MicroscopyStructurePropertyBenchmark - In active development - This Readme might not be updated!

What it is: Minimal active-learning loop for structure-property relationships in electron microscopy. Small, modular, easy to extend.

Pipeline : dataset → representation → model → acquisition → reward/metric → log

---

Quick start

uv sync --extra dev

uv run mspb --config configs/pca_gp_ei.yaml

Real STEM file: place test_stem.h5 at data/raw/, use configs/stem_pca_gp_ei.yaml.
Sweep all methods: uv run mspb-sweep --config configs/stem_all_methods.yaml

Plotting results? - only available for mspb-sweep (i.e all methods sweep option)

Plotting: 
uv run python scripts/postprocess_csv.py --csv outputs/stem_all_methods_100_steps.csv

---

Datasets — synthetic, stem_h5

Representations — pca, patches

Models — gpytorch_gp, dkl

Acquisition — EI, UCB, beacon, random

Rewards — dipole (0.35–0.55 eV), edge (0.60–0.75 eV), bulk (0.80–1.00 eV), zero

---

Outputs per run

• predictions_BO_step<N>.png/.pkl — mean, variance, true scalarizer per step
• Active_learning_statistics.pkl — acquisition order, seeds, traces, coords
• AL_traj.png — trajectory over image
• run.log + training_log.jsonl — human and structured logs
• checkpoints/model_step<N>.pt + latest.pt (optional)
• Sweep CSV: one row per method per step — mse, mae, nlpd, coverage, loss

---

Adding a method — implement representation, model.fit/predict, acquisition, runner. Add a config in configs/ and a smoke test in tests/.

---