Nautilus: Multifaceted Science Platform

• Nautilus is a cross-domain label representing diverse systems including modular space observatories, cryogenic resonant detectors, combined SIMS–SSAMS instruments, and advanced computational frameworks.
• It emphasizes modularity and scalability, as seen in designs like replicated telescopes for exoplanet spectroscopy yielding effective apertures around 50 m, and hyperclusters with thousands of GPUs.
• Its varied implementations drive innovations in exoplanet studies, gravitational-wave detection, trace isotope analysis, federated learning, and algorithmic methods across astrophysics and machine learning.

NAUTILUS is a recurrent name in the research literature for several unrelated observatory concepts, laboratory instruments, detectors, computational infrastructures, and algorithms. In current arXiv usage, it denotes a modular space observatory concept for exoplanet science, a cryogenic resonant-bar detector, a hybrid SIMS–AMS spectrometer, the National Research Platform Nautilus HyperCluster, and multiple software systems spanning Bayesian inference, robotics, tensor compilation, mesh generation, federated learning, and astrochemistry (Apai et al., 2019, Collaboration et al., 2012, Groopman et al., 2019, Hurt et al., 2024, Lange, 2023, Jin et al., 12 May 2026).

1. Scope of the name

The name is best understood as a cross-domain label rather than a single scientific entity. In some cases it denotes a physical instrument or observatory; in others it denotes a software package, compiler, or machine-learning framework.

Domain	Referent	Core function
Space astronomy	Nautilus Space Observatory	Modular telescope constellation for exoplanet spectroscopy and imaging
Detector physics	NAUTILUS resonant bar	Cryogenic resonant-mass detector for gravitational-wave and exotic-particle searches
Mass spectrometry	NRL NAUTILUS	Combined SIMS–SSAMS instrument for trace isotope imaging
Scientific computing	NRP Nautilus HyperCluster	Kubernetes-native multi-institutional compute resource
Computational methods	Multiple NAUTILUS frameworks	Bayesian sampling, robot-learning harnesses, tensor compilation, mesh generation, FL, astrochemical emulation

This plurality is not accidental. Across fields, the name is consistently attached to systems built around modularity, scalability, or hybridization. In the observatory literature, it refers to a scalable constellation of replicated telescopes; in mass spectrometry, to a combined SIMS–AMS platform; in computing, to infrastructures or algorithms that reduce the cost of scaling experiments or inference (Apai et al., 2019, Groopman et al., 2019, Hurt et al., 2024).

2. Nautilus as a space observatory concept

In exoplanet astronomy, Nautilus originated as a mission concept for a very large, modular space observatory composed of identical ultralight unit telescopes using multi-order diffraction engineered (MODE) lenses rather than conventional mirrors. The baseline concept uses about 35 unit telescopes, each with an 8.5 m diameter MODE lens, whose data are combined incoherently, yielding the photon-collecting power of a roughly 50 m class collector. The baseline science case is transmission spectroscopy of about 1,000 transiting, potentially Earth-like exoplanets, with low- to moderate-resolution spectra over approximately $0.45\text{–}1.0\,\mu\mathrm{m}$ and resolving power $R \approx 50\text{–}170$ (Apai et al., 2019).

The central architectural distinction is that Nautilus is an array, not an interferometer. Individual telescopes act as independent light buckets; intensities are co-added rather than beams being combined coherently. This avoids the metrology and phase-control requirements of interferometric formation flying while preserving scalability. The effective collecting diameter is expressed as

$$D_{\text{eff}} = \sqrt{35}\times 8.5\ \mathrm{m} \approx 50\ \mathrm{m}.$$

The design is coupled to a replication strategy: identical telescopes, replicated optics, and a build-out path from a small initial configuration toward a much larger survey facility (Apai et al., 2019).

The original science rationale was explicitly population-level. Nautilus was designed to survey transiting Earth-sized habitable-zone planets out to roughly 300 pc, rather than concentrating on a few tens of directly imaged systems. That emphasis on scalable sample size remains the common thread in later Nautilus white papers, which reuse the same modular observatory architecture for additional time-domain and stellar-astrophysics programs (Apai et al., 2019).

3. Derived space-science programs: stellar heterogeneities, flares, and exomoons

One proposed Nautilus program targets stellar photospheric heterogeneities as a limiting systematic in exoplanet transmission spectroscopy. The central problem is the Transit Light Source Effect, in which unocculted starspots, faculae, plages, and related structures imprint wavelength-dependent biases on transit depths. The white paper formulates the contaminated observation as

$$\delta_{\mathrm{obs}}(\lambda) = \delta_{\mathrm{planet}}(\lambda;\theta_p)\times S(\lambda;\theta_*),$$

and argues that a two-generation Nautilus program could empirically calibrate the stellar contamination factor $S(\lambda;\theta_*)$. Generation 1 uses a single-slit, medium-resolution spectrograph with $0.3\text{–}1.2\,\mu\mathrm{m}$ coverage, $R\sim 1000$ at $0.4\,\mu\mathrm{m}$, cadence $120\pm 60$ s, and white-light precision of $10\ \mathrm{ppm\ min^{-1}}$ to connect starspot-crossing events to disk-integrated diagnostics. Generation 2 then uses a slitless spectrograph with $R \approx 50\text{–}170$0, cadence $R \approx 50\text{–}170$1 s, and wavelength coverage $R \approx 50\text{–}170$2 centered on optimized diagnostics, monitoring hundreds of GKM stars to build population-level priors for exoplanet retrievals (Feinstein et al., 29 Jun 2026).

A second white paper uses Nautilus for fast, time-resolved spectroscopy of GKM stellar flares and their effects on planetary atmospheres. In that formulation, the observatory provides continuous, flux-calibrated spectroscopy from $R \approx 50\text{–}170$3 to $R \approx 50\text{–}170$4 at $R \approx 50\text{–}170$5, with $R \approx 50\text{–}170$6 s cadence and $R \approx 50\text{–}170$7 per spectral pixel at $R \approx 50\text{–}170$8. The stated aim is a statistical library of flare spectral templates indexed by flare energy, flare phase, and host-star spectral type, replacing fixed-temperature blackbody approximations commonly used in planetary photochemistry and escape calculations. The program is explicitly designed to recover wavelength- and time-resolved flare luminosities and to partition flare energy across the near-UV, optical, and near-infrared (Lin et al., 24 Jun 2026).

A third white paper applies the same architecture to exomoon astrometry. There the moon signal is the reflex motion of the directly imaged planet, with angular semi-amplitude

$R \approx 50\text{–}170$9

For this program, Nautilus again exploits scalability rather than higher angular resolution: individual units remain incoherent, so the inner working angle is set by a single 8.5 m telescope, while astrometric precision improves as $$D_{\text{eff}} = \sqrt{35}\times 8.5\ \mathrm{m} \approx 50\ \mathrm{m}.$$0. The white paper anchors the single-aperture precision at $$D_{\text{eff}} = \sqrt{35}\times 8.5\ \mathrm{m} \approx 50\ \mathrm{m}.$$1 for a 3 m telescope near 500 nm and adopts a five-year campaign detection floor of about $$D_{\text{eff}} = \sqrt{35}\times 8.5\ \mathrm{m} \approx 50\ \mathrm{m}.$$2. A staged program is proposed, beginning with the nearest imaged giant planets and extending, as the array grows, to a broader nearby-star sample of spectral type K and earlier (Wagner et al., 26 Jun 2026).

Taken together, these programs define Nautilus less as a single instrument than as a scalable observatory platform. The common logic is stable, long-baseline, space-based monitoring with modular growth: one program turns stellar contamination into a calibrated retrieval input, another builds empirical flare forcing libraries for planetary atmospheres, and another treats long-baseline relative astrometry as a route to exomoon discovery (Feinstein et al., 29 Jun 2026, Lin et al., 24 Jun 2026, Wagner et al., 26 Jun 2026).

4. Laboratory instruments and resonant detectors named NAUTILUS

In gravitational-wave physics, NAUTILUS is a cryogenic resonant-bar detector located in Frascati, Italy. It is an aluminum-alloy cylinder of length $$D_{\text{eff}} = \sqrt{35}\times 8.5\ \mathrm{m} \approx 50\ \mathrm{m}.$$3, diameter $$D_{\text{eff}} = \sqrt{35}\times 8.5\ \mathrm{m} \approx 50\ \mathrm{m}.$$4, and mass $$D_{\text{eff}} = \sqrt{35}\times 8.5\ \mathrm{m} \approx 50\ \mathrm{m}.$$5, cooled to about $$D_{\text{eff}} = \sqrt{35}\times 8.5\ \mathrm{m} \approx 50\ \mathrm{m}.$$6, with an auxiliary mechanical resonator and a dc SQUID readout. The detector responds to the odd longitudinal vibrational modes of the bar. In burst analyses with EXPLORER, about three years of coincident operation yielded 761 days of good live time and a null search for coincident short-duration gravitational-wave events; the resulting upper limit was reported as competitive, around $$D_{\text{eff}} = \sqrt{35}\times 8.5\ \mathrm{m} \approx 50\ \mathrm{m}.$$7, with interferometric-detector limits in that amplitude range (Collaboration et al., 2012). In a separate all-sky search for periodic signals from isolated neutron stars, half a year of 2001 data in the $$D_{\text{eff}} = \sqrt{35}\times 8.5\ \mathrm{m} \approx 50\ \mathrm{m}.$$8 band produced no detections and yielded upper limits on the dimensionless strain amplitude from $$D_{\text{eff}} = \sqrt{35}\times 8.5\ \mathrm{m} \approx 50\ \mathrm{m}.$$9 to $$\delta_{\mathrm{obs}}(\lambda) = \delta_{\mathrm{planet}}(\lambda;\theta_p)\times S(\lambda;\theta_*),$$0 (0809.0273).

The same detector was also repurposed for exotic-particle searches. Using the thermo-acoustic effect in the aluminum bar, NAUTILUS and EXPLORER were used to search for quark nuggets or nuclearites. For a vertical track through the middle of the bar, the quoted response is

$$\delta_{\mathrm{obs}}(\lambda) = \delta_{\mathrm{planet}}(\lambda;\theta_p)\times S(\lambda;\theta_*),$$1

with $$\delta_{\mathrm{obs}}(\lambda) = \delta_{\mathrm{planet}}(\lambda;\theta_p)\times S(\lambda;\theta_*),$$2 encoding the mass dependence of the effective cross section. Ten years of NAUTILUS data and seven years of EXPLORER data yielded limits of interest for nuclearites of mass less than $$\delta_{\mathrm{obs}}(\lambda) = \delta_{\mathrm{planet}}(\lambda;\theta_p)\times S(\lambda;\theta_*),$$3 grams, where the measured flux was reported as smaller than the flux predicted if nuclearites constituted the dark matter (Astone et al., 2013).

A wholly different physical instrument is the U.S. Naval Research Laboratory’s NAUTILUS, the NAval Ultra-Trace Isotope Laboratory’s Universal Spectrometer. This system combines a modified Cameca ims 4f secondary ion mass spectrometer with a $$\delta_{\mathrm{obs}}(\lambda) = \delta_{\mathrm{planet}}(\lambda;\theta_p)\times S(\lambda;\theta_*),$$4 NEC single-stage accelerator mass spectrometer into a unified SIMS–SSAMS beamline. The SSAMS serves as a molecule-filtering detector for the SIMS, enabling molecule-free raster ion imaging and spatially resolved trace isotope analysis. The instrument’s trace-element sensitivity is reported as at least $$\delta_{\mathrm{obs}}(\lambda) = \delta_{\mathrm{planet}}(\lambda;\theta_p)\times S(\lambda;\theta_*),$$5 better than commercial SIMS instruments because of near-zero background conditions, and the paper highlights applications in nuclear materials analysis, cosmochemistry, and geochemistry (Groopman et al., 2019).

The detector and spectrometer share little beyond the name. NAUTILUS in the resonant-bar literature denotes a cryogenic mechanical detector optimized around narrow-band strain sensitivity and thermo-acoustic event response; NAUTILUS at NRL denotes a hybrid microanalytical mass spectrometer built around molecular suppression and imaging of heterogeneous materials (Collaboration et al., 2012, Groopman et al., 2019).

5. Cyberinfrastructure and distributed research frameworks

In scientific computing, “Nautilus” can denote the National Research Platform Nautilus HyperCluster, a large Kubernetes-native, multi-institutional research compute resource. The cited case study describes it as an ever-growing Kubernetes cluster with more than 1,300 NVIDIA GPUs, 19,000 CPU cores, terabytes of available memory, and heterogeneous GPUs ranging from GTX 1080 (11 GB) to A100 (80 GB). In that work it served as the execution substrate for 234 deep-learning models across overhead object detection, burned-area segmentation, and deforestation detection, totaling 35,200 epochs, 8,084 million optimized parameters, 37.7 TB of imagery, and 4,040 hours of wall-clock compute (Hurt et al., 2024).

A separate usage appears in robot learning. The framework described as NAUTILUS is an open-source harness intended to turn a single natural-language prompt into ready-to-use reproduction, evaluation, fine-tuning, and deployment workflows across policy repositories, benchmarks, and real robots. It is organized around typed Policy, Benchmark, and Robot contracts, chambered execution environments, and a Tier-2 WebSocket transport layer. The paper frames its central scaling claim as a reduction of harness cost from $$\delta_{\mathrm{obs}}(\lambda) = \delta_{\mathrm{planet}}(\lambda;\theta_p)\times S(\lambda;\theta_*),$$6 to $$\delta_{\mathrm{obs}}(\lambda) = \delta_{\mathrm{planet}}(\lambda;\theta_p)\times S(\lambda;\theta_*),$$7, and in a leave-one-benchmark-out onboarding study reports aggregated task success of 98.7% / 98.0% for full NAUTILUS versus 95.4% / 93.2% for a vanilla agent under Opus/Sonnet settings (Jin et al., 12 May 2026).

In federated learning for vehicular-edge-cloud systems, Nautilus denotes a verifiable hierarchical FL framework for Internet of Vehicles settings. Its scheduler adapts local epochs, sparsification ratios, and quantization levels to node bandwidth, latency, and compute power, and a blockchain-backed verification layer commits scheduling decisions and spot-checks client compliance with Zero-Knowledge Proofs. In experiments at 50 nodes, the framework reduced total training time by 59.9% versus FedAvg, reduced communication to about 16.7% of FedAvg / zkFL, kept accuracy loss within about 1.5 percentage points, and held on-chain overhead below 2% of per-round time; the paper further states that 5% random spot checks achieve a malicious-behavior detection rate above 99% while incurring less than 5% of full-verification overhead (Wu et al., 22 Jun 2026).

Across these three cases, the shared motif is orchestration under heterogeneity. The HyperCluster exposes a heterogeneous GPU fleet through Kubernetes; the robotics harness stabilizes heterogeneous policies, simulators, and embodiments through typed contracts; the vehicular FL framework schedules heterogeneous clients while making the schedule itself auditable (Hurt et al., 2024, Jin et al., 12 May 2026, Wu et al., 22 Jun 2026).

6. Algorithms, compilers, and scientific software

Several NAUTILUS systems are purely algorithmic. In Bayesian computation, NAUTILUS is an importance nested sampling method and Python package that combines importance nested sampling with fully connected neural-network regression. It estimates the evidence

$$\delta_{\mathrm{obs}}(\lambda) = \delta_{\mathrm{planet}}(\lambda;\theta_p)\times S(\lambda;\theta_*),$$8

by treating all evaluated points as samples from a pseudo-proposal density $$\delta_{\mathrm{obs}}(\lambda) = \delta_{\mathrm{planet}}(\lambda;\theta_p)\times S(\lambda;\theta_*),$$9, then weighting them as $S(\lambda;\theta_*)$0. The paper reports substantially higher sampling efficiency than EMCEE, DYNESTY, ULTRANEST, and POCOMC, often by more than an order of magnitude, while retaining accurate posterior and evidence estimates and scaling well up to about 50 dimensions (Lange, 2023).

In compiler research, Nautilus is an end-to-end tensor compiler for tiled GPU kernels. Starting from a math-like algebraic operator description, it performs successive lowering together with a high-level auto-scheduler that searches sequences of algebraic rewrites and reduction fusions while preserving tile-friendly program structure. The headline claim is that it can automatically discover FlashAttention-3-like kernels from a high-level attention specification. Across five transformer-based models and 150 evaluation configurations, the reported throughput gains reach up to 23% over state-of-the-art compilers on NVIDIA GH200 and up to 42% on RTX 5090, while matching or exceeding manually written cuDNN kernels on many long-sequence configurations (Zhao et al., 16 Apr 2026).

In 3D generation, Nautilus is a locality-aware autoencoder for direct triangle-mesh generation. It introduces a shell-based tokenization that preserves face proximity and compresses sequence length through locally shared vertices and edges, together with a Dual-stream Point Conditioner for multi-scale geometric guidance. The model reports generation of meshes with up to 5,000 faces, a compression ratio of 0.275 versus 0.462 for AMT and 0.474 for EdgeRunner, and point-cloud-conditioned results with Chamfer distance 0.087 and Hausdorff distance 0.176, compared with 0.106 / 0.248 for MeshAnything V2; in the cited user study it was preferred 88.68% of the time (Wang et al., 24 Jan 2025).

In astrochemistry, Nautilus refers to the underlying three-phase chemical code emulated by a later neural model. Nautilus 1.1 is described there as a gas–grain code tracking gas-phase, grain-surface, and grain-mantle chemistry. The emulator paper uses conditional neural fields to reproduce the abundances of 192 species for arbitrary times between 1 and $S(\lambda;\theta_*)$1 years, with uncertainties well below 0.2 dex for all species and a computing time of order $S(\lambda;\theta_*)$2 smaller than the original code. As a scientific example, the emulator’s feature-importance analysis found that increasing the initial sulphur abundance from a depleted scenario to the cosmic abundance enhances the electron density by about an order of magnitude in low-density gas (Ramos et al., 2024).

These algorithmic usages show the broadest semantic spread of the name. Here NAUTILUS does not denote an instrument or observatory at all, but rather a family of methods for importance sampling, tensor-kernel synthesis, artist-like mesh generation, and accelerated surrogate modeling of stiff astrochemical kinetics (Lange, 2023, Zhao et al., 16 Apr 2026, Wang et al., 24 Jan 2025, Ramos et al., 2024).