Auto Volume And Minsep

ap run search can derive buildcell #VARVOL, #MINSEP, and #NFORM directives from a volume/minsep estimate while preserving unrelated seed settings.

The first implementation is intentionally lightweight. It supports exact lookup from a user-supplied curated JSON dataset, a baseline JSON bundle trained from that dataset, or reference structures. It does not require torch, and the large curated dataset is not bundled with airsspy.

Preview Generated Seed Text

Use --diagnose before a long search:

ap run search --seed seed --build-only --formula SiO2 \
  --volume-minsep-source dataset \
  --volume-minsep-dataset generation/minsep_vol_dataset_curated.json \
  --diagnose 3

The diagnostic output shows the selected formula, estimate provenance, volume per atom, total volume, generated minsep ranges, resolved #NFORM, and the final buildcell input.

Use A Baseline Bundle

After training a baseline bundle, point search at the generated baseline_bundle.json:

ap run search --seed seed --formula SiO2 \
  --volume-minsep-source baseline \
  --volume-minsep-bundle artifacts/baseline/baseline_bundle.json

--volume-scale, --minsep-scale-low, --minsep-scale-high, --max-atoms, and --max-nform tune the generated buildcell directives.

Generate Training Data

If you have a Materials Project document dump with structures and energy_above_hull, build and curate the dataset offline:

ap tools volume-minsep-build-dataset \
  --mp-docs generation/mp-stable-docs.json \
  --output generation/minsep_vol_dataset.json

ap tools volume-minsep-curate-dataset \
  --dataset generation/minsep_vol_dataset.json \
  --mp-docs generation/mp-stable-docs.json \
  --output generation/minsep_vol_dataset_curated.json

The curation step keeps the lowest energy_above_hull row for each reduced formula.

Train The Baseline Predictor

Train the non-torch ridge baseline from the raw or curated dataset:

ap tools volume-minsep-train-baseline \
  --dataset generation/minsep_vol_dataset_curated.json \
  --output-dir formula_model_artifacts

The command writes baseline/baseline_bundle.json, aggregated target JSON files, and a training summary under the output directory.