This directory contains density functional theory (DFT) convex hull calculations, materials data used to train the polaritype-based recommender model, and candidate compounds predicted by the recommender model. Directory contents are described below.

---

Convex Hull DFT Data

hull_calcs.tar.gz

Compressed archive of DFT input and output files for convex hull stability analysis. Calculation directories are sorted into the following directory structure:

.
├── hull_calcs/
│   ├── comp_A4B2C/
|   |   ├── calc_A2B3C4_<identifier>--<AFLOW prototype designation>/
|   |   |    ├── relax
|   |   |    └── scf
|   |   └── calc_.../
│   ├── comp.../
|   └── incomplete_members/
|       ├── calc_AxBy_<identifier>--<AFLOW prototype designation>/
|       |   ├── relax
|       |   └── scf
|       └── calc_.../

The comp_A4B2C directories contain calculations of all ternaries in the same chemical system of the candidate compound A4B2C. Directories with compknown_ represent calculations for known compounds in the polaritype of interest. Each calc_* directory contains a subdirectory for the ionic relaxation and the self-consistent field (SCF), or static calculation. The calc_* directories are labeled with the chemical formula, the structure identifier (or "candidate" for the predicted compounds), and the AFLOW prototype designation. Output files include the std_isX_ibY_itZ.out which are the VASP standard output files for the relaxation iteration Z with ISIF = X and IBRION = Y. The incomplete members directory contains the cumulative group of all convex hull-relevant compounds with less than 3 elements, e.g. A3B2. When calculating convex hulls for each candidate compound, the appropriate unaries and binaries from within the candidate compound's chemical system were selected for analysis from this directory. Since prototyping/decoratyping was not important for the incomplete members, the prototypes all show up as "None".

---

Training Data (Materials Properties)

known_data_compounds.json

Augmented dataset containing materials properties for experimentally validated structures/compounds from the Materials Project and the ICSD. Also includes assigned polaritypes. This dataset contains element-resolved, site-resolved, and aggregate statistical descriptors of crystalline compounds. Column names follow consistent naming conventions that encode their meaning.

---

Aggregate Statistical Descriptors Across Elements

Columns of the form:

<statistic>_<property>

describe aggregate statistics computed across all distinct elements in the compound.

1.1 Supported Statistics

1.2 Supported Properties

1.3 Example Interpretation

---

2. Site-Resolved Elemental Descriptors (typeX_*)

Columns of the form:

typeX_<property>

describe properties of the element occupying Wyckoff site X, where:

2.1 Supported Site Properties

These mirror the elemental properties listed in Section 1.2, but apply to a specific crystallographic site.

2.2 Example Interpretation

---

3. Multi-Site Vector Descriptors (type_*)

These columns store ordered vectors corresponding to Wyckoff sites, consistent with the typeX_* ordering.

---

4. Composition & Structural Metadata

---

5. Electronic & Bonding Descriptors

---

known_data_prototypes.json

Augmented dataset containing structural prototype grouping and polaritype population information.

Column Descriptions

training_data.pkl

Cleaned, imputed, and scaled version of the known data -- used to train the polaritype classifier. Serialized as a Python pickle object. Columns in this dataset are described in the sections above with the exception of the scaled_* columns which are versions of the columns described above after being passed through the provided scaler.

scaler.joblib.gz

Scikit-learn StandardScaler object saved with joblib. Used to scale the data in training_data.pkl. The *.joblib.gz files can be unpacked by importing joblib (https://joblib.readthedocs.io/en/stable/) and running the following command:
scaler = joblib.load(path/to/scaler.joblib.gz).

scaler_means.npy

Numpy array storing the means applied by the scaler.

scaler_stds.npy

Numpy array storing the standard deviations applied by the scaler.

---

Candidate Compounds

candidate_data_0_feat.csv

Featurized representation of candidate compounds evaluated by the ML discovery pipeline.

Column Descriptions
See the columns descriptions in the above sections for the majority of the columns in this table. Additional columns are described below. The cousin modifier indicates that the property refers to the known structure (from known_data_compounds.json) which was most closely related to the candidate compound according to the maximum compound substitution probability (described below).

candidate_data_1_imputed.pkl

Preprocessed candidate dataset with missing values imputed and values scaled. Serialized as a Python pickle object. Columns in this dataset are described in the sections above with the exception of the scaled_* columns which are versions of the columns described above after being passed through the provided scaler.