xpanceo-team/optical-anisotropy

Description

Contains structures for training generative models with dn / k condition. Concatenation of NOEMD dataset, vdW dataset and dft relaxed generated structures. Filtered with <=20 atoms.

Building a dataset

Version after 23.10.2025:

import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
from datasets import Dataset, DatasetDict, load_dataset, concatenate_datasets
from pymatgen.analysis.structure_matcher import StructureMatcher
from pymatgen.core import Structure
from tqdm.auto import tqdm

noemd_ds = load_dataset("xpanceo-team/noemd_n_k_dn_new_1064")
dft_ds = load_dataset("xpanceo-team/dn-k-generated-dft-relaxed", split="train")
vdw_ds = load_dataset("xpanceo-team/vdw-n-k-dn-1064", split="train")

dft_ds = dft_ds.rename_columns({"dft_relaxed_structure": "structure", "dn_max": "dn", "k_avg": "k_1064"})
vdw_ds = vdw_ds.rename_columns({"k_avg": "k_1064"})

dft_ds = dft_ds.remove_columns(['dn_xy', 'dn_yx', 'dn_xz','dn_yz',])
vdw_ds = vdw_ds.remove_columns(['k_exx', 'k_eyy', 'k_ezz', 'n_exx', 'n_eyy', 'n_ezz'])
noemd_ds = noemd_ds.remove_columns(['nx', 'ny', 'nz', 'kx', 'ky', 'kz'])


def get_matches(s, ds):
    matcher = StructureMatcher(
        ltol=0.1, stol=0.1, angle_tol=2,
    )
    
    s = Structure.from_str(s, fmt="json")
    
    # add value for datasets to determine types
    matches = [-1, ]
    for idx, row in enumerate(ds):
        row_s = Structure.from_str(row["structure"], fmt="json")
        if matcher.fit(s, row_s):
            matches.append(idx)
            
    return matches
            
    
noemd_ds = noemd_ds.map(lambda x: {"matches_vdw": get_matches(x["structure"], vdw_ds)}, num_proc=96)
noemd_ds = noemd_ds.map(lambda x: {"matches_dft": get_matches(x["structure"], dft_ds)}, num_proc=96)
noemd_ds_matched = noemd_ds.filter(lambda x: len(x["matches_vdw"]) > 1 or len(x["matches_dft"]) > 1)

noemd_ds_matched = concatenate_datasets([noemd_ds_matched["train"], noemd_ds_matched["valid"], noemd_ds_matched["test"]])
noemd_ds_matched = noemd_ds_matched.map(
    lambda x: {"matches_vdw": x["matches_vdw"][1:], "matches_dft": x["matches_dft"][1:]}
)

drop_dft_ds = []
for row in noemd_ds_matched["matches_dft"]:
    drop_dft_ds.extend(row)
    
drop_vdw_ds = []
for row in noemd_ds_matched["matches_vdw"]:
    drop_vdw_ds.extend(row)

dft_ds = dft_ds.select([i for i in range(len(dft_ds)) if i not in drop_dft_ds])
vdw_ds = vdw_ds.select([i for i in range(len(vdw_ds)) if i not in drop_vdw_ds])
noemd_ds = noemd_ds.remove_columns(["matches_dft", "matches_vdw"])

ds = DatasetDict({
    'train': concatenate_datasets([noemd_ds["train"], vdw_ds, dft_ds]),
    'valid': noemd_ds["valid"],
    'test': noemd_ds["test"],
})

ds = ds.filter(lambda row: Structure.from_str(row["structure"], fmt="json").num_sites <= 20)

ds.push_to_hub("xpanceo-team/optical-anisotropy", private=True)

Version before 23.10.2025: Contains some structures from JARVIS, some from NOEMD with deduplication attempt