Bayesian Optimization for PEM Fuel Cells with Gaussian Processes

This repository helps you choose which PEMFC gas diffusion layer experiment to run next. It loads CSV data, fits a Gaussian Process surrogate model, and ranks candidate experiments with Expected Improvement.

Note: the repo name says physics-guided but the control equations aren't implemented yet. I want to iron out the details before starting to encode the physics

One-Stop Configuration

The main knobs now live in src/config.py.

Edit that file when you want to change:

• FEATURE_COLUMNS: which input variables are active
• TARGET_COLUMN: which metric the model predicts
• OPTIMIZATION_GOAL: whether the optimizer should maximize or minimize the target
• FEATURE_LABELS and TARGET_LABEL: notebook-friendly names for plots and summaries
• FEATURE_BOUNDS: optimizer search ranges, in the same order as the active features

If you add a new feature such as Catalyst, add it in all relevant config sections so the notebooks can both verify it and optimize over it.

Quick Start

Create a Conda environment, activate it, and install the required packages:

conda create -n pemfc-gp python=3.11
conda activate pemfc-gp
pip install -r requirements.txt

Then start Jupyter from the project root:

jupyter notebook

Data Requirements

Put your CSV files inside data/raw/. Every CSV must contain the active feature columns plus the active target column declared in src/config.py.

With the current defaults, that means:

• TiO2_Loading
• RH
• Voltage_1.5A

Notebook Workflow

Most users should run the notebooks rather than calling the src/ modules directly.

1. Open notebooks/01_data_check.ipynb to verify which features and target are active, inspect per-feature ranges, and look for obvious data issues.
2. Open notebooks/02_active_loop.ipynb to fit the GP, inspect the learned kernel, and generate experiment suggestions.

The notebooks now print:

• the active feature list
• the active target
• X_train and y_train shapes
• per-feature min, max, unique-count, mean, and scaling information
• whether any feature is constant in the training data
• whether the configured optimizer bounds are complete

Important limitation: the backend can read any configured feature count, but the contour-style notebook visualization still only works when exactly two active features are configured.

Project Layout

.
├── data/
│   ├── raw/          # original CSV files
│   └── processed/    # generated NumPy arrays
├── notebooks/        # notebook workflow
├── src/              # core Python code
│   └── config.py     # one-stop feature, target, label, and bounds settings
└── requirements.txt  # Python dependencies

Mathematical Background

The model assumes an unknown function f(x) that maps the chosen inputs to the chosen target. A Gaussian Process provides both a prediction and an uncertainty estimate at any candidate point.

The current surrogate uses:

$$k(x, x') = C \cdot k_{\mathrm{Matern}\,5/2}(x, x') + k_{\mathrm{white}}(x, x')$$

and the optimizer ranks candidates with Expected Improvement.

Current Limitations

• The contour visualization is intentionally limited to exactly two active features.
• The optimizer still uses a dense grid, so higher-dimensional searches will get expensive quickly.
• A constant feature can be loaded and verified, but it will not add information to the GP until the data vary along that dimension.
• There is still no automated test suite.