Moonatlas

Online

Moonatlas / Science · science-0.3.1

The science behind
MOONATLAS

Real lunar observations. Real model outputs. Reproducible processing.

MOONATLAS runs publicly released NASA-IBM Lunar Foundation Model checkpoints over the public SomBench benchmark samples, offline and on one GPU, then publishes the results as a static catalogue you can explore geographically. The web application never runs a model: it reads what the pipeline produced, with the provenance of every value.

The Moon, LROC colour map from the NASA SVS CGI Moon Kit
Surface: NASA SVS CGI Moon Kit · LRO LROC

WAC tiles analyzed

1,000

SomBench crater strips

Crater model detections

80,454

not unique craters

ICE patches analyzed

156

78/81 north · 78/81 south

IMP observations

130

104 targets

Model predictions

80,740

detections, maps and masks

Curated discoveries

97

42 held-out · 55 in-sample

01

Why MOONATLAS exists

MOONATLAS was built when the ingredients became publicly available. What happened upstream is marked UPSTREAM FACT; why MOONATLAS exists is our own motivation, and is marked as such.

Chapter 01

The release

Upstream fact

NASA

IBM

Lunar Foundation Model

Multimodal · multiresolution

SomBench

Lunar task benchmark

Public models

Datasets

Checkpoints

Open scientific building blocks

Reusable outside the original teams

NASA and IBM released a multimodal, multiresolution foundation model for lunar remote sensing, together with SomBench — a benchmark of lunar science tasks — and trained checkpoints for those tasks.

That release is what made this project possible: lunar observations, task datasets and trained models became usable by people outside the original teams. MOONATLAS is one of those outside uses. NASA and IBM did not build it, review it or endorse it.

Upstream fact

Models
Apache-2.0
SomBench datasets
CC-BY-4.0
Published by
NASA-IBM AI4Science

Fraccaro, P. et al. (2026); Patil, H. et al. (2026). Full citations in data & sources.

Chapter 02

The opportunity

Lunar missions have returned enormous quantities of imagery and remote-sensing data. Working with it normally means specialist tools, geospatial processing, machine-learning infrastructure and familiarity with each individual dataset.

Reusable lunar AI changes what can be built on top of that archive. The training, the benchmarks and the task definitions are already public. What was missing was an interface.

What the missions returned

Lunar Reconnaissance Orbiter Camera

LROC WAC and NAC imagery

Lunar Orbiter Laser Altimeter

LOLA topography (slope, aspect, curvature)

Diviner Lunar Radiometer Experiment

Diviner thermal observations (maximum temperature, ice stability)

What it normally takes to use it

Specialist tools

Geospatial processing

ML infrastructure

Domain knowledge

Foundation models made another interface possible

Chapter 03

The question

MOONATLAS motivation

What if you could
point at the Moon
and ask:

“What do we know about this place?”

No GeoTIFF to download. No notebook to run. Point at the Moon and get an answer — including the answer “nothing has been analyzed here”.

This question became MOONATLAS.

The Moon rendered by the MOONATLAS explorer with the camera centred on 31.134°S 52.578°E

Target locked

31.134°S

52.578°E

Catalogued here · MM-CR-000014

Still from the MOONATLAS explorer · camera centred on this coordinate · reticle drawn by the explorer at that location

Chapter 04

The Moon as the interface

Normally encountered as

  • GeoTIFF rasters
  • Model checkpoints
  • JSON outputs
  • Benchmark tables
  • Disconnected imagery

Put back onto

The Moon itself
becomes the interface

Lunar missions

LRO instruments

Scientific observations

Imagery, topography, thermal

NASA-IBM lunar AI

Foundation model + task training

Public models + SomBench

Apache-2.0 · CC-BY-4.0

MOONATLAS processing

Offline inference, normalization

Geographic model outputs

Georeferenced, provenance kept

The Moon

+

Human exploration

MOONATLAS deliberately avoids presenting the science as files. Every output is placed back where it was observed, so the archive can be read the way the Moon is read: by looking at a place.

  1. RotateTurn the globe, nearside to farside
  2. TargetLock any coordinate, covered or not
  3. ExploreAsk what the catalogue holds there
  4. InspectOpen the tile, the field, the mask
  5. DiscoverFollow a curated candidate

Chapter 05

Why build it

MOONATLAS motivation

01

Accessibility

Lunar analysis without a geospatial toolchain

02

Exploration

A place to go, not a table to read

03

Scientific curiosity

Questions a visitor can ask themselves

04

Transparency

Every value says where it came from

MOONATLAS is not intended to replace lunar scientists. It explores whether advanced lunar AI outputs can be turned into an interface that makes people want to investigate the Moon themselves.

Make the science accessible
without pretending
it says more than it does.

No coverage

The report says there is no NASA-IBM LFM coverage here

Unmatched crater

Never called new, unknown or discovered

Ice prospectivity

Never called a probability of ice

Uncalibrated score

Labelled uncalibrated everywhere it appears

Where this stands

Today

Available now

Select any place

Check real coverage

Query catalogued model outputs

MOONATLAS report

80,740 catalogued model outputs, georeferenced onto the Moon. The report states the coverage that actually exists at the coordinate you picked, even when there is none.

The next frontier

Future vision

Select any place

Acquire lunar observations

Run lunar AI

New inference

MOONATLAS report

A future version could acquire the observations covering an arbitrary target and run a new inference for it. This does not exist today, and no date is promised.

02

The Lunar Foundation Model

A foundation model is trained once on a large amount of data and then adapted to specific tasks. The NASA-IBM Lunar Foundation Model (LFM) is trained on lunar remote sensing across several instruments and resolutions; task checkpoints adapt it to detect craters, to regress polar ice prospectivity and to segment irregular mare patches.

SomBench is the public benchmark that accompanies it: the lunar samples those tasks are trained and evaluated on. SomBench samples are the only places MOONATLAS analyses, which is also why coverage is limited — see what MOONATLAS has analyzed.

  1. Lunar observation

    LROC WAC / NAC, LOLA, Diviner

  2. NASA-IBM LFM

    Shared multimodal backbone

  3. Task model output

    Boxes · regression field · mask

  4. MOONATLAS normalization

    Georeferencing, raw + display

  5. Geographic exploration

    Layers on the Moon

  6. Discoveries

    Curated, permanent ids

Inference ran locally on a single GPU with the upstream configurations and strict checkpoint loading. Nothing in the interface is generated by a language model, and no value is invented or interpolated.

03

Three analysis systems

01

Crater detection

  1. LROC WAC imagery
  2. NASA-IBM LFM
  3. Crater candidates

80,454 detections

over 1,000 analyzed tiles

02

Polar ice prospectivity

  1. Multimodal polar observations
  2. NASA-IBM LFM
  3. Dense prospectivity field

156 patches

both poles, full valid grid

03

Irregular mare patch segmentation

  1. LROC NAC imagery
  2. NASA-IBM LFM
  3. Segmentation mask

130 observations

104 target groups

04

Craters: prediction and reference

The model predicts crater regions on a WAC tile as boxes with a confidence. MOONATLAS converts each box into lunar coordinates and an approximate diameter using the tile's own projection, then compares predictions with the Robbins (2019) global crater catalogue.

A model prediction and a reference annotation are different things and are never merged. A match means the prediction overlaps a catalogued reference crater under the configured criterion (IoU ≥ 0.5, greedy by confidence).

No match does not mean a new crater. It means no reference crater matched under that criterion — the reference catalogue has its own completeness limits, the box may be imprecise, and the feature may not be a crater at all. MOONATLAS labels those candidates “unmatched with Robbins” and never “new”, “unknown” or “discovered”.

Tabs switch the same tile between the source observation, the model prediction and the reference annotation; the highlighted box is the catalogued discovery.

LROC WAC tile M1244734512CE
AI prediction · NASA-IBM LFM1.0×

M1244734512CE · 117 predicted boxes · 144 Robbins (2019) reference boxes

Held-out WAC tile · predicted boxes and Robbins reference · MM-CR-000009

05

Polar ice: prospectivity, not probability

The ice task is a regression: from multimodal polar observations — illumination, thermal, topography — the model produces a dense field that emulates a knowledge-driven expert prospectivity map (Coyan et al., 2025) on a 0–1 target scale.

A value of 0.93 means high modelled prospectivity. It is not a 93% chance of ice, and it is not a measurement of ice. Prospectivity says “this terrain resembles the terrain the expert map favours”.

The regression is unconstrained, so raw outputs can fall slightly outside the target scale. MOONATLAS preserves every raw value and clips only the colours — hover the field and the readout quotes the raw model output, and says when the colour was clipped.

AI prediction · NASA-IBM LFM1.0×

patch_0003_0002_S_80S · 256 × 256 px · hover for the raw model value

Held-out polar patch · predicted prospectivity field · MM-ICE-000006

Raw model output · preserved

-0.136011.119

Display · clipped to the 0–1 target scale

0.001.00
Across the 156 analyzed polar patches the raw prediction runs from -0.1363 to 1.1189: 153 patches dip below 0 and 149 rise above 1. Readouts quote the raw value; only the colour is clipped, and the inspector says so when it happens.

06

Irregular mare patches: segmentation and an uncalibrated score

The IMP task segments the pixels of a 256 m LROC NAC tile that belong to an irregular mare patch candidate. From that mask MOONATLAS derives the predicted area, the pixel count, the share of the tile it covers and its centroid — all MOONATLAS derived metrics computed from the model output.

The mean class-1 softmax reported in the inspector is uncalibrated. It orders pixels; it is not a probability, and it must not be read as one.

An IMP candidate is not an active volcano. Irregular mare patches are small, unusual mare features whose origin is debated in the literature; the model segments a candidate feature, nothing more. Reference polygons come from Hargitai et al. (2025), as packaged in SomBench.

LROC NAC tile M1126915118RE
AI prediction · NASA-IBM LFM1.0×

M1126915118RE · 256 m tile · 1.00 m/px · masks at native resolution

Held-out NAC tile · predicted segmentation mask · MM-IMP-000001

07

MOONATLAS by the numbers

Everything below is read from the published catalogue at build time. Note the wording: 80,454 model detections, not craters discovered.

WAC tiles analyzed

1,000

SomBench crater strips

Crater model detections

80,454

not unique craters

Polar ice patches

156

both poles

Valid grid cells

78 / 81 · 78 / 81

north · south

IMP observations

130

LROC NAC tiles

SomBench target groups

104

repeat observations grouped

Model predictions

80,740

detections, maps and masks

Curated discoveries

97

42 held-out · 55 in-sample

08

What has MOONATLAS analyzed?

The models only ran where SomBench has samples. The Moon is fully explorable in MOONATLAS, but model coverage is not global, and the interface says so instead of implying analysis everywhere.

90°N90°S180°W180°E
WAC tiles · held-out testWAC tiles · in-sampleNAC IMP observations2° cells · polar ice is mapped separately on its own grids

Surface covered

3.88%

1.47M km²

Summed tile area

2.62M km²

tiles overlap each other

Geographic areas

12

strips, 300 km apart

Held-out test ground

0.66%

0.25M km²

The 926 distinct footprints of the 1,000 analyzed WAC tiles cover about 1.47M km² — 3.88% of the lunar surface — in 12 strips. Their areas add up to 2.62M km², and the difference is overlap: SomBench tiles share ground with each other.

That overlap is also why a detection count is not a crater count. 53,842 of the 80,454 detections (66.92%) have a similar detection in another tile. MOONATLAS does not deduplicate them, because each tile prediction is an authentic model output on its own observation.

Polar ice coverage

The polar task runs on a fixed stereographic grid per pole. MOONATLAS processed 78 of 81 cells in the north and 78 of 81 in the south: every cell that contains valid pixels inside the dataset extent. The remaining three cells per pole contain none, and they stay empty rather than being filled in.

This means all valid coverage available to this model and dataset configuration has been processed. It does not mean lunar ice has been mapped.

north pole78 / 81 cells

3 cells hold no valid pixels · cell brightness follows its valid-pixel share

south pole78 / 81 cells

3 cells hold no valid pixels · cell brightness follows its valid-pixel share

Held-out testValidationTrainingNo valid pixels

IMP coverage

130 NAC observations, grouped into 104 SomBench target groups — several tiles can image the same annotated target, and MOONATLAS groups them by that upstream target id for navigation only. The term is observation group, not “site”: the upstream data establishes which annotation a tile was framed on, not that two observations show the identical physical feature. Their masks are never merged.

These are geographically limited observations on the nearside mare, marked in amber on the map above. Nothing here is a global volcanic survey.

09

What happens when you explore an area

You can lock any coordinate on the Moon, including ground no model has ever seen. MOONATLAS queries the catalogue it published and reports what is actually there.

  1. You select a target

    Any lunar coordinate

  2. Coverage is checked

    Which models reach that ground

  3. Model outputs retrieved

    Catalogued predictions there

  4. Nearby analyses searched

    Nearest tiles and discoveries

  5. Exploration report

    Including “no coverage”

Where there is no NASA-IBM LFM coverage, the report says so: it names the surface imagery it can show, the distance to the nearest analyzed tiles and the nearest curated discoveries. It never invents a value for a place no model looked at.

In this version, exploring queries precomputed real model outputs. Pressing the button does not start a new GPU inference. The autonomous scan and the exploration feed are replays of catalogued results for the same reason.

Next frontier · not available today

Live on-demand lunar analysis

  1. Select any place

  2. Acquire lunar data

  3. Run lunar AI

  4. Analyze

  5. Report

A future version could acquire the observations covering an arbitrary target and run a new inference for it. This is a direction of travel, not a feature of this release, and no date is promised.

10

Held-out test and in-sample

Held-out test

Observations that were not used as training samples for the downstream task, kept aside for independent evaluation. Performance numbers — ours and the model cards' — come only from here.

135 analyzed samples · 42 discoveries

Training and validation · in-sample

Equally real observations and equally real model outputs, published because they extend geographic coverage and are worth exploring. They are not independent evidence of model performance, and MOONATLAS labels them everywhere they appear.

1,151 analyzed samples · 55 discoveries

Held out does not mean physically more important, and in-sample does not mean invalid or fake. The badge describes evaluation provenance, nothing else.

11

Reproduction check

Before publishing anything, MOONATLAS re-runs the released checkpoints with the upstream configuration over the held-out test split and compares the result with the published model cards. This is an engineering validation of the pipeline — evidence that inference, georeferencing and metrics were wired correctly — not a competition with the original teams.

WAC crater detection

100 held-out tiles
MetricModel cardMOONATLAS held-out
mAP0.25810.2602
AP@500.61830.6239

Upstream test configuration, recomputed over the 100 held-out tiles of this build.

IMP segmentation

10 held-out tiles
MetricModel cardMOONATLAS held-out
IoU class 10.57090.5824
F1 class 10.72680.7361

Ten held-out tiles: a small sample, so small differences from the card are expected in both directions.

The ice checkpoint publishes regression errors rather than a per-split benchmark table, so MOONATLAS reports its own RMSE and MAE against the reference map per patch instead of claiming a reproduction score.

Train and validation metrics exist in the same reports and are deliberately not shown here: they are in-sample and would overstate what the pipeline demonstrates.

12

Raw, normalized, display, derived

Four kinds of value exist in MOONATLAS and they never blur into each other. This is the rule the whole product is built on.

  1. Raw model output

    Exactly what the model produced

  2. Normalized science data

    Georeferenced, structured, validated

  3. Display representation

    Clipped and coloured for rendering

  4. MOONATLAS derived metric

    Areas, ranks, distances, navigation

Raw is the model result as produced, preserved even when it falls outside a target scale. Normalized is that result placed on the Moon with its projection, units and provenance, and validated against the published schemas.

Display is the transformation needed to draw it — clipping, colormaps, opacity. Derived is arithmetic MOONATLAS does on top, such as a predicted area or a diameter rank.

Display values are never reported as if they were raw model output, and MOONATLAS-derived values are never presented as NASA or IBM outputs. Every metric in the interface carries its origin badge.

Provenance carried by every object

Model
NASA-IBM Lunar Foundation Model
Task
WAC crater detection
Checkpoint
WAC_ni_lfm_ps8_lora_s46.ckpt
Dataset
sombench-wac-crater-detection
Split
Held-out test
Source
M1244734512CE_r4336_c240.tif
Processing
science-0.3.1
Data origin
Real model output

The same stack appears on every discovery page and in the explorer's inspector.

13

Limitations

What this system does not do, stated plainly. This list is part of the product, not a disclaimer.

  • Crater coverage is geographically limited

    Analyzed WAC tiles cover 3.88% of the surface in 12 strips. Elsewhere there is no crater model coverage.

  • Detections are not unique craters

    Source tiles overlap: 66.92% of detections have a similar detection in another tile, and they are not deduplicated.

  • Unmatched is not new

    A prediction without a Robbins match under the configured criterion is not evidence of a previously unknown crater.

  • Ice prospectivity is not probability

    It is a regression against a knowledge-driven expert map, not a measurement or a likelihood of ice.

  • Ice values extend outside 0–1

    Raw outputs run from -0.136 to 1.119; MOONATLAS preserves them and clips only the display.

  • IMP mean softmax is uncalibrated

    It ranks pixels within a mask. It is not a probability that a feature is an irregular mare patch.

  • An IMP candidate is not an active volcano

    The model segments a candidate volcanic feature on one NAC tile; origin and activity are not claimed.

  • Train and validation predictions are in-sample

    They extend coverage and exploration, and are never independent evidence of model performance.

  • References are not model output

    Robbins (2019), Coyan et al. (2025) and Hargitai et al. (2025) are published expert work, shown for comparison only.

  • Explore Area queries a catalogue

    In this version it searches precomputed model outputs; it does not execute live inference.

14

Data & sources

Everything MOONATLAS analyses belongs to someone else. Upstream models are Apache-2.0, SomBench datasets are CC-BY-4.0 (attribution required, changes indicated), and reference annotations belong to their authors. MOONATLAS's own code is Apache-2.0.

Models

Datasets

Reference annotations

Instruments

Imagery

Changes MOONATLAS makes to upstream data — crops, normalization, georeferencing, display encodings — are indicated by the “modified” flag above and described throughout this page. No NASA or IBM insignia are used, and no endorsement is implied.

Ready to look at the Moon? Open the explorer or browse the discoveries.