fgh123654/3dsc-mp-tc-pilot

3DSC MP superconducting temperature pilot data

This release contains the 5,773 records actually used in Guohuan Feng's superconductivity pilot on 1 October 2026: critical temperature labels, fixed split assignments, numerical input features, and matched crystal structures. It is a documented research snapshot derived from the existing 3DSC MP database, with a new preparation and grouping protocol. The original data authors retain credit for the underlying database. This release does not establish new superconductors or independent predictive validation.

Code, workflow, results and limitations are in the GitHub research report. The retrospective_test cohort had been analyzed before this pilot and must be treated as a historical diagnostic cohort.

Data and fixed splits

Split	Records	Connected groups	Role in the pilot
train	3,764	1,117	Fit models and preprocessing
validation	869	280	Adaptive descriptor selection
retrospective_test	1,140	350	Historical diagnostic evaluation after freezing
Total	5,773	1,747	Complete released cohort

Groups are connected components linking records that share a chemical system or an original Materials Project parent. The existing strict split was reused with seed 20260929. Across the three splits, the preparation audit found no overlap in material row IDs, connected groups, chemical systems, MP parents, or canonical compositions. This grouping reduces these specified forms of overlap; it does not remove all possible dataset or measurement biases.

Choose records in the online viewer for the readable 14-column label and metadata tables. Choose features for the same records joined one to one with 109 composition features and 28 conventional structure features: 151 columns in total. These 137 numerical features are the prepared baseline inputs, not the later agent-generated descriptor values.

Fields and predictor separation

Column	Meaning
material_id	Unique source-row ID of the form 3dsc_mp_row_000000
row_id	Row index in the exact upstream table
split	Fixed role: train, validation or retrospective_test
tc	Critical temperature target in kelvin
weight	Inverse number of matched rows per canonical composition in the original full cohort
group	Connected component used for splitting
chemical_system	Element set used for grouping
parent_id	Original MP parent; it is not a unique material record
formula	Matched material composition
class	Upstream superconductor family metadata
canonical_composition	Canonical composition key used for grouping and weights
synth_doped	Whether the matched structure was artificially doped upstream
upstream_weight	Original 3DSC weight, retained separately from this pilot's weight
cif	Matched CIF filename in the structure archive

For model inputs, select only the raw_features and/or structure_features arrays in feature_schema.json. Treat the 14 columns above as labels, identifiers, weights or diagnostic metadata, not as numerical predictors. In particular, tc is the target; row IDs, class, formula, parent IDs and grouping keys were not passed into descriptor generation. Learn imputation or other preprocessing from training records only. Do not use the historical cohort for adaptive selection.

tc = 0 preserves the upstream encoding for non-superconductors. It is a database label, not a new experimental verification under every pressure, composition or temperature condition. The pilot includes both these zero labels and positive temperatures.

Structures and downloadable evidence

Partial occupancies and mixed species are preserved. Of the released records, 3,435 have disordered structures and 3,437 have artificially doped structures; these categories overlap. The supplied cells were parsed without primitive reduction, ordering, relaxation or idealization. Synthetic doping reflects the upstream matching procedure and is not a new structure calculation.

The prepared JSONL structure payloads contain geometry and species only, with complete directed periodic neighbors within 6 Å, including nonzero self images and image multiplicities. Exact occupancies are in species_occupancy; a mixed site symbol does not select a dominant species. Site-averaged atomic properties can be noninteger, and unknown properties remain null. Read descriptor_schema.json before constructing chemistry-sensitive descriptors.

File	Contents
records_*.csv.gz	Labels and metadata, separated by fixed split
features_*.csv.gz	The same records plus 137 prepared numerical features
frozen_data.zip	Original prepared data files, retained byte for byte
cifs.zip	All 5,773 matched source CIF files
upstream_3DSC_MP.csv.gz	Exact upstream table after decompression
cif_manifest.csv	CIF filenames and SHA-256 hashes
original_source_manifest.json	Upstream provenance and source hashes
dataset_manifest.json	Preparation, weighting, split and limitation details
feature_schema.json, descriptor_schema.json	Numerical feature lists and structure field definitions
split_assignments.csv	Original fixed split and grouping assignments
verification_report.json	Export, identity, split-overlap and archive checks
FILE_MANIFEST.json	Release file hashes; excludes itself
UPSTREAM_LICENSE.md	License and attribution details

No trained model files are included in this dataset release. The historical pilot found only a small overall MAE difference with uncertainty spanning zero, while positive-temperature and high-temperature errors worsened. A selected agent descriptor also failed a rotation-invariance audit. See the report for the measured results and proposed corrections; publication of these inputs does not resolve those findings.

Loading

from datasets import load_dataset

records = load_dataset("fgh123654/3dsc-mp-tc-pilot", "records")
features = load_dataset("fgh123654/3dsc-mp-tc-pilot", "features")
print(len(records["train"]))  # 3764

Provenance and attribution

The source is aimat-lab/3DSC, pinned at commit 2471dd51a298a854cb4f365ebd39e72c7cbf3634. The upstream file is superconductors_3D/data/final/MP/3DSC_MP.csv, SHA-256 353b86f9c60505d11a3e25df8312da02ff88fd8e27af05b632dec4739d52eb24.

The upstream license statement licenses 3DSC MP data under CC BY 4.0 and identifies the source SuperCon/Stanev and Materials Project data. This release contains the MP branch only; it contains no ICSD structures. Changes made for this snapshot are the row identifiers, prepared features and structure payloads, strict grouped split, evaluation weights, joined viewer tables, and documentation. These prepared data are released under CC BY 4.0 with upstream attribution. Upstream software has a separate MIT license.

Please acknowledge the original data authors when reusing this snapshot:

• Sommer, T., Willa, R., Schmalian, J. and Friederich, P. 3DSC — A New Dataset of Superconductors Including Crystal Structures. arXiv:2212.06071 (2022).
• Stanev, V. et al. Machine learning modeling of superconducting critical temperature. npj Computational Materials 4, 29 (2018).
• Materials Project and SuperCon.

For the preparation and fixed split, also cite this repository by its exact Hugging Face revision and link the GitHub report. Release curator: Guohuan Feng (fgh123654).

中文说明

这是本轮超导临界温度研究实际使用的 5,773 条 3DSC MP 数据整理快照。原始数据库来自 3DSC 作者，数据按 CC BY 4.0 发布。这里公开的是本轮使用的标签、分组划分、特征、结构及校验记录；原始数据来源和我们新增的处理步骤均在上文说明。

在页面的配置选项中选择 **records**，可以查看化学式、临界温度 tc（K）、MP 父结构、数据划分等 14 列；选择 **features**，可以查看同一批材料加上 109 个成分特征和 28 个常规结构特征，共 151 列。tc 是预测目标，标识符、类别、化学式和分组信息不应直接作为模型输入。实际可用特征以 feature_schema.json 的两个特征列表为准。

训练集 3,764 条，验证集 869 条，历史诊断集 retrospective_test 1,140 条。化学体系或 MP 父结构相同的材料通过连通分量共同分组，各划分之间没有这些分组键和规范化成分的交叉。历史诊断集此前已经分析过，因此本轮结果不构成新的独立测试证据。

全部 5,773 个 CIF 可在 cifs.zip 下载，部分占位、混合元素和上游人工掺杂均保留。人工掺杂结构来自数据库的匹配流程，不能当作新计算或新实验确认的结构。tc = 0 是上游数据库的非超导标签，也不能推断该材料在所有实验条件下均无超导性。

本数据集服务于复现和后续改进；目前没有发现新超导材料。历史诊断集上的总体误差改善很小，置信区间跨零，正温度和高温材料的误差反而增加，且一个所选描述符未通过旋转不变性检查。具体流程图、代码、结果和下一步修正方案见 GitHub 汇报。