correlation_heatmap.png
:alt: Feature correlation matrix
:width: 70%

Feature correlation matrix for the training data. Pressure and temperature are strongly correlated with voltage (r > 0.85), while current has moderate correlation (r = 0.56). The near-perfect H₂/O₂ pressure correlation (0.997) reflects symmetric operation.
```

```{figure} ../images/input_distributions.png
:alt: Input feature distributions
:width: 100%

Distribution of input features. Note the non-uniform coverage — the model must extrapolate to operating regimes not seen during training.
```

```{figure} ../images/voltage_response.png
:alt: Voltage response surfaces
:width: 100%

Voltage response surfaces showing the nonlinear relationships: (Left) current vs temperature, (Center) current vs pressure, (Right) temperature vs pressure. Higher current and pressure increase voltage.
```

### Data Filtering

Two filtering levels are used:

**STRICT filtering** (training data only):
- Current ≥ 5 A (exclude low-current transients)
- H₂/O₂ pressure > 10 bar (exclude startup/shutdown)
- Temperature 70-85°C (operational range)
- Voltage 1.5-2.0 V (exclude anomalies)

**MINIMAL filtering** (OOD evaluation):
- Remove NaN values
- Remove negative pressures
- Remove cold startup (I < 1A AND T < 60°C)
- Voltage 1.0-2.5 V

### Why Two Filtering Levels?

Strict filtering for training ensures the model learns from clean, steady-state data within the electrolyzer's nominal operating range. The narrow training window (current > 5A, pressure > 10 bar) means the model never sees low-pressure or low-current data during training.

Minimal filtering for evaluation preserves the full range of operating conditions, including regions entirely absent from training data. This is critical for testing OOD generalization — the model must predict accurately at pressures below 10 bar and currents below 5A despite never training on them.

After strict filtering, only ~7.9% of Test4 raw samples pass (15,234 out of ~193K), because most samples are at low current during ramp-up/down transients.

### Out-of-Distribution (OOD) Protocol

Test4 is used for training, while Test2 and Test3 are used for evaluation. These datasets were collected under different operating protocols:
- **Test2**: Systematic current sweeps at varying pressures (0-30 bar) — tests current extrapolation
- **Test3**: Pressure swap experiments at fixed current (6A), 0-40 bar — tests pressure extrapolation
- **Test4**: Long-term stability at various conditions, bimodal 15/35 bar — training data

Key OOD challenges:
- **Low-pressure regime** (< 10 bar): Entirely absent from training data
- **Pressure range**: Training sees bimodal 15/35 bar; Test3 sweeps 0-40 bar continuously
- **Current patterns**: Training uses discrete steps; Test2 uses continuous sweeps

This ensures the model is evaluated on truly unseen operating conditions.

```{figure} ../images/comprehensive_correlation.png
:alt: Comprehensive correlation analysis
:width: 100%

Comprehensive correlation analysis across all sensor channels, revealing the complex multivariate relationships in the electrolyzer data.
```

```{figure} ../images/cell_voltage_analysis.png
:alt: Cell-to-cell voltage analysis
:width: 90%

Cell-to-cell voltage analysis from the 4-cell stack: (Top-left) Voltage distributions per cell. (Top-right) Cell voltage variation over time. (Bottom-left) Cell-to-cell correlation matrix. (Bottom-right) Mean vs individual cell distributions.
```

```{figure} ../images/feature_importance.png
:alt: Feature importance by correlation with voltage
:width: 90%

Feature importance ranked by absolute correlation with voltage. The top features (PS-V-MON, T-ELY-avg, W-FM1) have the strongest predictive power, guiding feature selection for the model.
```

```{figure} ../images/temperature_analysis.png
:alt: Temperature effects on voltage
:width: 100%

Temperature analysis showing how voltage varies with temperature at different operating points. Higher temperature generally reduces voltage, improving electrolyzer efficiency.
```

### Keep-Out Validation

For training, a temperature-based keep-out validation split is used:
- **Training**: T < 76°C OR T > 80°C
- **Validation**: 76°C ≤ T ≤ 80°C

This forces the model to interpolate to an unseen temperature range, which is critical for OOD generalization.

```{keypoints}
- SWVF technology differs from liquid-fed PEM — explains higher ohmic resistance
- Data from NORCE Bergen: Proton OnSite M400 stack (50 cm² active area, 4 cells)
- Three test campaigns: Test4 (training, ~172K samples), Test2/Test3 (OOD evaluation)
- STRICT filtering for training (7.9% pass rate), MINIMAL for evaluation
- Low-pressure regime (< 10 bar) is entirely absent from training data
- Keep-out validation (76-80°C) is critical for OOD generalization
- H₂ and O₂ pressures are kept separate for correct Nernst equation
```