21,263 superconducting materials, each given by its chemical formula and one reported critical temperature, with 81 features computed from the formula. Compiled by Kam Hamidieh from the SuperCon database and released by the UCI Machine Learning Repository.
This page follows the data from the literature to the 21,263 × 81 feature matrix. There is no raw instrument output: the steps up to UCI's release are described by Hamidieh and UCI, and the 81 features are rebuilt here from each formula and checked against the released file.
Kam Hamidieh.
There is no single instrument. Each critical temperature is a laboratory measurement reported in the literature and compiled into SuperCon; the release does not record how, or under what conditions, each value was measured.
From the laboratory to the held files
- DocumentedLaboratory measurements
Critical temperatures reported in journal articles
- DocumentedSuperCon (NIMS)
Japan’s National Institute for Materials Science compiles the reports into a database
- DocumentedHamidieh’s snapshot
Downloaded 24 July 2017: 31,611 rows, cleaned to 21,263
- ReproducedFeature recipe
Formula → 81 features; rebuilt here for every material
- ReproducedUCI release
Two files: formulas and features; the held features file is byte-identical
- HeldHeld files
21,263 rows: features (82 columns) and formulas
Labels: Documented means described by Hamidieh (2018) and UCI; Reproduced means rerun or checked on this page; Held means the files this page reads.
One acquisition unit: one material
A record is one material: a chemical formula and one reported critical temperature.
What each record holds
| File | Column(s) | Unit | Meaning |
|---|---|---|---|
| formulas | 86 element columns, H … Rn | — | Amount of each element in the formula (0 if absent) |
| formulas | material | — | The formula as text |
| both | critical_temp | K | Reported critical temperature |
| features | number_of_elements | — | Number of distinct elements in the formula |
| features | 80 columns: 10 statistics × 8 properties | property unit | Summaries of the elements’ atomic mass, first ionisation energy, atomic radius, density, electron affinity, heat of fusion, thermal conductivity, valence |
The files, opened
The formulas file, first rows (elements with amount 0 omitted):
| material | critical_temp (K) | elements |
|---|---|---|
| Ba0.2La1.8Cu1O4 | 29 | O 4 Cu 1 Ba 0.2 La 1.8 |
| Ba0.1La1.9Ag0.1Cu0.9O4 | 26 | O 4 Cu 0.9 Ag 0.1 Ba 0.1 La 1.9 |
| Ba0.1La1.9Cu1O4 | 19 | O 4 Cu 1 Ba 0.1 La 1.9 |
| Ba0.15La1.85Cu1O4 | 22 | O 4 Cu 1 Ba 0.15 La 1.85 |
| Ba0.3La1.7Cu1O4 | 23 | O 4 Cu 1 Ba 0.3 La 1.7 |
| Ba0.5La1.5Cu1O4 | 23 | O 4 Cu 1 Ba 0.5 La 1.5 |
The features file, first rows (the first 9 of 81 feature columns):
| number_of_elements | mean_atomic_mass | wtd_mean_atomic_mass | gmean_atomic_mass | wtd_gmean_atomic_mass | entropy_atomic_mass | wtd_entropy_atomic_mass | range_atomic_mass | wtd_range_atomic_mass | critical_temp |
|---|---|---|---|---|---|---|---|---|---|
| 4 | 88.9445 | 57.8627 | 66.3616 | 36.1166 | 1.1818 | 1.0624 | 122.906 | 31.7949 | 29 |
| 5 | 92.7292 | 58.5184 | 73.1328 | 36.3966 | 1.4493 | 1.0578 | 122.906 | 36.1619 | 26 |
| 4 | 88.9445 | 57.8852 | 66.3616 | 36.1225 | 1.1818 | 0.976 | 122.906 | 35.7411 | 19 |
| 4 | 88.9445 | 57.874 | 66.3616 | 36.1196 | 1.1818 | 1.0223 | 122.906 | 33.768 | 22 |
| 4 | 88.9445 | 57.8401 | 66.3616 | 36.1107 | 1.1818 | 1.1292 | 122.906 | 27.8487 | 23 |
| 4 | 88.9445 | 57.795 | 66.3616 | 36.0989 | 1.1818 | 1.2252 | 122.906 | 20.6875 | 23 |
Critical temperatures in the release
Critical temperatures of all 21,263 materials, on a logarithmic scale; each bar counts the materials in one temperature range (hover to read it). 58% are below 30 K; the rest reach 185 K.
What the release covers
Two charts: how many materials contain each element (the 30 most common of the 77 elements), and how many formulas have each number of elements.
What the raw record does not give you directly
- The elemental property values are not in the release. The features depend on eight properties per element, but no table of them is published with the data. This page recovers it from the features themselves.
- 6,093 rows repeat another row’s features exactly, and 66 also repeat its critical temperature. Formulas with the same elements in the same proportions get the same features.
- Critical temperature spans six orders of magnitude, from about 0.0002 K to about 185 K.
- Measurement conditions are not recorded: no method, pressure or sample quality per value.
Steps 1–2, Hamidieh’s selection and cleaning, are described in his paper; the SuperCon snapshot they start from is not part of the release. Steps 3–7, the feature recipe, are rebuilt here from each formula and compared with the released features.
| Step | Evidence | Operation | Result |
|---|---|---|---|
| 1 | Documented | Hamidieh downloads SuperCon’s oxide and metallic collection on 24 July 2017 | 31,611 rows |
| 2 | Documented | Cleaning: fix temperature columns, remove missing or zero temperatures and problem formulas, approximate oxygen content, drop elements above atomic number 86, remove repeated rows | 21,263 materials |
| 3 | Reproduced | Split each formula into element proportions p (amount ÷ total amount) | 1–9 elements per material |
| 4 | Reproduced | Give each element eight property values t; recovered here from the file, one value per element and property | 77 elements × 8 properties |
| 5 | Reproduced | For each property compute ten statistics over the material’s elements: mean, weighted mean, geometric and weighted geometric mean, entropy and weighted entropy, range and weighted range, standard deviation and weighted standard deviation | 80 features |
| 6 | Reproduced | Add the number of elements | 81 features |
| 7 | Reproduced | Keep the critical temperature apart as the target | 21,263 × 81 matrix and target |
The ten statistics
| Statistic | Definition (p = proportion, t = property value, over the material’s elements) |
|---|---|
| mean / weighted mean | average of t / Σ p·t |
| geometric / weighted geometric mean | (Π t)^(1/n) / Π t^p |
| entropy / weighted entropy | −Σ w ln w with w = t / Σ t / the same with w ∝ p·t |
| range / weighted range | max t − min t / max(p·t) − min(p·t) |
| standard deviation / weighted | population SD of t / √Σ p (t − weighted mean)² |
Where the records go
The drop from 31,611 to 21,263 is Hamidieh’s cleaning, as described in his paper. After the release nothing is dropped: every material becomes a row of the matrix.
Follow one material
Pick a material and a property to see the recipe at each step, from the formula to the row the matrix stores. The last table compares the ten statistics rebuilt here with the values in the released file.
Checks against the release
| Check | Result |
|---|---|
| Held features file vs UCI’s published file | byte-identical |
| Critical temperature, formulas file vs features file | identical for every row |
| Number of elements, from the formula | 100% of rows identical |
| Features rebuilt from formula and recovered property table | 81 of 81 identical for all 21,263 materials |
| Recovered atomic masses (sample) | H 1.0079, O 15.9994, Cu 63.546, Nb 92.9064 |
The property table is recovered from the unweighted means, which fix one value per element; the other nine statistics of each property then serve as an independent check. The recovered atomic masses equal the standard atomic weights. In the released file, the weighted range is the spread of p·t across the elements.
Each row is one material, ordered here by critical temperature (labels on the left mark 1, 10, 30 and 90 K); each column is one feature. Features have different units, so colour shows each value’s percentile within its column. Hover to read a row: its formula and temperature appear under the chart.
Colour is a display transform. The matrix itself holds the released values in their own units.
81 features: number_of_elements and ten statistics (mean_, wtd_mean_, gmean_, wtd_gmean_, entropy_, wtd_entropy_, range_, wtd_range_, std_, wtd_std_) of each property below. Entropies are dimensionless; the other statistics carry the property’s unit.
| Property | Column suffix | Unit |
|---|---|---|
| Atomic mass | …_atomic_mass | amu |
| First ionisation energy | …_fie | kJ mol⁻¹ |
| Atomic radius | …_atomic_radius | pm |
| Density | …_Density | kg m⁻³ |
| Electron affinity | …_ElectronAffinity | kJ mol⁻¹ |
| Heat of fusion | …_FusionHeat | kJ mol⁻¹ |
| Thermal conductivity | …_ThermalConductivity | W m⁻¹ K⁻¹ |
| Valence | …_Valence | — |
Kept beside the matrix
| Column | Unit | Meaning |
|---|---|---|
critical_temp | K | The target: reported critical temperature |
material | — | The formula, from the formulas file; an identifier, never an input |
UCI distributes Superconductivty Data under Creative Commons Attribution 4.0 International (CC BY 4.0). Credit Kam Hamidieh and UCI; this page adds descriptive visualizations.
- The elemental property values are recovered from the released features, not taken from Hamidieh’s source table. They reproduce every released feature, but individual values are not checked against an independent reference here.
- 6,093 rows share all 81 features with another row, so a random train/test split puts identical inputs on both sides.
- Hamidieh’s cleaning (31,611 → 21,263 rows) is documented, not reproduced; the SuperCon snapshot is not held.
- No measurement conditions accompany the critical temperatures.
Hamidieh, K. (2018). Superconductivty Data. UCI Machine Learning Repository. DOI 10.24432/C53P47. Method: A data-driven statistical model for predicting the critical temperature of a superconductor, Computational Materials Science (2018), arXiv:1803.10260.
Superconductor regression explorer: regression of critical temperature on the 81 features; and its Insight page.