Train Split of the CBED Dataset for Parameter Identification. Part 2

# CBED-ML Training Dataset (Part 2): Synthetic Data for Uncertainty-Aware Inverse Problems in Electron Diffraction

This repository contains **Part 2 of the training split** of a high-fidelity synthetic dataset for **Convergent Beam Electron Diffraction (CBED)** parameter inference in magnesium diboride (MgB₂). The data was generated to train uncertainty quantification (UQ) methods for scientific inverse problems, where parameters vary in their observability—from tightly constrained (e.g., sample geometry) to weakly constrained (e.g., thermal vibration tensors).

**For the corresponding test split**, see: https://zenodo.org/records/18340683

## Dataset Format

All images are provided as **320×320 center crops** of the original simulated CBED patterns.

### File Structure
“`
f11b/
├── f11b_cbedINDEX.ity.h5
├── …
└── labels.csv
“`

### HDF5 Files (`*.h5`)
Each `.h5` file contains **1,000 samples**:
– `images`: `(1000, 320, 320)` — simulated CBED patterns (float16), center-cropped to 320×320 pixels  
– `filenames`: `(1000,)` — string identifiers (e.g., `”sample_00000.npz”`), used to join with `labels.csv`

### Labels (`labels.csv`)
Columns:
– `file`: base name of the corresponding `.h5` file (e.g., `f11b_cbedINDEX.ity`)
– `conf_0` to `conf_11`: **12 normalized physical parameters** (see below)
– `label`: **sample thickness** in Ångströms (Å)

> **Full 13-parameter vector** = `[conf_0, …, conf_11, label]`

All values in `labels.csv` are in **physical units**, not normalized.

## Parameter Definitions

| Index | Symbol | Description | Units |
|——-|——–|————-|——-|
| `conf_0–2` | `u₁₁(Mg)`, `u₃₃(Mg)`, `u₁₂(Mg)` | Mg Debye-Waller factors (thermal vibration) | pm² |
| `conf_3–5` | `u₁₁(B)`, `u₃₃(B)`, `u₁₂(B)` | B Debye-Waller factors | pm² |
| `conf_6–9` | `F₀₀₁`, `F₁₀₀`, `F₁₀₁`, `F₀₀₂` | Low-order structure factors (valence electron distribution) | Å |
| `conf_10–11` | `θₓ`, `θᵧ` | Beam tilt angles | degrees |
| `label` | `t` | Sample thickness | Å |

Parameter ranges follow uniform sampling over physically realistic intervals (see paper Appendix, Table 1).