matgen-BO

Reproduction code, data, and figures for the paper's three subsystems: probabilistic surrogates on pretrained atomistic embeddings (GP / MTGP / DGP × MACE/ORB/UMA/SOAP), a closed-loop generative discovery loop (MatterGen / CrystalFlow / ADiT, GP-routed Expected-Improvement + DPP gate), and an ORB-PU synthesizability classifier.

Layout

01_static_benchmark/   surrogate benchmark — elastic, dielectric, phonon-thermodynamics (DFPT + Pheasy)
02_closed_loop_bo/     discovery loop + oracles (heat capacity, bulk modulus)
03_synthesizability/   ORB-PU classifier (checkpoints on HuggingFace — see docs/EXTERNAL.md)
shared/                common.py, plot_style.py, data_loaders.py
docs/                  EXTERNAL.md (Zenodo/HF), PROVENANCE.md, fetch_external.sh
archive/               verbatim copies of every source tree (completeness; not needed to reproduce)
REPRODUCE.md           which command regenerates each paper Figure/Table

Install

conda env create -f environment.yml   # creates the `matinvent` env
conda activate matinvent
# (or: pip install -r requirements.txt)

Reproduce

Everything needed to regenerate the figures is in-tree. Each generator resolves the repo root from its own location and reads the complete verbatim data mirror under archive/<source-tree>/ (e.g. archive/ASE-native-surrogates/analysis_v3/…), writing freshly rendered figures to 0*/figures/regenerated/. The paper-reference figures under 0*/figures/rendered/ are never overwritten. No external downloads required for figures.

conda activate matinvent
bash reproduce.sh          # regenerates figures for all three subsystems from in-tree data
# or per subsystem:
bash 01_static_benchmark/run.sh

See REPRODUCE.md for the Figure/Table → command map.

Self-contained: the generators are path-portable — no machine-specific paths, no manual edits. They discover the repo root via Path(__file__) and read in-tree archive/ data, so a fresh clone reproduces the figures as-is. s13_s14_regen also self-checks its reproduced R² values against the published numbers.

External (full from-scratch rerun only)

Raw 79 GB phonon deposit (Zenodo 20196565) and model checkpoints (HuggingFace SheikhAhnaf/...) are not needed to reproduce figures — see docs/EXTERNAL.md / docs/fetch_external.sh.

License

MIT — see LICENSE.