COSMOS: Mapping Complex Systems
- COSMOS is a polysemous research designation representing both the mapped universe in astronomy and unified infrastructures in blockchain, AI, and numerical relativity.
- In astronomy, COSMOS encompasses a 2 deg² survey field and employs Bayesian reconstruction to visualize galaxy dynamics and dark matter distribution.
- In technical domains, COSMOS names robust software systems and world models that simulate complex physical processes and ensure deterministic consensus.
COSMOS is a polysemous research designation whose meaning depends on disciplinary context. In astronomy it denotes both the cosmos as the mapped universe and the COSMOS legacy survey field; in other technical domains it names software systems, model architectures, and datasets in blockchain engineering, numerical relativity, machine learning, and technology studies (Pomarede et al., 2017, Weaver et al., 2021, Surmont et al., 2023, Yoo et al., 1 Jun 2026, Aditi et al., 1 Jun 2026, Gong et al., 15 May 2025). A plausible common theme is systematic representation of complex structure: matter flows in the nearby universe, galaxy populations across cosmic time, consensus-critical execution paths, multimodal embodied environments, or the frontier of emerging technologies.
1. Conceptual scope and principal usages
In philosophical and scientific usage, cosmos can denote the totality of physical reality. Nicholas Maxwell uses “cosmos” or “universe” broadly for total physical reality and argues, within aim-oriented empiricism, that science operates with a hierarchy of assumptions culminating in physical comprehensibility, formalized as Physicalism(8,1), the thesis that a unified true theory of everything exists (Maxwell, 2012). In astronomy, by contrast, cosmography is defined as “the study and making of maps of the universe or cosmos,” a descriptive and cartographic enterprise distinct from theoretical cosmology (Pomarede et al., 2017).
| Domain | COSMOS referent | Source |
|---|---|---|
| Philosophy of science | Cosmos as total physical reality | (Maxwell, 2012) |
| Extragalactic mapping | Cosmography of the local universe | (Pomarede et al., 2017) |
| Survey astronomy | The COSMOS field and its catalogs | (Weaver et al., 2021) |
| Blockchain systems | Cosmos appchain framework and SDK | (Surmont et al., 2023) |
| Numerical relativity | COSMOS code for PBH formation | (Yoo et al., 1 Jun 2026) |
| AI and world models | Cosmos 3 omnimodal world models | (Aditi et al., 1 Jun 2026) |
| Technology intelligence | Cosmos 1.0 emerging-technology map | (Gong et al., 15 May 2025) |
This multiplicity is not accidental. Several of these usages are explicitly cartographic or infrastructural, while others are acronymic project names. The astronomical usages are historically primary in the supplied literature, but the computational usages are now comparably prominent.
2. Cosmography and mapping of the nearby universe
In extragalactic astronomy, cosmography is a data-driven program for recovering the three-dimensional structure and dynamics of the local universe. “Cosmography and Data Visualization” describes a workflow centered on the Cosmicflows Project, which assembles catalogs of galaxy distances and peculiar velocities and reconstructs continuous velocity, density, gravitational potential, and velocity-shear fields with a Bayesian Wiener Filter (Pomarede et al., 2017). The kinematic core is the decomposition
so that peculiar velocities trace departures from pure Hubble expansion and thereby encode the underlying gravitational field, including the influence of dark matter (Pomarede et al., 2017).
The paper emphasizes that the raw observables are sparse, noisy, and largely line-of-sight, whereas the scientific targets are volumetric. The Wiener Filter reconstruction therefore regularizes the inversion and yields the most probable three-dimensional velocity and density fields, together with derived quantities such as the gravitational potential and the velocity shear tensor. In the same framework, the velocity field can be separated into local divergent and tidal components, and the cosmic web can be classified dynamically via the eigenstructure of the shear tensor, producing the Cosmic V-Web (Pomarede et al., 2017).
Visualization is not treated as a presentational afterthought. The SDvision software, written in IDL with OpenGL and GLSL interfaces, provides ray-casting volume rendering, isosurfaces, slicing, streamlines, hedgehog arrows, and point-cloud rendering in supergalactic coordinates. These techniques were applied to galaxy catalogs, peculiar velocities, reconstructed scalar and vector fields, and envelope surfaces of basins of attraction (Pomarede et al., 2017). The resulting images, videos, and interactive viewers supported studies of the Great Attractor, the mapping of the cosmic web, the identification of attractors and repellers, and the dynamical definition of the Laniakea supercluster as a basin of attraction rather than a simple overdensity (Pomarede et al., 2017).
A recurring implication is methodological rather than merely visual. Because filaments, voids, and convergent or divergent flows are intrinsically three-dimensional, the paper argues that projections onto two-dimensional sky maps obscure continuity and connectivity. This suggests that modern cosmography is best understood as an overview of peculiar-velocity inference, Bayesian field reconstruction, and interactive volumetric analysis rather than as an updated star chart (Pomarede et al., 2017).
3. The COSMOS field as an observational infrastructure
As a proper name in observational astronomy, COSMOS refers to the Cosmic Evolution Survey field: a contiguous, nearly square extragalactic field centered at
$\mathrm{RA}=10^\mathrm{h}00^\mathrm{m}27.92^\mathrm{s},\qquad \mathrm{Dec}=+02^\circ12'03\farcs5,$
covering about on the sky (Weaver et al., 2021). Its combination of large area and deep multiwavelength coverage has made it a reference field for large-scale structure, weak lensing, rare galaxy populations, and photometric-redshift calibration (Weaver et al., 2021).
The current photometric reference release is COSMOS2020, which provides source detection and multi-wavelength photometry for 1.7 million sources across the field using two complementary catalogs: the aperture-based Classic catalog and the profile-fitting Farmer catalog (Weaver et al., 2021). Photometric redshifts are computed with two independent codes, and the reported performance reaches sub-percent photometric-redshift accuracy for sources, while even the faintest sources at $25
The spectroscopic counterpart is the COSMOS Spectroscopic Redshift Compilation DR1, which assembles 165,312 redshifts of 97,929 unique objects from 108 observing programs up to (Khostovan et al., 28 Feb 2025). The compilation is 50% spectroscopically complete by and 0 mag, with 1 completeness within the CANDELS area and 2 completeness in the outer regions of COSMOS for 3 mag and 4 mag, separately (Khostovan et al., 28 Feb 2025). It also identifies spectroscopically under-sampled galaxy subpopulations by projecting the compilation onto self-organizing maps trained on COSMOS2020 (Khostovan et al., 28 Feb 2025).
Recent JWST spectroscopy has extended the field’s role. The NIRISS PASSAGE pure-parallel survey contributes 2,183 emission-line sources at 5 within COSMOS, of which 1,955 are new spectroscopic redshifts relative to earlier compilations (Huberty et al., 5 Jan 2026). The same study finds excellent agreement between COSMOS photometric redshifts and PASSAGE spectroscopic redshifts for strong multi-line emitters, while single-line emitters are likely mis-identified around 18% of the time, or about 19% for originally assumed H6 single-line sources after the paper’s correction exercise (Huberty et al., 5 Jan 2026). At still higher redshift, JWST COSMOS-3D reports a spectroscopically selected catalog of 237 [O III] emitters at 7 over 8, with the first constraints on the cosmic variance of [O III] emitters, estimating a 15% relative uncertainty for the 9 [O III] luminosity function in that field (Meyer et al., 13 Oct 2025).
The field is therefore not merely a sky area. It is an observational infrastructure: a deeply cross-calibrated environment in which photometry, spectroscopy, morphology, lensing, and environmental reconstruction can be jointly developed and validated.
4. Reconstruction, simulation, and population inference in the COSMOS field
Several recent projects use COSMOS not only as an observing field but as a target for forward models and constrained simulations. “BIRTH of the COSMOS Field” reconstructs primordial and evolved dark-matter density fields over $\mathrm{RA}=10^\mathrm{h}00^\mathrm{m}27.92^\mathrm{s},\qquad \mathrm{Dec}=+02^\circ12'03\farcs5,$0 by applying the extended COSMIC BIRTH algorithm to five spectroscopic surveys in COSMOS: zCOSMOS-deep, VUDS, MOSDEF, ZFIRE, and FMOS-COSMOS (Ata et al., 2020). The reconstruction explicitly accounts for gravitational matter displacements, peculiar velocities, galaxy bias, and survey completeness, and reveals a holistic view of proto-clusters and the growth of the cosmic web during cosmic high noon (Ata et al., 2020).
The simulation project cosmosTNG takes a related but complementary approach. It simulates a $\mathrm{RA}=10^\mathrm{h}00^\mathrm{m}27.92^\mathrm{s},\qquad \mathrm{Dec}=+02^\circ12'03\farcs5,$1 patch of the COSMOS field at $\mathrm{RA}=10^\mathrm{h}00^\mathrm{m}27.92^\mathrm{s},\qquad \mathrm{Dec}=+02^\circ12'03\farcs5,$2 using initial conditions inferred from galaxy redshift surveys and the CLAMATO Lyman-$\mathrm{RA}=10^\mathrm{h}00^\mathrm{m}27.92^\mathrm{s},\qquad \mathrm{Dec}=+02^\circ12'03\farcs5,$3 forest tomography survey, reconstructed by the TARDIS algorithm (Byrohl et al., 2024). The suite evolves eight realizations of this constrained volume with the IllustrisTNG galaxy-formation model at a baryonic mass resolution of $\mathrm{RA}=10^\mathrm{h}00^\mathrm{m}27.92^\mathrm{s},\qquad \mathrm{Dec}=+02^\circ12'03\farcs5,$4, equal to TNG100-1 (Byrohl et al., 2024). In this setup, the COSMOS subvolume is found to be overdense relative to random regions, with a larger abundance of high-mass galaxies and a significantly lower quenched fraction at fixed mass, suggesting accelerated stellar-mass growth and a higher cosmic star-formation-rate density in that specific region (Byrohl et al., 2024).
A third layer is statistical rather than dynamical. The pop-cosmos model is calibrated to 140,938 $\mathrm{RA}=10^\mathrm{h}00^\mathrm{m}27.92^\mathrm{s},\qquad \mathrm{Dec}=+02^\circ12'03\farcs5,$5 galaxies from COSMOS with photometry in 26 bands from the ultraviolet to the infrared (Alsing et al., 2024). It combines a score-based diffusion population model, a Prospector/FSPS-based stellar population synthesis model, and an explicit data model for observation, calibration, noise, and selection, fitting the model by minimizing an optimal-transport distance between synthetic and real data (Alsing et al., 2024). The resulting model predicts the mass function and redshift distribution, the mass-metallicity-redshift and fundamental metallicity relations, the star-forming sequence, and dust and gas-ionization relations, while faithfully reproducing COSMOS color distributions across a broad redshift and wavelength range (Alsing et al., 2024).
Taken together, these projects shift COSMOS from a passive legacy field to an actively reconstructed one. The field becomes an object of inference: a specific realization of the cosmic web whose initial conditions, galaxy populations, and environmental histories can be modeled jointly.
5. COSMOS as software infrastructure in blockchain and numerical relativity
Outside observational astronomy, Cosmos is the name of a prominent blockchain framework for application-specific blockchains, or appchains (Surmont et al., 2023). The Cosmos SDK is a modular framework in Go, while Tendermint for SDK versions below v0.47.0 and CometBFT for v0.47.0 and above provide BFT consensus through ABCI (Surmont et al., 2023). In this setting, deterministic execution is a consensus requirement: store-affecting code must produce identical key-value store updates on all honest nodes for the same transaction and block sequence (Surmont et al., 2023).
The security study on static application security testing focuses on sources of non-determinism in consensus-critical Go code, including map iteration, goroutines, floating-point arithmetic, system time, unsafe packages, hardcoded Bech32 prefixes, and platform-dependent types (Surmont et al., 2023). It evaluates an 11-chain corpus and introduces a refactored CodeQL strategy that identifies consensus-critical functions by call-graph reachability from BeginBlock, EndBlock, and DeliverTx rather than by fragile package blacklists (Surmont et al., 2023). Across all analyzed chains, the paper reports a precision increase per chain between 42.5% and 100%, with an average increase of 80.64%, thereby making Cosmos-specific SAST tooling substantially more practical for CI integration (Surmont et al., 2023). A notable controversy remains the treatment of panics in BeginBlock and EndBlock: the paper records community disagreement about “expected” panics, while maintaining that unexpected panics are always problematic (Surmont et al., 2023).
In numerical relativity, COSMOS is a compact, dependency-free C++ package for solving the Einstein equations in $\mathrm{RA}=10^\mathrm{h}00^\mathrm{m}27.92^\mathrm{s},\qquad \mathrm{Dec}=+02^\circ12'03\farcs5,$6 dimensions, specialized for primordial black hole formation (Yoo et al., 1 Jun 2026). It implements a massless scalar field and a perfect fluid with a linear equation of state $\mathrm{RA}=10^\mathrm{h}00^\mathrm{m}27.92^\mathrm{s},\qquad \mathrm{Dec}=+02^\circ12'03\farcs5,$7, uses non-Cartesian scale-up coordinates and fixed mesh refinement to resolve collapse inside an expanding cosmological background, and employs OpenMP for parallelization (Yoo et al., 1 Jun 2026). The public code includes examples for a small sinusoidal perturbation, adiabatic spherical collapse, and spherical isocurvature collapse, and identifies PBH formation by apparent-horizon finding (Yoo et al., 1 Jun 2026). Its design deliberately omits external dependencies and elliptic initial-data solvers, reflecting its specialization for long-wavelength cosmological initial data rather than generic asymptotically flat spacetimes (Yoo et al., 1 Jun 2026).
These two software usages share a structural concern with the astronomical one: reliable evolution of complex systems under strict dynamical constraints. In Cosmos appchains the constraint is deterministic consensus; in COSMOS numerical relativity it is stable evolution of the Einstein–matter system in a cosmological setting.
6. COSMOS in machine learning and technology intelligence
In contemporary AI research, Cosmos 3 denotes a family of omnimodal world models for Physical AI that jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture (Aditi et al., 1 Jun 2026). The architecture separates an autoregressive subsequence handled by a “Reasoner” tower from a diffusion subsequence handled by a “Generator” tower, while allowing the diffusion stream to attend to the autoregressive context (Aditi et al., 1 Jun 2026). The system is positioned as a single backbone that subsumes vision-LLMs, video generators, world simulators, and world-action models, and the report states that its post-trained models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written (Aditi et al., 1 Jun 2026). The paper also notes limitations, including sim-to-real gaps, hallucination and physical errors in generated videos, and substantial compute requirements (Aditi et al., 1 Jun 2026).
A different machine-learning usage appears in “Continuous Simplicial Neural Networks,” where the method itself is named COSMOS (Einizade et al., 17 Mar 2025). This architecture replaces discrete polynomial simplicial filters with PDE-derived continuous diffusion on simplicial complexes, using matrix exponentials of lower and upper Hodge Laplacians to define propagation (Einizade et al., 17 Mar 2025). The paper provides theoretical stability bounds under simplicial perturbations and analyzes over-smoothing through the Dirichlet energy, arguing that the continuous diffusion times $\mathrm{RA}=10^\mathrm{h}00^\mathrm{m}27.92^\mathrm{s},\qquad \mathrm{Dec}=+02^\circ12'03\farcs5,$8 and $\mathrm{RA}=10^\mathrm{h}00^\mathrm{m}27.92^\mathrm{s},\qquad \mathrm{Dec}=+02^\circ12'03\farcs5,$9 provide finer control of smoothing than discrete simplicial neural networks (Einizade et al., 17 Mar 2025).
In technology studies, Cosmos 1.0 is a multidimensional map of the emerging technology frontier (Gong et al., 15 May 2025). The dataset contains 23,544 technology-related entities, or ET23k, organized into three meta clusters and seven theme clusters and represented by 100-dimensional Wikipedia2vec embeddings (Gong et al., 15 May 2025). A curated ET100 subset of 100 emerging technologies is manually verified, and the resource adds indices such as the Technology Awareness Index, Generality Index, Deeptech Index, Age of Tech Index, and a classifier-derived Technology Proximity Index (Gong et al., 15 May 2025). The paper explicitly notes limitations arising from English Wikipedia bias, 2018 static embeddings, and uneven third-party data coverage, so the map is comprehensive within its construction pipeline rather than exhaustive in a global sense (Gong et al., 15 May 2025).
Across these AI and technology usages, COSMOS designates systems that attempt to unify heterogeneous modalities or domains inside a common representational frame. This suggests a modern, acronymic extension of the older cartographic idea: not only to map a world, but to learn a space in which disparate observations become jointly navigable.