Starrydata2: Curated Scientific Data Platform

• Starrydata2 is a curated, LLM-assisted infrastructure that aggregates experimental thermoelectric and astronomical datasets for advanced data mining.
• It offers unmatched scale with nearly 50,000 samples and over 130,000 data curves from thousands of publications, ensuring statistical rigor.
• Its integration of automated extraction and domain-specific filtering enables reproducible benchmarking and facilitates experimental validation.

Searching arXiv for papers mentioning Starrydata2 to ground the article in the latest relevant literature. I’m checking arXiv for direct mentions of “Starrydata2,” including thermoelectric dataset work and related curation papers. Starrydata2 is described in the recent thermoelectric literature as the world’s largest thermoelectric database and, more specifically, as an LLM-assisted, plot-based extraction resource for experimental transport data. In parallel, several astronomy data-release papers explicitly frame their products as relevant to “Starrydata2” or to “Starrydata2”-style access, indicating a broader association with curated, queryable scientific data products. Taken together, the literature presents Starrydata2 as a data infrastructure concept centered on large-scale, machine-usable scientific measurements, with its clearest direct realization in thermoelectric materials informatics and a secondary use as a reference model for astronomical catalog integration (Kumari et al., 8 Jul 2025, Athar et al., 21 Dec 2025, Yang et al., 2023).

1. Thermoelectric database identity

In thermoelectric research, Starrydata2 is explicitly characterized as a large experimental database assembled from the published literature. One study reports that, as of December 2021, Starrydata2 contained about 43,601 sample property sets from 7,994 publications, and that filtering the raw GitHub data yielded a validated subset of 13,469 sample datasets from 3,142 unique publications. Each valid sample contained temperature-dependent thermoelectric properties including the Seebeck coefficient $S$, electrical conductivity $\sigma$, thermal conductivity $\kappa$, and figure of merit $zT$, after which the dataset was separated into p-type and n-type classes according to the sign of the Seebeck coefficient (Kumari et al., 8 Jul 2025).

A later curation study describes the resource at a larger scale, stating that Starrydata2 contains roughly 50,000 physical samples, data from around 10,000 publications, and more than 130,000 curve traces spanning Seebeck coefficient, electrical conductivity, thermal conductivity, and $zT$. That study also clarifies the underlying data model: Starrydata2 is an LLM-assisted, plot-based extraction database in which numerical values are mined from published figures and associated with a sample ID, figure ID, DOI, and the reported temperature dependence of the property (Athar et al., 21 Dec 2025).

This dual description is significant. The first paper emphasizes Starrydata2 as a source for statistically filtered, composition-aware benchmarking, whereas the second emphasizes it as a raw, large-scale corpus whose size is attractive for machine learning but whose automatic curation pipeline introduces nontrivial reliability constraints. A plausible implication is that Starrydata2 should be understood not as a single “finished” benchmark table, but as an evolving experimental evidence base whose utility depends on downstream filtering and task-specific validation (Kumari et al., 8 Jul 2025, Athar et al., 21 Dec 2025).

2. Composition-space analysis and reference selection

A central use of Starrydata2 is the identification of reference compositions that are representative of the literature rather than merely high performing. In the bismuth telluride study, the authors restricted attention to BiTe-based systems and used two statistical tools: frequency counting to identify commonly reported third elements, and 2D Kernel Density Estimation (KDE) with a Gaussian kernel and automatically chosen bandwidth to locate the most densely populated region of composition space. Only compositions in which a third-element atomic fraction was 10\% or greater were counted as relevant substitutions (Kumari et al., 8 Jul 2025).

Within the validated dataset, the p-type subset contained 721 material compositions. Among these, Sb appeared in 511 of 721 p-type compositions, identifying BiSbTe as the relevant ternary host system. KDE over the Bi–Sb distribution yielded a highest-density point at Bi: 9.29 at.\% and Sb: 30.71 at.\%, corresponding to $\mathrm{Bi_{0.46}Sb_{1.54}Te_3}$, which the paper identifies as the most frequently observed p-type composition. The n-type subset contained 510 material compositions; Se appeared in 248 of 510 n-type compositions, identifying BiTeSe as the n-type host system. KDE over the Te–Se distribution produced a densest point at Te: 54.09 at.\% and Se: 6.06 at.\%, corresponding to $\mathrm{Bi_2Te_{2.7}Se_{0.3}}$, more precisely $\mathrm{Bi_2Te_{2.71}Se_{0.29}}$, as the most frequently studied n-type reference composition (Kumari et al., 8 Jul 2025).

The same paper formalizes host-versus-dopant character in quaternary systems through the definitions

$\mathrm{HF} := \frac{x+y+z}{x+y+z+\delta}$

and

$\mathrm{DF} := \frac{\delta}{x+y+z+\delta},$

where $\sigma$0 are host-element stoichiometric amounts and $\sigma$1 is the dopant amount. For both the p-type and n-type families, compositions with host fraction $\sigma$2 were treated as close to the reference ternary host. This matters because the selected compositions were justified as modal compositions in the literature database: they occur most often, lie at the high-density center of the distribution, and are therefore positioned as reproducible and well benchmarked rather than simply optimal in peak $\sigma$3 (Kumari et al., 8 Jul 2025).

3. Experimental validation and benchmark module design

The reference-composition study used Starrydata2-derived compositions as the basis for direct experimental validation. The selected materials were $\sigma$4 for p-type and $\sigma$5 for n-type. Synthesis proceeded by melting high-purity Bi, Sb, Te, and Se, all >99.99\% purity, in 30 g batches sealed in quartz ampoules under Ar at 0.5 atm, heating at 800°C for 10 h in a rocking tube furnace, and then water quenching. Subsequent ball milling used a stainless steel jar and 10 mm balls, a 1:15 material-to-ball ratio, and 300 rpm for 1 h, followed by sieving to <45 μm with a 325-mesh sieve. Consolidation was then carried out by hot pressing (HP) and spark plasma sintering (SPS) at 420°C, 50 MPa, and 5 min under argon (Kumari et al., 8 Jul 2025).

Because $\sigma$6-based materials are strongly anisotropic, the study measured transport properties along the A-axis (perpendicular to pressing direction) and C-axis (parallel to pressing direction). Temperature-dependent $\sigma$7, $\sigma$8, thermal diffusivity, $\sigma$9, power factor, and $\kappa$0 were measured from room temperature to 300°C in 50°C increments, using ULVAC ZEM-3 for Seebeck coefficient and electrical conductivity and NETZSCH Laser Flash Analysis LFA 447 for thermal diffusivity. Thermal conductivity was calculated as

$\kappa$1

with $\kappa$2 estimated from the Dulong–Petit limit as 0.19 J g$\kappa$3 K$\kappa$4 for p-type and 0.157 J g$\kappa$5 K$\kappa$6 for n-type. The electronic component was estimated via

$\kappa$7

with Lorenz number $\kappa$8 to $\kappa$9 W·$zT$0·K$zT$1, and the lattice contribution was obtained from

$zT$2

The experimentally measured properties were then overlaid on Starrydata2 KDE plots; the reported result is that the selected compositions lie near the high-density region for $zT$3, $zT$4, $zT$5, and $zT$6, especially for SPS A-axis data (Kumari et al., 8 Jul 2025).

The same work reports concrete benchmark outcomes. For $zT$7, the maximum $zT$8 was about 1.2, achieved around 100°C in SPS A-axis; for $zT$9, the maximum $zT$0 was about 0.9, achieved at 150°C in SPS A-axis. The paper further estimated module-level performance with a 200-pair thermoelectric generator module (TGM) modeled in COMSOL Multiphysics using a simplified 2D uni-couple finite element model and HP A-axis material properties. In the realistic Model F, which included Cu electrodes, Ni diffusion barriers, solder layers, and Al$zT$1O$zT$2 substrates, the benchmark at $zT$3 K with hot side 150°C and cold side 30°C yielded $zT$4 V, $zT$5 W, $zT$6 W, and $zT$7. The idealized Model A at the same $zT$8 gave $zT$9 W and $\mathrm{Bi_{0.46}Sb_{1.54}Te_3}$0, illustrating the impact of contact and interfacial losses (Kumari et al., 8 Jul 2025).

4. Reliability constraints and curation methodology

The strongest methodological caution in the literature is that Starrydata2’s size does not make it automatically suitable for robust machine learning. The half-Heusler curation study argues that the database contains inaccuracies arising from LLM-assisted extraction, ambiguous nomenclature, complex formulas and misidentified compositions, multi-source experimental data for the same composition, high standard deviations and non-normal spread, and missing synthesis information. The paper explicitly states that these issues “cannot be filtered with conventional dataset curation workflows” (Athar et al., 21 Dec 2025).

Its case studies illustrate distinct failure modes. For TiNiSn, Starrydata2 contains multiple $\mathrm{Bi_{0.46}Sb_{1.54}Te_3}$1 curves with very different $\mathrm{Bi_{0.46}Sb_{1.54}Te_3}$2 values; the paper contrasts a source with $\mathrm{Bi_{0.46}Sb_{1.54}Te_3}$3 attributed to highly optimized synthesis against another with $\mathrm{Bi_{0.46}Sb_{1.54}Te_3}$4 for a sample lacking a standard densification step. For NbFeSb, curves from doped compounds such as $\mathrm{Bi_{0.46}Sb_{1.54}Te_3}$5 are reported as if they belonged to pure NbFeSb because of host-system naming in the source literature. For TiCoSb, source-figure misassignment caused some TiNiSn curves to be assigned to TiCoSb, creating a bimodal spread in apparent performance. The paper also distinguishes $\mathrm{Bi_{0.46}Sb_{1.54}Te_3}$6, the arithmetic mean of instantaneous $\mathrm{Bi_{0.46}Sb_{1.54}Te_3}$7 over a temperature range, from $\mathrm{Bi_{0.46}Sb_{1.54}Te_3}$8, the engineering figure of merit, emphasizing that these are not interchangeable with $\mathrm{Bi_{0.46}Sb_{1.54}Te_3}$9 (Athar et al., 21 Dec 2025).

To address these problems, the authors propose a round-robin bin filtering method based on an interlaboratory uncertainty of approximately $\mathrm{Bi_2Te_{2.7}Se_{0.3}}$0 for half-Heuslers. The procedure begins from the working assumption that the average $\mathrm{Bi_2Te_{2.7}Se_{0.3}}$1 is close to the representative maximum $\mathrm{Bi_2Te_{2.7}Se_{0.3}}$2 of a composition. Reported $\mathrm{Bi_2Te_{2.7}Se_{0.3}}$3 values are binned using a width equal to $\mathrm{Bi_2Te_{2.7}Se_{0.3}}$4 of $\mathrm{Bi_2Te_{2.7}Se_{0.3}}$5; bins are then ranked by the number of unique DOIs they contain, ties are resolved by the number of full $\mathrm{Bi_2Te_{2.7}Se_{0.3}}$6 curves, and a representative DOI and reference pure-material curve are chosen accordingly. Additional doped-material curves are retained only under specific consistency conditions. The paper is explicit that this method entails some data loss, that the 15\% threshold is domain specific, and that the procedure depends on an initial assumption that may fail if the dataset is heavily contaminated (Athar et al., 21 Dec 2025).

The same study therefore advocates a hybrid dataset creation workflow, termed ICGM, combining manual extraction using PlotDigitizer with filtered Starrydata2 entries. The workflow retains only densified bulk samples, specifically HP or SPS materials, removes phase-separated or multiphasic materials, applies the round-robin filter, and limits each curve to at most six experimental $\mathrm{Bi_2Te_{2.7}Se_{0.3}}$7 points without interpolation or extrapolation. Quantitatively, the resulting half-Heusler subset contains 108 materials and 256 compositions, compared with 132 materials and 267 compositions in the Starrydata2 subset, while preserving nearly the same UMAP chemical space. The paper’s conclusion is unambiguous: Starrydata2 is indispensable as a starting resource, but it should not be used blindly as a finished training set for thermoelectric ML (Athar et al., 21 Dec 2025).

5. Astronomical data products explicitly framed as Starrydata2-relevant

The astronomy papers in the supplied literature do not define Starrydata2 as an astronomical survey. They do, however, explicitly present several catalog releases as relevant to Starrydata2 or to “Starrydata2”-style access, which suggests a second usage centered on curated stellar photometry and derived metadata.

Resource	Core product	Reported relevance to Starrydata2
AST3-2 survey release (Yang et al., 2023)	$\mathrm{Bi_2Te_{2.7}Se_{0.3}}$8-band source catalogue, light-curve catalogue, preprocessed FITS images	“directly useful for ‘Starrydata2’-style access to time-domain stellar data”
Bright-star Sloan catalog v2 (Mallama, 2018)	Sloan magnitudes for the brightest stars	“highly relevant to Starrydata2” as a reference source for bright-star Sloan photometry
Bayestar17 (Green et al., 2018)	3D Galactic reddening map	“the product most relevant to Starrydata2”

The AST3-2 release is the clearest example of this astronomical framing. It presents the 2016 supernova-survey season of the Antarctic Survey Telescope AST3-2, observed from 2016 March 24 to May 16, spanning 565 sky fields over about 2200 deg$\mathrm{Bi_2Te_{2.7}Se_{0.3}}$9 with roughly 30 visits per field at cadences of a half to a few days. The release consists of 22000 scientific images and photometric time-series products, while the FITS image archive contains 22576 preprocessed images. It includes an $\mathrm{Bi_2Te_{2.71}Se_{0.29}}$0-band source catalogue with over 7 million sources, a light-curve catalogue with per-epoch measurements after quality filtering, and photometric performance summarized by a median $\mathrm{Bi_2Te_{2.71}Se_{0.29}}$1 limiting magnitude of 17.8 mag in $\mathrm{Bi_2Te_{2.71}Se_{0.29}}$2 band and 4 mmag light-curve precision for bright stars. From these data, over 3,500 variable stars were detected, including 70 newly discovered variables classified into LPV, PUL, EC, ROT, and pROT groups. The paper explicitly states that this scale and structure make the release directly useful for “Starrydata2”-style access to time-domain stellar data (Yang et al., 2023).

The other two astronomy resources play complementary roles. The bright-star Sloan catalog provides $\mathrm{Bi_2Te_{2.71}Se_{0.29}}$3 magnitudes for stars between $\mathrm{Bi_2Te_{2.71}Se_{0.29}}$4 and $\mathrm{Bi_2Te_{2.71}Se_{0.29}}$5, along with $\mathrm{Bi_2Te_{2.71}Se_{0.29}}$6 and $\mathrm{Bi_2Te_{2.71}Se_{0.29}}$7, and is described as a reliable source of comparison-star magnitudes for astronomical photometry and as a bright-star extension to APASS. Its stated uncertainties are 0.02 mag for Johnson-Cousins colors, 0.03 mag for Sloan bands except $\mathrm{Bi_2Te_{2.71}Se_{0.29}}$8, which is 0.08 mag (Mallama, 2018). Bayestar17, by contrast, is a probabilistic 3D Galactic reddening map built from 806 million PS1+2MASS sources divided among 3.42 million sightlines, covering sky north of $\mathrm{Bi_2Te_{2.71}Se_{0.29}}$9 with typical angular resolutions from $\mathrm{HF} := \frac{x+y+z}{x+y+z+\delta}$0 to $\mathrm{HF} := \frac{x+y+z}{x+y+z+\delta}$1; after a 15\% renormalization of Planck14, it agrees with far-infrared reddening maps at the $\mathrm{HF} := \frac{x+y+z}{x+y+z+\delta}$2 level out to a depth of 0.8 mag in the comparison reddening measure (Green et al., 2018). This suggests that, in the astronomy context, Starrydata2 is being invoked as a query-oriented layer that would benefit from calibrated source catalogues, time-series measurements, bright-star photometric standards, and distance-dependent extinction information.

6. Interpretation, misconceptions, and scientific role

One common misconception is that Starrydata2’s primary value lies simply in its size. The literature does not support that view. In thermoelectrics, the database is valuable precisely because it is large and experimentally grounded, but the same papers stress that data volume and chemical diversity are insufficient without data quality, composition verification, and synthesis-aware filtering. The half-Heusler curation study therefore treats Starrydata2 as a raw source to be curated rather than as an immediately reliable ML benchmark (Athar et al., 21 Dec 2025).

A second misconception is that the reference BiTe compositions were chosen because they deliver the highest reported performance. The bismuth telluride study states the opposite: $\mathrm{HF} := \frac{x+y+z}{x+y+z+\delta}$3 and $\mathrm{HF} := \frac{x+y+z}{x+y+z+\delta}$4 were selected because they are the modal compositions in the literature, identified by KDE as the most densely populated regions of composition space. Their subsequent experimental validation matters because the measured properties align with the median or central literature tendency, not because they define a universal upper bound (Kumari et al., 8 Jul 2025).

A third misconception is that Starrydata2 refers only to thermoelectric materials. The supplied literature suggests a narrower direct meaning and a broader analogical one. Directly, Starrydata2 is a thermoelectric database with sample-level and curve-level data structures. More broadly, astronomy papers invoke it as the kind of platform for which large calibrated survey archives, bright-star reference photometry, and 3D reddening maps are directly useful. A plausible implication is that Starrydata2 functions, across domains, as a model of curated scientific data infrastructure whose scientific utility depends on explicit metadata, statistical traceability, and compatibility with downstream analysis rather than on simple record count alone (Yang et al., 2023, Mallama, 2018, Green et al., 2018, Athar et al., 21 Dec 2025).