THEIA: A Multifaceted Scientific Term

• THEIA is a designation used in various fields, including astronomy where it describes a Gaia-discovered stellar structure and a tidally disrupted star cluster.
• In planetary science, THEIA represents the Moon-forming impactor, with studies detailing its isotopic, dynamical, and compositional characteristics.
• THEIA also names proposed experiments in neutrino detection, high-precision astrometry missions, and advanced AI systems, reflecting its wide interdisciplinary impact.

THEIA is a recurrent designation applied to several unrelated research entities across astronomy, planetary science, detector physics, astrometry, medicine, and computer science. In recent arXiv literature, the name denotes a Gaia-discovered stellar structure in the Milky Way, the Moon-forming impactor in Giant Impact studies, a proposed hybrid optical neutrino detector, a proposed high-precision astrometry mission, and multiple software or AI systems for medical screening, robot learning, neural reasoning, deep-learning debugging, and mobile distributed search (Tregoning et al., 2024, Branco et al., 2 Jul 2025, Askins et al., 2022, Malbet et al., 2022, Vaghefi et al., 2021). The meaning of “Theia” is therefore entirely context-dependent.

1. Domain-specific uses of the name

Across the cited literature, the designation does not refer to a single scientific object or platform.

Domain	“Theia” denotes	Representative source
Galactic astronomy	Theia 456 / COIN-Gaia-13, a Gaia-discovered stellar structure	(Tregoning et al., 2024)
Planetary science	The Moon-forming impactor, or the last giant impactor onto Earth analogues	(Branco et al., 2 Jul 2025)
Neutrino physics	A hybrid Cherenkov/scintillation detector concept based on WbLS	(Askins et al., 2022)
Space astrometry	A pointed, differential astrometry mission concept beyond Gaia	(Malbet et al., 2022)
Medicine and computing	Several unrelated AI or software systems named THEIA	(Vaghefi et al., 2021)

This multiplicity is explicit in the literature. For example, the stellar-dynamics study of Theia 456 states that in that paper “Theia” refers to a Gaia-discovered stellar structure rather than the planet-formation impactor, while planetary-science papers use the same name for the body involved in the Moon-forming collision (Tregoning et al., 2024, Branco et al., 2 Jul 2025).

2. Theia 456 in Galactic stellar dynamics

In Galactic astronomy, Theia 456 is a nearby, young, low-density stellar structure in the Milky Way thin disk, also known as COIN-Gaia-13. It was first identified in Gaia clustering searches, with Cantat-Gaudin et al. designating it COIN-Gaia-13 and Kounkel & Covey subsequently including it in their Gaia “Theia” catalog as Theia 456 (Tregoning et al., 2024). Previous work had already shown coherence in kinematics, chemistry, and gyrochronology, and the 2024 dynamical study strengthened that case with precise follow-up spectroscopy.

The structure is currently very extended and only loosely bound at best. The final membership catalog contains 321 stars, of which 43 have full 6D phase-space information and 278 have Gaia 5D astrometry only. It spans roughly 120 pc, about $20^\circ$ on the sky, at a distance of about 500 pc, and morphologically consists of two overdense lobes connected by a diffuse bridge. Only about $\sim 20\%$ of stars lie inside the quoted Jacobi radius, so the majority of the system is effectively unbound. The interpretation advanced in the paper is therefore not that of a classical compact open cluster today, but of a dispersed, tidally disrupted remnant.

The central methodological advance was the acquisition of MMT/Hectochelle spectroscopy across six fields. The spectra, modeled with MINESweeper, delivered radial velocities with median precision of about $\sim 0.1\ {\rm km\,s^{-1}}$, compared with the much poorer Gaia DR2 and LAMOST radial-velocity precisions available previously. Orbit integration with gala in a Milky Way potential showed that the two lobes, now about 100 pc apart, undergo a strong convergence at about 250 Myr in the past, where the lobe separation shrinks to about $\simeq 20$ pc.

A Bayesian forward model treated the birth configuration as a spherically symmetric 3D Gaussian in phase space with parameter vector

$$\psi_{0} : \{X, Y, Z, V_X, V_Y, V_Z, \sigma_v, R, T\}.$$

From this, the study derived a kinematic age of $245 \pm 3\ {\rm Myr}$, an initial Gaussian scale $R = 6 \pm 1\ {\rm pc}$, an initial one-dimensional velocity dispersion $0.14 \pm 0.02\ {\rm km\,s^{-1}}$, and an inferred half-mass radius of $9 \pm 2\ {\rm pc}$. Using a Kroupa IMF and Gaia completeness for $G<17$, the authors estimated an initial total population of about $\sim 20\%$0 stars and total stellar mass $\sim 20\%$1, implying a characteristic birth density of only about $\sim 20\%$2.

The astrophysical significance of Theia 456 lies in its interpretation as a system disrupted by Galactic tides before full virialization. The paper estimates a tidal disruption timescale of about $\sim 20\%$3 and a crossing time of about 50 Myr, consistent with the claim that Galactic tides can disperse such a diffuse cluster before it dynamically relaxes. The dynamical age estimate is presented as independent of isochrones, stellar-evolution details, and internal cluster dynamics, with statistical precision comparable to the most precise age-dating techniques currently available (Tregoning et al., 2024).

3. Theia as the Moon-forming impactor

In planetary science, Theia denotes the impactor involved in the final giant collision that formed the Moon, or, in accretion simulations, the body involved in the final embryo-scale collision onto the Earth analogue (Branco et al., 2 Jul 2025). This usage anchors a long-running debate over Theia’s origin, composition, and the extent to which the Moon-forming impact can explain the Earth–Moon isotopic similarity.

A recent dynamical study modeled terrestrial planet formation with a narrow annulus of non-carbonaceous material and an added population of carbonaceous bodies scattered inward by Jupiter. In the favored mixed scenario, Earth’s last giant impactor contains a carbonaceous component in about half of viable systems: 38.5% of cases as a pure carbonaceous embryo and 13.5% as a non-carbonaceous embryo that had previously accreted a carbonaceous embryo (Branco et al., 2 Jul 2025). The same study argues that the scenario works best if the inward-scattered carbonaceous reservoir had total mass $\sim 20\%$4, an embryo-to-planetesimal mass ratio of at least 8, and carbonaceous embryo masses in the $\sim 20\%$5 range. In that framework, Theia is plausibly, but not certainly, carbonaceous.

Isotopic and chemical constraints point in a different direction when the Giant Impact is analyzed via source mixing. Meier et al. parameterized the difference in proto-Earth contribution to Earth and Moon through

$\sim 20\%$6

and showed that the canonical impact corresponds to large $\sim 20\%$7, whereas high-angular-momentum models permit much smaller values (Meier et al., 2014). Their representative ranges are about $\sim 20\%$8 for the canonical case, average $\sim 20\%$9 for hit-and-run, average $\sim 0.1\ {\rm km\,s^{-1}}$0 for impact-fission, and average $\sim 0.1\ {\rm km\,s^{-1}}$1 for merger. In isotopic terms, the high-angular-momentum models allow, by a narrow margin, a CI-chondritic or Mars-like Theia. However, once the Earth–Moon mantle FeO difference is included, the paper argues that the simplest reconciliation is an Earth-like isotopic composition for Theia together with a higher mantle FeO content of about $\sim 0.1\ {\rm km\,s^{-1}}$2 (Meier et al., 2014).

A short 2015 follow-up on Theia’s accretion reanalyzed two apparently conflicting simulation studies and argued that much of the discrepancy came from analog-selection choices rather than from different simulations. Under the analysis favored there, the fraction of Theia analogs consistent with the canonical Giant Impact hypothesis remains in the $\sim 0.1\ {\rm km\,s^{-1}}$3 range after accounting for simulation granularity uncertainty and terrestrial contribution to lunar material (Kaib et al., 2015). That result supports the view that a canonical, isotopically Earth-like Theia is statistically disfavored but not excluded.

A distinct 2026 proposal addresses the lunar isotopic crisis through rheology rather than source composition. In that work, Theia is modeled as a high-viscosity body with

$\sim 0.1\ {\rm km\,s^{-1}}$4

colliding with a low-viscosity magma-ocean proto-Earth of mass $\sim 0.1\ {\rm km\,s^{-1}}$5; the impactor mass is $\sim 0.1\ {\rm km\,s^{-1}}$6, the impact angle is $\sim 0.1\ {\rm km\,s^{-1}}$7, and the impact speed is set to the mutual escape speed (Liu, 18 Jun 2026). The presented SPH run yields debris that is about 70% proto-Earth-derived after 1.68 hr, suggesting a route to an Earth-like Moon without invoking an initially high-angular-momentum Earth–Moon system. The paper treats this as a possible resolution of the lunar isotopic crisis, not as a complete final solution.

4. Theia as a hybrid optical neutrino detector

In neutrino physics, Theia is a proposed large underground detector intended to combine the strengths of water Cherenkov and liquid scintillator technologies in a single hybrid optical platform (Askins et al., 2022). The central medium is water-based liquid scintillator (WbLS), with separation of prompt directional Cherenkov light and delayed isotropic scintillation light enabled by fast timing, angular reconstruction, and in some configurations spectral separation. The literature describes several representative scales, including a 50-kiloton concept, Theia-25, and Theia-100 (Fischer, 2018, Gann, 2015).

The detector concept is explicitly multipurpose. The 2022 Snowmass summary describes physics goals spanning long-baseline oscillations and CP violation, mass ordering, solar neutrinos, supernova burst neutrinos, the diffuse supernova neutrino background (DSNB), nucleon decay, neutrinoless double beta decay, geo-neutrinos, reactor antineutrinos, and sterile-neutrino searches (Askins et al., 2022). Representative instrumentation in the Theia-100 simulations includes 86% coverage with standard 10-inch PMTs and 4% coverage with LAPPDs, while the earlier 50-kiloton concept proposed more than 100,000 photosensors and more than 90% effective photocoverage (Askins et al., 2022, Fischer, 2018). The same concept papers emphasize tunable WbLS composition, with lower scintillator fractions for beam physics and higher fractions or isotope-loaded inner volumes for low-energy and rare-event programs (Gann, 2015, Fischer, 2018).

Quantitative reach estimates are extensive. The Snowmass summary quotes $\sim 0.1\ {\rm km\,s^{-1}}$8 sensitivity to CP violation for 30% of $\sim 0.1\ {\rm km\,s^{-1}}$9 values with 524 kt-MW-yr, supernova pointing of $\simeq 20$0 for 100 kt and $\simeq 20$1 for 25 kt at 10 kpc, and expected supernova burst counts of 20,000 events in Theia-100 and 5,000 in Theia-25 (Askins et al., 2022). For neutrinoless double beta decay, it quotes

$\simeq 20$2

for tellurium and

$\simeq 20$3

for xenon. The proton-decay benchmark for $\simeq 20$4 is

$\simeq 20$5

for 800 kton-yr (Askins et al., 2022).

The DSNB program is one of the best-developed cases. A dedicated 2020 study of WbLS-based Theia configurations finds that the full analysis chain—IBD-like coincidence selection, cosmogenic veto, fiducial cut, Cherenkov ring counting, Cherenkov/scintillation-ratio cut, and delayed-decay veto—can deliver signal efficiency above 80% and reduce the residual atmospheric neutral-current background to about 1.3% of its initial level (Sawatzki et al., 2020). Under the paper’s fiducial DSNB model, a $\simeq 20$6 DSNB discovery is achievable with about $\simeq 20$7 live exposure.

Low-energy antineutrino performance has also been quantified for a 17.8-ktonne fiducial Theia-25 configuration at SURF with $\simeq 20$8 free target protons. After one year, the expected fitted event counts are $\simeq 20$9 geoneutrinos and $$\psi_{0} : \{X, Y, Z, V_X, V_Y, V_Z, \sigma_v, R, T\}.$$0 reactor antineutrinos, with fit precisions of 6.72% and 8.55%, respectively (Zsoldos et al., 2022). The same study reports separate one-year fits of $$\psi_{0} : \{X, Y, Z, V_X, V_Y, V_Z, \sigma_v, R, T\}.$$1 and $$\psi_{0} : \{X, Y, Z, V_X, V_Y, V_Z, \sigma_v, R, T\}.$$2, corresponding to $$\psi_{0} : \{X, Y, Z, V_X, V_Y, V_Z, \sigma_v, R, T\}.$$3, and infers a mantle signal of

$$\psi_{0} : \{X, Y, Z, V_X, V_Y, V_Z, \sigma_v, R, T\}.$$4

or

$$\psi_{0} : \{X, Y, Z, V_X, V_Y, V_Z, \sigma_v, R, T\}.$$5

when systematic uncertainties are included (Zsoldos et al., 2022).

5. Theia as a high-precision astrometry mission

In space astronomy, Theia is a proposed visible-light astrometric observatory designed for high-precision differential astrometry in small pointed fields, at accuracies well beyond Gaia (Malbet et al., 2022). The current baseline described in the 2022 mission-profile study is a single-spacecraft, 0.8 m observatory using a Korsch three-mirror anastigmat at Sun–Earth L2, operating in the $$\psi_{0} : \{X, Y, Z, V_X, V_Y, V_Z, \sigma_v, R, T\}.$$6–$$\psi_{0} : \{X, Y, Z, V_X, V_Y, V_Z, \sigma_v, R, T\}.$$7 nm band over a nominal 4-year science mission plus about 6 months for transfer and commissioning. The 2017 M5 proposal presented a related point-and-stare concept in the $$\psi_{0} : \{X, Y, Z, V_X, V_Y, V_Z, \sigma_v, R, T\}.$$8–$$\psi_{0} : \{X, Y, Z, V_X, V_Y, V_Z, \sigma_v, R, T\}.$$9 nm range, with about 15% of mission time allocated to an open observatory (Collaboration et al., 2017).

The science case is organized around small-field relative astrometry rather than Gaia-like global scanning. The flagship themes are dark matter, nearby habitable terrestrial exoplanets, and compact-object astrophysics (Malbet et al., 2022, Collaboration et al., 2017). For dark matter, the mission is designed to constrain dwarf-spheroidal inner density profiles, detect kinematic perturbations from low-mass dark subhalos, and recover Milky Way halo axis ratios to 5% using hypervelocity stars (Malbet et al., 2022). For exoplanets, Theia is intended to survey about 60 nearby solar-type stars, with about 100 visits per target, and the paper quotes a median detectable mass across the full habitable zone of $245 \pm 3\ {\rm Myr}$0; combined with an occurrence rate estimate of

$245 \pm 3\ {\rm Myr}$1

this implies an expected yield of 9 to 57 true Earth-like planets (Malbet et al., 2022).

The measurement principle is differential calibration against reference stars in the same field. The detector-to-sky mapping is written as

$245 \pm 3\ {\rm Myr}$2

with coefficients fitted from reference-star positions tied to the Gaia frame (Malbet et al., 2022). A central conclusion of the mission-profile paper is that the ultimate astrometric accuracy can be met without drastic constraints on telescope stability, because field distortions can be solved from the reference stars themselves.

A laboratory precursor to this mission concept is DICE (“Detector Interferometric Calibration Experiment”), built to test whether detector calibration can support Theia’s centroiding requirement (Crouzier et al., 2016). The science driver there is Earth-like exoplanet astrometry, with a centroiding target of $245 \pm 3\ {\rm Myr}$3 pixel. The DICE testbed uses pseudo stars projected onto a CCD together with Young fringes from a metrology system. After upgrades, it achieved pixel-position calibration accuracy estimated at $245 \pm 3\ {\rm Myr}$4 pixel, astrometric accuracy of $245 \pm 3\ {\rm Myr}$5 pixel for PSF motion over more than 5 pixels, and $245 \pm 3\ {\rm Myr}$6 pixel in static mode with less than $245 \pm 3\ {\rm Myr}$7 pixel jitter (Crouzier et al., 2016). Those results are explicitly presented as a proof-of-concept step toward the mission requirement rather than its final fulfillment.

6. Computational and AI systems named THEIA

Several unrelated computational systems also use the name THEIA.

In medical AI, THEIA is a New Zealand-developed cloud-based clinical decision-support system for diabetic eye screening. In a prospective multi-center evaluation of 902 adults screened in both an urban DHB eye service and a semi-rural optometrist-led clinic, it achieved 100% sensitivity, 98.33% specificity, 98% accuracy, and 100% negative predictive value for patient-level referable disease, with no missed referable patients (Vaghefi et al., 2021). Agreement with the adjudicated gold standard was $245 \pm 3\ {\rm Myr}$8, compared with clinician-against-aggregate kappas of 0.9881, 0.9557, and 0.9175. The same study reports that all 11 patient-level binary errors were false positives, mostly due to overgrading of maculopathy (Vaghefi et al., 2021).

In software engineering, THEIA is a dataset-aware debugger for structural bug localization in deep learning programs written in Keras and PyTorch. It analyzes executable models together with training-data characteristics such as problem type, number of classes, image channels, and input range, and intervenes at the beginning of training rather than after prolonged optimization (Manke et al., 2024). On a benchmark of 40 buggy programs containing 75 structural bugs, it localized 57/75 bugs, compared with 17/75 for NeuraLint. After applying actionable fixes, the paper reports average performance improvement of 41% in 34/40 buggy programs for THEIA, versus 30% in 19/40 for NeuraLint (Manke et al., 2024).

In robot learning, Theia is a vision foundation model obtained by distilling multiple off-the-shelf vision foundation models into a compact student encoder. The student is trained on 1.2M ImageNet-1k images for 50 epochs using a multi-teacher loss

$245 \pm 3\ {\rm Myr}$9

with default $R = 6 \pm 1\ {\rm pc}$0 and $R = 6 \pm 1\ {\rm pc}$1 (Shang et al., 2024). The best configuration, Theia-B, has 86M parameters, was pretrained in 152 H100 GPU hours, and achieved a $R = 6 \pm 1\ {\rm pc}$2 average on the 14-task CortexBench evaluation. On real robots it reached, for example, 92% / 66% on Door Opening open/fully-open and 85% / 100% on Drawer Opening frozen/fine-tuned (Shang et al., 2024).

In neural reasoning, THEIA (“Three-valued Hybrid Engine for Inference Architecture”) is a pure-neural modular architecture for learning complete Kleene three-valued logic (K3) without an external symbolic solver. It processes arithmetic, order, set membership, and propositional logic through separate engines in a 128-dimensional latent space, with a total of 2.75M parameters (Li, 13 Apr 2026). On a 2M-sample dataset with input space of approximately $R = 6 \pm 1\ {\rm pc}$3, it achieved 12/12 targeted Kleene K3 rule coverage across 5 seeds in $R = 6 \pm 1\ {\rm pc}$4 minutes. In a mod-3 sequential composition experiment, it generalized from 5-step training to 500-step evaluation at $R = 6 \pm 1\ {\rm pc}$5, while matched flat MLP baselines collapsed to chance by 50 steps (Li, 13 Apr 2026).

In mobile distributed systems, Theia is a platform for crowd-sourced real-time content search over smartphone photos. Its two central mechanisms are Incremental Search, which expands search scope incrementally and exploits user feedback, and Partitioned Search, which splits execution between phone and cloud to reduce mobile energy consumption (Sani et al., 2011). In evaluation, the system reduced the cost per relevant photo by an average of 59%, reduced smartphone search energy consumption by up to 55% relative to full offloading and 81% relative to full local execution, and returned search results from smartphones in seconds (Sani et al., 2011).

Taken together, these usages show that “THEIA” functions less as a stable scientific term than as a reused project name across disparate domains. Its referent may be a disrupted open-cluster remnant, a Moon-forming impactor, a detector or mission concept, or a domain-specific AI or software system. The primary encyclopedia task is therefore disambiguation by field, with each usage carrying its own technical vocabulary, performance criteria, and scientific context.