SWORDlib

SWORDlib provides tools for generating SWORD labels from ordered and disordered crystal structures. A SWORD label combines symmetry, standardized Wyckoff sequence information, and site-occupancy disorder information into a compact representation that can be used for structure grouping, disorder curation, and order-disorder family matching.

This package is the installable Python package form of SWORD:

Installation

pip install SWORDlib

Core SWORD Labelling

import sword

Use get_sword_label when you only need the compact label.

from sword import get_sword_label

label = get_sword_label("path/to/structure.cif")
print(label)

The input can be a CIF string, a CIF path, or a pymatgen.core.Structure.

Use get_sword_info when you also need the parsed entry and diagnostic metadata.

from sword import get_sword_info

entry, info = get_sword_info(
    cif_text,
    parser_occ_tolerance=1.05,
    occ_tolerance=1.0,
    site_tolerance=1e-4,
    vac_tolerance=1e-2,
    frac_tolerance=1e-4,
)

print(info["disorder_label"])
print(info["is_disorder"])
print(info["degree_of_mixing"])

The main labelling parameters are:

• symprec: symmetry tolerance used during structure standardization. See pymatgen/spglib documentation for detailed behavior.
• angle_tolerance: angular tolerance, in degrees, used during symmetry standardization. See pymatgen documentation for details.
• symprec_scan: optional adaptive symmetry-tolerance scan. It is disabled by default (False); when enabled, the default symprec_scan_mode is "loose".
• site_tolerance: distance tolerance for deciding whether two sites should be treated as the same disorder site.
• occ_tolerance: SWORD's post-parsing occupancy tolerance for each grouped orbit/site. Occupancy sums above this value are reported as occupancy-orbit errors.
• vac_tolerance: minimum missing occupancy needed to record a VAC component in the disorder label. For example, an occupancy sum of 0.995 does not produce VAC when vac_tolerance=1e-2.
• frac_tolerance: rounding tolerance used in fractional-coordinate and disorder-label operations. Values like 1e-4 or 1e-5 are recommended.
• conventional_struct: whether to use pymatgen's conventionalized structure during the SWORD parsing path.
• refine_struct: whether to let pymatgen refine the structure before labelling. This can change labels and should be used deliberately.

Two occupancy-related parameters are intentionally separate. parser_occ_tolerance: occupancy tolerance passed to pymatgen's CIF parser. Acts before SWORD, while the CIF is being read. occ_tolerance: SWORD's post-parsing occupancy tolerance for each grouped orbit/site. Occupancy sums above this value are reported as occupancy-orbit errors. For ICSD curation, a common choice is parser_occ_tolerance=1.05 and occ_tolerance=1.0.

ICSD Curation Pipeline

run_icsd_dedup_pipeline provides a high-level workflow for ICSD dataframes containing a CIF text column and a stable entry ID column.

import pandas as pd
from sword import run_icsd_dedup_pipeline

df = pd.read_pickle("ICSD2025_summary.pkl")

result = run_icsd_dedup_pipeline(
    df,
    cif_col="cif",
    id_col="CollectionCode",
    mode="from_collection_code",
    prescreen_params={
        "parser_occ_tolerance": 1.05,
        "excluded_elements": ("He", "Ne", "Ar", "Kr", "Es"),
        "exclude_hydrogen": True,
    },
    sword_params={
        "parser_occ_tolerance": 1.05,
        "occ_tolerance": 1.0,
        "site_tolerance": 1e-4,
        "vac_tolerance": 1e-2,
        "frac_tolerance": 1e-4,
    },
    family_info=False,
    disorder_refinement_policy="dom_vector",
    composition_tol=0.05,
    dom_tol=0.03,
    output_dir="sword_icsd_results",
)

The result object contains:

• result.prescreen_rejected: entries removed before labelling, with reject reasons.
• result.label_results: the labelled dataframe with SWORD labels and compact diagnostics.
• result.anomalies: detailed warning/error tables.
• result.label_groups: grouped SWORD-label summary table.
• result.refined: curated and deduplicated entries.

Disordered entries use the dom_vector refinement policy by default. Its default tolerances are composition_tol=0.05 and dom_tol=0.03; smaller values make refinement stricter. The representative-site occupancy vector used by the policy is available in result.label_results["dom_site_occ_vector"].

result.prescreen_rejected["reject_reason"] records why a row did not enter the labelling stage. Possible reasons include:

• parse_error: the CIF block could not be read by pymatgen.
• hydrogen: the CIF site labels contain H, D, or T when exclude_hydrogen=True.
• non_element: the CIF site labels contain unsupported/non-element symbols or elements listed in excluded_elements.
• coordinate_error: fractional coordinates are missing or malformed.
• wyckoff_error: Wyckoff symbols are missing or malformed.
• multiplicity_error: site multiplicities are missing or malformed.
• occupancy_error: a raw site occupancy is missing, non-positive, or larger than parser_occ_tolerance.

result.label_results["processing_issue"] records semicolon-separated warning or error flags generated during labelling. None means no issue was recorded. The detailed rows are stored in result.anomalies and, when output_dir is provided, saved under anomalies/:

• label_errors: hard failures during structure parsing or SWORD label generation. The compact flag: structure_parse_failed or label_generation_failed.
• wyck_float_warn: Wyckoff coordinate expansion gives more positions than the declared multiplicity, usually indicating floating-point or refinement issues in the CIF. The compact flag: wyckoff_float_warning.
• occ_err_sites: grouped orbit/site occupancy exceeds occ_tolerance. The compact flag: occupancy_orbit_error.
• equivalent_but_distinct_sites: sites are symmetry-equivalent but have distinct representative coordinates, often because the CIF coordinates are not exactly on the expected special position. The compact flag: equivalent_sites_warning.
• same_valence_sites: same-element/same-valence co-occupancy was detected and merged for labelling. The compact flag is same_valence_site_warning.
• intersect_orb_errors: the positional-disorder check failed. The compact flag is positional_check_failed. Positional disorder itself is recorded in is_positional_disorder.

By default, the refined table removes hard parse/label errors, positional disorder, occupancy-orbit errors, Wyckoff floating-point warnings, and equivalent-site warnings. Same-valence warnings are recorded but not removed by default.

Set family_info=True to append a SWORD_family_dic column to result.label_results. This is slower and usually not needed for a first ICSD curation pass.

result = run_icsd_dedup_pipeline(
    df,
    family_info=True,
    family_params={
        "parser_occ_tolerance": 10.0,
        "fill_vacancy": False,
    },
)

family_params must be a dictionary accepted by SWORDFamilyMatcher. Use fill_vacancy=False for most large reference-table precomputations such as ICSD/COD. Set fill_vacancy=True mainly for query-side ordered structure datasets/databases that may contain vacancy ordering or missing-site defects, e.g. Materials Project, LeMat-Bulk, where filling possible vacancies can help expose candidate disorder parents.

Generic Database Labelling

For non-ICSD databases, use label_dataframe. The dataframe only needs a CIF text column and a stable ID column.

from sword import label_dataframe

labelled, anomalies = label_dataframe(
    df,
    cif_col="cif",
    id_col="material_id",
    sword_params={
        "parser_occ_tolerance": 1.05,
        "occ_tolerance": 1.0,
        "site_tolerance": 1e-4,
        "vac_tolerance": 1e-2,
        "frac_tolerance": 1e-4,
    },
    family_info=False,
)

print(labelled[["material_id", "SWORD_label", "is_disorder"]])

This interface is appropriate for MP, LeMat, or user-built structure tables when each row can provide a CIF string.

Set family_info=True if you also want family dictionaries:

labelled, anomalies = label_dataframe(
    df,
    cif_col="cif",
    id_col="material_id",
    family_info=True,
    family_params={
        "symprec_child": 1e-2,
        "symprec_search": 1.0,
        "parser_occ_tolerance": 10.0,
        "fill_vacancy": False,   #turn on to True if this dataset may contain vacancy ordering and you want to find disordered parent structure of this vacancy-type ordered structure
    },
)

print(labelled[["material_id", "SWORD_label", "SWORD_family_dic"]])

Order-Disorder Family Matching

SWORDFamilyMatcher can generate a SWORD family dictionary for an ordered query structure and search for possible disordered parent labels.

from sword import SWORDFamilyMatcher

matcher = SWORDFamilyMatcher(
    symprec_child=1e-2,
    symprec_search=1.0,
    parser_occ_tolerance=10.0,
    fill_vacancy=False,
)

family = matcher.get_sword_dic("path/to/ordered_query.cif")

print(family["child_label"])
print(family["parent_labels"])

For repeated searches, precompute a family dictionary column for the reference table. This can be slow for large datasets, so it is usually done once and saved. For large reference tables, keep fill_vacancy=False unless you explicitly want the reference-side vacancy-filling expansion.

from sword import SWORDFamilyMatcher

matcher = SWORDFamilyMatcher(fill_vacancy=False)

reference_df["SWORD_family_dic"] = reference_df["cif"].apply(
    matcher.get_sword_dic
)

Then compare a query against the reference table:

matches = matcher.fit_many(
    query_structure,
    reference_df,
    disorder_refinement_policy="dom_vector",
    composition_tol=0.30,
    dom_tol=0.30,
)

print(matches["matched_disordered_ids"])
print(matches["matched_disordered_labels"])

By default, fit_many reads reference_df["SWORD_family_dic"]. If the query is a dataframe row with SWORD_family_dic and SWORD_family_dic_vac, both query columns are used and matches["matched_source_by_id"] records which query column produced each match.

For comparable direct disorder matches, fit_many also returns dom_distance_by_id and dom_match_by_id, together with matching_mode_by_id and sg_relation_by_id for concise match diagnostics. fit_many uses the dom_vector policy by default with composition_tol=0.30 and dom_tol=0.30. Use smaller tolerances when a stricter parent-composition match is required.

fill_vacancy=True can be useful when ordered structures may represent vacancy ordering variants. In practice, this is most useful for query structures that are ordered but may be vacancy-ordered or vacancy-deficient variants. For large ICSD-style reference precomputation, fill_vacancy=False is usually a better default because it is faster and avoids generating extra vacancy-filled candidates for every entry.

Notes

SWORD labels are intended for symmetry- and Wyckoff-aware grouping of materials structures. They are especially useful for comparing ordered and disordered entries under a common structural representation, but they do not replace manual crystallographic judgment for ambiguous or low-quality CIF records.

Citation

If you use SWORDlib, please cite our papers:

@article{huang2026sword,
  title   = {SWORD: Symmetry and Wyckoff-sequence of Ordered and Disordered crystals},
  author  = {Huang, Yuyao and Nong, Wei and Yamazaki, Shuya and Petersen, Martin Hoffmann and Wang, Jianghai and Zhu, Ruiming and Hippalgaonkar, Kedar},
  journal = {arXiv preprint arXiv:2604.17994},
  year    = {2026},
  url     = {https://arxiv.org/abs/2604.17994},
  doi     = {10.48550/arXiv.2604.17994}
}

For family-matching functions, please cite:

@article{yamazaki2026orderdisorder,
  title   = {Navigating Order-(Dis)Order Family Trees via Group-Subgroup Transitions},
  author  = {Yamazaki, Shuya and Huang, Yuyao and Petersen, Martin Hoffmann and Nong, Wei and Hippalgaonkar, Kedar},
  journal = {arXiv preprint arXiv:2604.21386},
  year    = {2026},
  url     = {https://arxiv.org/abs/2604.21386},
  doi     = {10.48550/arXiv.2604.21386}
}