EarthMM: Multiple Research Frontiers

• EarthMM is a polysemous term defining a curated, large-scale multimodal remote sensing dataset that underpins generative Earth observation models like MetaEarth-MM.
• In dark matter research, EarthMM refers to an Earth-matter-enhanced axion search framework that leverages terrestrial effects to boost sensitivity to axion gradients.
• In exoplanet and infrastructure studies, EarthMM groups methods for detecting Earth-mass planets and designs systems for Earth–Mars transport by minimizing resource assumptions.

EarthMM is a field-dependent research label rather than a single standardized concept. In the supplied arXiv literature, it denotes several distinct constructs: a large-scale multimodal remote sensing dataset introduced for the generative model MetaEarth-MM; an Earth-matter-enhanced axion dark matter search framework; and, in broader interpretive use, a family of Earth-mass or Earth-like planetary detection, formation, and infrastructure problems spanning astrometry, microlensing, direct thermal detection, minimum-mass disk reconstruction, and reusable Earth–Mars logistics [(Yu et al., 19 May 2026); (Huang et al., 23 Feb 2026); (Malbet et al., 2011); (Chiang et al., 2012); (Janhunen et al., 2014)]. The term therefore has to be read contextually, with its meaning fixed by the surrounding research program rather than by a universal acronym expansion.

1. Terminological scope across research areas

In the supplied corpus, EarthMM is used in at least four technically separate ways. The most explicit data-centric use is in Earth observation, where EarthMM is the dataset underpinning MetaEarth-MM. A second explicit use is in dark-matter detection, where EarthMM denotes an Earth-matter-enhanced axion search. Additional summaries use EarthMM more loosely for Earth-mass exoplanet science and for architectures coupling Earth and Mars through in-space resource logistics.

Context	Meaning of EarthMM	Reference
Remote sensing	Large-scale multimodal Earth observation dataset	(Yu et al., 19 May 2026)
Dark matter	Earth-matter-enhanced axion wind search	(Huang et al., 23 Feb 2026)
Exoplanets	Earth-mass / Earth-like planet detection and characterization problem	[(Malbet et al., 2011); (Abe et al., 2013); (Saito et al., 2011)]
Space infrastructure	Earth–Mars transport architecture enabled by orbital depots and E-sails	(Janhunen et al., 2014)

This plurality matters methodologically. A remote-sensing paper uses EarthMM to denote a curated dataset and a training substrate for joint generative modeling, whereas the axion paper uses it to denote an environment-aware signal model in which terrestrial matter reshapes the local field profile. In exoplanet and mission-design contexts, the label is not a single mission acronym but a convenient way of grouping Earth-mass detection goals or Earth-centered logistical architectures. A plausible implication is that EarthMM should be treated as a polysemous term whose meaning must always be resolved by domain.

2. EarthMM as a multimodal remote sensing dataset

In the remote-sensing usage, EarthMM is the large-scale dataset introduced to support MetaEarth-MM, a unified generative foundation model for cross-modal Earth observation imagery. The dataset comprises 2.8 million images and 2.2 million aligned modality pairs across five modalities—RGB, SAR, NIR, PAN, and OSM—with a resolution range of 0.5 to 10 m/pixel and global coverage over many land areas and cities. All data are normalized by cropping or resampling to 256 × 256 patches. The dataset also spans multiple resolution bands, specifically ≤ 0.5 m/pix, 1–5 m/pix, and 8–10 m/pix (Yu et al., 19 May 2026).

EarthMM was built from 15 public high-quality pair-wise aligned datasets together with a newly collected RGB-OSM dataset covering 42 representative cities worldwide such as London, Beijing, New York, and Tokyo. The integrated public sources listed in the supplied summary include OpenEarthMap-SAR, DFC23, DDHRNet, fMoW, MultiResSAR, OSDataset, QXS-SAROPT, SEN12MS, SpaceNet-3, SpaceNet-5, SpaceNet-6, SSL4EO-S12, WHU-OPT-SAR, and 3MOS. Curation is described as strict and consistent across sources: it removes severe noise, corrupted observations, samples with large-area cloud occlusion, and highly repetitive regions such as homogeneous bare land, while ensuring that RGB-OSM pairs are geographically aligned by coordinates (Yu et al., 19 May 2026).

The dataset was created because existing remote-sensing generative datasets were described as too limited in one or more of the following respects: scale, number of modality combinations, geographic diversity, or spatial-resolution diversity. EarthMM is therefore intended to provide large-scale aligned multimodal supervision for joint generation, five-modality coverage for unified modeling, global scene diversity for generalization, and multi-resolution data for learning across spatial scales. In this sense, EarthMM is not merely a benchmark; it is the data backbone for a specific modeling hypothesis about multimodal Earth observation.

3. Scene-centered joint modeling and the MetaEarth-MM framework

MetaEarth-MM treats multimodal remote sensing images as different observations of the same underlying scene rather than as isolated visual domains. The core probabilistic statement in the supplied summary is

$$p(\mathbf{x}_i, \mathbf{x}_j \mid \mathbf{s}) = p(\mathbf{x}_i \mid \mathbf{s})\,p(\mathbf{x}_j \mid \mathbf{s}),$$

where $\mathbf{s}$ is a latent scene variable. This “scene-centered joint modeling” replaces direct appearance-level mappings with a decoupled architecture that first infers a latent scene representation from available observations and then generates target modalities conditioned on that intermediate state. The model therefore supports paired joint generation, conditional translation, and any-to-any translation within a unified system (Yu et al., 19 May 2026).

The training formulation is based on flow matching / rectified flow. For each modality latent $\mathbf{z}_i^0$, noise $\mathbf{z}_i^1 \sim \mathcal{N}(0, I)$, and time $t$, the interpolation is

$$\mathbf{z}_i^t = (1-t)\mathbf{z}_i^0 + t\mathbf{z}_i^1,$$

with target velocity

$$\mathbf{v}_i = \frac{d\mathbf{z}_i^t}{dt} = \mathbf{z}_i^1 - \mathbf{z}_i^0.$$

A scene inference module estimates latent scene tokens from paired noisy latents, and a modality-aware routed generator predicts modality-specific velocity fields while using modality-specific FFN branches and shared attention layers. To enforce consistency between inferred scene tokens, the model adds a symmetric InfoNCE scene-consistency term, with overall objective

$$\mathcal{L} = \mathcal{L}_{\text{joint}} + \lambda \mathcal{L}_{\mathrm{sc}},$$

and $\lambda = 1$ in training (Yu et al., 19 May 2026).

The reported findings are that MetaEarth-MM exhibits strong cross-modal translation on SAR $\leftrightarrow$ RGB, OSM $\mathbf{s}$0 RGB, NIR $\mathbf{s}$1 RGB, and PAN $\mathbf{s}$2 RGB; supports zero-shot generation on unseen modality pairs such as OSM–NIR and OSM–SAR; and performs better in joint generation than a cascade pipeline in which one modality is generated and then translated to a second. The study also reports downstream benefits in generative data augmentation, image-level domain adaptation, and zero-shot representation transfer, and describes the approach as computationally heavy, at around 600M parameters trained for 900K iterations. This suggests that EarthMM is intended not only as a dataset for image synthesis but also as infrastructure for multimodal Earth observation foundation modeling (Yu et al., 19 May 2026).

4. EarthMM as an Earth-matter-enhanced axion dark matter search

In the dark-matter usage, EarthMM refers to an axion “wind” search that explicitly incorporates the way Earth’s ordinary matter modifies the local axion field when the axion has quadratic couplings to matter. The central claim is that Earth is not merely a passive laboratory environment: for such couplings, terrestrial nucleon density shifts the axion’s effective mass in matter, creating an effective potential well that suppresses field amplitude near the surface while strongly enhancing the spatial gradient to which spin-based sensors are sensitive. The relevant matter-induced shift is summarized by

$\mathbf{s}$3

so that inside Earth

$\mathbf{s}$4

with a larger in-medium wavenumber $\mathbf{s}$5 (Huang et al., 23 Feb 2026).

The search targets the derivative axion-nucleon interaction, which in the nonrelativistic limit is written as

$\mathbf{s}$6

Experimentally, the platform is a self-compensation alkali–noble-gas comagnetometer comprising coupled K, Rb, and $\mathbf{s}$7Ne spin ensembles. The sensitive axis is aligned radially with respect to Earth, and operation near the compensation point suppresses ordinary magnetic perturbations while preserving sensitivity to anomalous pseudo-fields associated with the axion gradient. Calibration is performed with controlled rotations and applied magnetic fields, using the response to mechanical rotation as a proxy for the expected axion response. The experiment collected 132 hours of active data, with calibration every four hours (Huang et al., 23 Feb 2026).

The search covers axion masses

$\mathbf{s}$8

corresponding to a Compton-frequency window of roughly $\mathbf{s}$9–$\mathbf{z}_i^0$0 Hz. The analysis accounts for finite axion linewidth, finite frequency resolution, correlations among neighboring mass hypotheses, and the look-elsewhere effect over more than 8 million test points. No signal candidates survived the global-significance analysis, so the result is a null detection. The paper reports 95% C.L. limits on the axion-neutron derivative coupling reaching down to $\mathbf{z}_i^0$1, with improvement over earlier laboratory axion-wind searches by up to three orders of magnitude for some masses. A central methodological conclusion is that environmental effects cannot be treated as negligible corrections: under quadratic couplings, the Earth itself becomes part of the detector response (Huang et al., 23 Feb 2026).

5. EarthMM in Earth-mass exoplanet detection and characterization

Several supplied summaries connect EarthMM to the problem of detecting or characterizing planets at the Earth-mass or Earth-like scale. In this usage, the term denotes a scientific regime rather than a single instrument. One route is space astrometry: the proposed NEAT mission aims to detect and characterize planetary systems exhaustively down to 1 Earth mass in the habitable zone and beyond around nearby F, G, and K stars. Its stated performance goal is 0.05 $\mathbf{z}_i^0$2as astrometric precision, achieved with a 1 m class off-axis parabolic telescope, a focal plane 40 m away, formation flying of two spacecraft, and a metrology system projecting dynamic Young’s fringes onto the focal plane. The paper gives the benchmark that at 10 pc an Earth at 1 AU around a Sun-like star induces about 0.3 $\mathbf{z}_i^0$3as, whereas Jupiter induces about 500 $\mathbf{z}_i^0$4as; with the target precision, an Earth-mass planet at 1 AU around a Sun-like star at 10 pc would be detectable with SNR $\mathbf{z}_i^0$5 (Malbet et al., 2011).

A second route is microlensing at moderately high magnification. The supplied summary argues that the scientifically useful regime is not the most extreme peaks but events with

$\mathbf{z}_i^0$6

monitored intensively only around their peaks with 1–2 m class telescopes. In this regime, planets of a few Earth masses can produce deviations of about 5% lasting about 0.7–3 hr in events with magnification around 100, provided the projected separation lies in the annulus $\mathbf{z}_i^0$7; Earth-mass planets produce similar deviations if they lie in $\mathbf{z}_i^0$8. The perturbation duration is described by

$\mathbf{z}_i^0$9

and is noted to be independent of magnification because higher magnification both lengthens the Einstein arc and makes it move faster around the ring (Abe et al., 2013).

A third route is direct thermal detection in the millimeter/submillimeter regime. The proposed VLMSA concept is motivated by the requirement to directly detect thermal radiation from a 2nd Earth in a solar-system-like planetary system at 2 pc, in less than 24 hours of observation. The design assumptions are concrete: 64 antennas, 50 m diameter each, 25 $\mathbf{z}_i^1 \sim \mathcal{N}(0, I)$0m surface accuracy, placement within an area of 300 km, dual-polarization SSB receivers, and IF bandwidth of 128 or 256 GHz, with the possibility that a few antennas would be located at up to 3000 km to reach the 20 $\mathbf{z}_i^1 \sim \mathcal{N}(0, I)$1arcsec angular resolution stated as potentially necessary for black-hole imaging (Saito et al., 2011).

Ground-based transit work supplies a complementary, smaller-star pathway. The MEarth project operates from the Fred Lawrence Whipple Observatory in Arizona and Cerro Tololo Inter-American Observatory in Chile, with eight 0.4 m robotic telescopes at each site, and targets approximately 3000 nearby mid-to-late M-dwarfs with $\mathbf{z}_i^1 \sim \mathcal{N}(0, I)$2. Its science logic follows the standard depth relation

$\mathbf{z}_i^1 \sim \mathcal{N}(0, I)$3

under which small stars amplify transit signals, while their low luminosities move habitable zones inward and shorten periods. The survey revisits each star every 20–30 minutes and uses real-time triggering to capture ongoing transits lasting only 0.5–2 hours (Irwin et al., 2014).

Taken together, these programs show that the EarthMM problem in exoplanet science is methodologically plural: astrometry targets true masses and inclinations, microlensing probes cold planets near the Einstein ring, direct mm/submm concepts aim at thermal emission, and transit surveys exploit the favorable geometry of nearby low-mass stars.

6. Planet formation and Earth–Mars infrastructure interpretations

In another supplied interpretation, EarthMM refers to the Minimum-Mass Extrasolar Nebula (MMEN): the minimum disk of solids plus solar-composition gas that could have produced the observed population of close-in super-Earths in situ if formation were nearly 100% efficient and migration negligible. Using $\mathbf{z}_i^1 \sim \mathcal{N}(0, I)$4 Kepler planets with $\mathbf{z}_i^1 \sim \mathcal{N}(0, I)$5 and $\mathbf{z}_i^1 \sim \mathcal{N}(0, I)$6 d, and the mass–radius relation

$\mathbf{z}_i^1 \sim \mathcal{N}(0, I)$7

the inferred solid surface density is fit by

$\mathbf{z}_i^1 \sim \mathcal{N}(0, I)$8

with corresponding gas surface density

$\mathbf{z}_i^1 \sim \mathcal{N}(0, I)$9

The fiducial disk is described as gravitationally stable, with

$t$0

and able to form super-Earth cores rapidly enough to accrete modest H/He envelopes before gas dispersal (Chiang et al., 2012).

A separate, infrastructure-oriented usage appears in the E-sail facilitated Manned Mars Initiative (EMMI), where the supplied summary describes an EarthMM-style architecture linking Earth and Mars through reusable logistics. The concept uses propellantless electric solar wind sails to move asteroid-mined water into high orbits around Earth and Mars, where the water is converted into LOX/LH$t$1 cryogenic fuel in orbital depots. The E-sail thrust scales as

$t$2

and a large E-sail producing about 1 N at 1 au is stated to be able to travel from Earth to the asteroid belt in about one year and return three tonnes of water in three years. A representative system mass table lists 50,000 kg/year water production and transportation, an extractor vehicle of 2000 kg, a transporter E-sail of 500 kg, and a manned vehicle of 50,000 kg. The paper argues that once the infrastructure exists, the recurrent cost of continuous bidirectional Earth–Mars traffic could ultimately approach the recurrent cost of operating the ISS (Janhunen et al., 2014).

These two interpretations are conceptually different but structurally similar: both are minimum-resource reconstructions. The MMEN reconstructs the least massive disk compatible with close-in super-Earth formation, while EMMI reconstructs a minimum recurrent-logistics chain for sustained human transport between Earth and Mars.

7. Nomenclature, adjacent terms, and common confusions

Because EarthMM is not unique across fields, confusion with visually similar or conceptually adjacent terms is easy. MEarth is a dedicated transit survey of nearby M-dwarfs; despite the orthographic similarity, it is a specific exoplanet survey name rather than a generic EarthMM designation. MilkyWay@home uses an Earth-Mover Distance method to compare normalized histograms of stellar density along tidal streams, where the “Earth-Mover” label refers to a transport-based goodness-of-fit metric and not to EarthMM as a dataset, mission, or axion framework. The paper’s quoted description is that the similarity of two normalized histograms is measured by the number of items that must be moved and the distances they must move, with an additional cost function for different total counts [(Irwin et al., 2014); (Newberg et al., 2014)].

Other Earth-centered research directions in the supplied corpus are likewise separate. The Blender-based virtual morphometric globes for Earth, Mars, and the Moon use 15′-gridded global DEMs, a UV sphere, TIFF morphometric textures, and the Blender Game Engine to display curvature and catchment-area maps on interactive planetary globes; this is a geomorphometric visualization system, not an EarthMM framework. Testing MOND on Earth is an Earth-based low-acceleration laboratory program centered on the MOND scale

$t$3

with proposals such as the SHLEM effect, CCC, special trajectories, rotating-ring setups, and satellite tests; it is Earth-based, but not an EarthMM usage [(Florinsky et al., 2015); (Ignatiev, 2014)].

The main interpretive caution is therefore straightforward. When EarthMM appears in current technical writing, it cannot be assumed to name a single object class. In remote sensing it is a multimodal dataset; in axion physics it is an environment-aware search framework; in several astronomy and space-systems summaries it functions as a contextual shorthand for Earth-mass planets or Earth-centered architectures. Any rigorous reading must disambiguate the term from neighboring labels such as MEarth, MMEN, Earth-Mover Distance, or Earth-based MOND tests before drawing substantive conclusions.