NASA-IBM Lunar Foundation Model (NASA-IBM LFM)

A multimodal, multi-resolution foundation model for lunar remote sensing.

A ViT-B encoder–decoder trained from scratch on SomBench — ~2 million co-registered lunar tile bundles spanning 11 modalities at two spatial scales (LROC NAC ~1 m/px and LROC WAC ~100 m/px). It adapts the TerraMind masked-token recipe to the Moon with two extensions:

Acquisition geometry as explicit context — per-tile illumination angles, solar-frame anchors, and tile footprint are sequence-tokenized encoder inputs. Lunar surface appearance is governed more by illumination geometry than by intrinsic surface variation, so this hands the model the dominant confound instead of making it recover a quantity already recorded for every tile.
Joint mixed-resolution pretraining — NAC- and WAC-anchored tiles train in a single mixed-batch loop at native resolution, so one set of weights covers both resolution families across a 100× scale gap.

FlexiViT patch embedding lets the checkpoint be fine-tuned at other patch sizes without retraining the backbone; modality-wise tokenization lets you drop or add modalities at fine-tuning time.

📄 Full method, ablations, and per-baseline results: paper · 💻 Fine-tuning code: GitHub

Modality-wise tokenization (left/center-left) feeds a TerraMind-style any-to-any masked-token objective (center-right); the pretrained encoder then supports multimodal generation, multimodal fine-tuning, and FlexiViT fine-tuning at new patch sizes (right).

Model at a glance


Encoder / decoder	ViT-B (768 dim, 12 layers, 12 heads) / 12-layer Transformer, shared width
Input size / patch size	256 × 256 px / 16 × 16 at pretraining (FlexiViT-resizable; results below use ps8)
Modalities	11 — 9 dense image-like, plus per-tile optical metadata and static-map context
Objective	Cross-entropy over discrete token vocabularies at sampled target positions
Tokenizers	9 modality-specific VQ-VAE, FSQ quantization (levels 8,8,8,6,5), DDPM decoder
Pretraining	16 × H100, 150k steps, global batch 1,536, peak LR 1e-4 cosine, bf16, ~1.1k GPU-h
Data	SomBench — 963,609 WAC + 1,000,113 NAC bundles, split geographically by LTM zone / LPS cap
Downstream adaptation	TerraTorch
License	Apache-2.0

Repository contents

backbone/         pretrained checkpoint and config to be used for finetuning/generation
tokenizers/       checkpoints and configs for all tokenizers

Intended use

A reusable representation for lunar remote sensing: fine-tune or LoRA-adapt the encoder for detection, segmentation, and dense regression on LROC WAC/NAC imagery and co-registered terrain products, with modality subsets and patch grids that differ from pretraining. Any-to-any multimodal generation is supported as a qualitative probe of learned cross-modal structure.

Out of scope. Not a scientific-grade generative product — generated fields are not calibrated predictions and are no substitute for instruments, stereo photogrammetry, or geodetic solutions. The model maintains no geodetic reference frame: it recovers local structure but not absolute values (DTM shape is right at a shifted elevation offset; generated lat/lon can be off by tens of degrees). Not validated for operational decisions such as landing-site certification or hazard clearance. Ice-prospectivity outputs regress a knowledge-driven fuzzy-overlay prospectivity map, not measured ice. Not evaluated beyond the Moon or on products absent from SomBench.

Modalities

Dense modalities are patch-embedded and tokenized as images. A sample is either NAC- or WAC-centered — the two families are never mixed within a sample.

Family	Modality	Key	Native res.	Ch.	Source
WAC (51.2 km tiles)	Visible reflectance	`vis`	100 m	5	LROC WAC
	Ultraviolet reflectance	`uv`	500 m	2	LROC WAC
	Topography / slope / aspect	`dtm`, `slope`, `aspect`	60 m	1 / 1 / 2	SLDEM2015
NAC (512 m tiles)	Panchromatic imagery	`nac`	1 m	1	LROC NAC
	Topography / slope / aspect	`dtm_3m`, `slope_3m`, `aspect_3m`	3 m	1 / 1 / 2	NAC-stereo DTM

Aspect is stored as a sine–cosine pair to avoid angular wrap-around.

Two sequence-like modalities carry scalar context, binned and stringified before text tokenization (e.g. with 0.25° longitude bins, a tile-center longitude of 87.49° becomes C_LON=87.25-->87.50):

Optical metadata (8 fields) — solar incidence, emission, phase and azimuth angles; sub-solar and tile-center coordinates; ground sampling distance. Available at inference whenever the image is, so no label leakage.
Static-map context (28 fields, tile-footprint averages) — Diviner thermophysics (TREG, TBOL, ROCK_ABUND, HPAR, DICE), LOLA products (ROUGHNESS, PSR, AVG_ILLUM, ALBEDO), Mini-RF radar (MINIRF_*), Kaguya mineralogy (MI_*, SW_FE, SP_MINER), LROC WAC (nr643, TIO2, WAC_*), GRAIL (GRAVITY), and Lunar Prospector (HYDROGEN). Coverage varies by instrument footprint, from near-global to polar-only (~6,900 polar tiles per track).

Usage

Fine-tuning runs through TerraTorch, configured from YAML. The companion repo adds the SomBench benchmark datamodules, backbone wrappers, support for registering new input modalities into the pretrained backbone, and FlexiViT patch-embedding interpolation.

git clone https://github.com/NASA-IMPACT/NASA-IBM-Lunar-Foundation-Model && cd $_
pip install -e .
terratorch fit --config configs/finetune/ice_prospectivity.yaml

Point the backbone checkpoint field at the downloaded checkpoint from this repo. Exact registry names and full hyperparameters are in the GitHub.

Checkpoint can be downloaded using Python:

from huggingface_hub import snapshot_download

# Only download model weights and config
snapshot_download(repo_id="nasa-ibm-ai4science/NASA-IBM-Lunar-Foundation-Model", allow_patterns="backbone/*", local_dir="./")

# Download entire HuggingFace directory
snapshot_download(repo_id="nasa-ibm-ai4science/NASA-IBM-Lunar-Foundation-Model", local_dir="./")

Which adaptation strategy? From our experiments:

LoRA is the sensible default (rank 16, α = 32, on encoder attention + MLP linears). It matches or beats full fine-tuning on crater detection, stays competitive on segmentation, trains a small fraction of the encoder parameters, and shows narrower seed spread.
Full fine-tuning — it keeps an edge on our two smallest benchmarks. Use layer-wise LR decay with a lower encoder LR than head LR, so pretrained features aren't pushed away early.
Frozen encoder only after validating on your task: best on 100-tile IMP segmentation, but below every baseline on crater detection.

Evaluation

Four SomBench benchmarks via TerraTorch, with loaders, splits, augmentations, loss, and metric held fixed across backbones — only encoder weights and initialization differ. Mean ± std over 5 seeds (3 for the ablation). Baselines are ResNet-50, ViT-B MAE, ConvNeXt-B, ConvNeXt-V2-B, SwinV2-B, DaViT-B (ImageNet-1k/22k), plus DeepLabV3+ and SegFormer on segmentation. Full tables are in the paper.

Benchmark	Metric	Best NASA-IBM LFM	Best baseline	Random-init LFM
Robbins craters, WAC, 50% data	mAP ↑	0.2541 ± 0.0018 (full FT)	0.2313 ± 0.0027 (SwinV2-B)	0.2197 ± 0.0027
Robbins craters, WAC, 100% data	mAP ↑	0.2581 ± 0.0017 (LoRA)	0.2420 ± 0.0047 (SwinV2-B)	0.2289 ± 0.0037
NAC craters, meter scale	mAP ↑	0.1543 ± 0.0098 (LoRA)	0.1552 ± 0.0086 (SwinV2-B)	0.1274 ± 0.0151
Irregular Mare Patches	IoU₁ ↑	0.5709 ± 0.0114 (frozen)	0.5687 ± 0.0181 (ConvNeXtV2-B)	0.3142 ± 0.0746
Polar ice prospectivity	RMSE ↓	0.0293 ± 0.0013 (full FT)	0.0377 ± 0.0004 (SwinV2-B)	0.0397 ± 0.0004

Label efficiency. On WAC craters the pretrained variants at 50% training data already match or exceed SwinV2-B trained on the full set. The ranking is stable across fractions and the gap exceeds seed spread.
Widest margin: ice prospectivity, and it decomposes. The LFM gives each modality its own pretrained patch adapter and concatenates tokens along the sequence axis (native token-level late fusion), whereas baselines can only stack the 8 layers as channels at a shared stem. The random-init control shares that tokenization without lunar pretraining and already beats 5 of 6 ImageNet baselines (0.0397); pretraining accounts for the rest. In a modality-count ablation the LFM with only 3 modalities (aspect, slope, DICE) matches ConvNeXt-B using the full 8-layer stack (0.0434 vs 0.0437).
NAC craters and IMP segmentation: treat leaders as comparable — margins are smaller than the seed spread. On NAC, scores are low across the board (part of the benchmark is annotated at 5 m/px and visibly blurrier) and the frozen variant performs like random init, so some encoder adaptation is needed at meter scale. On IMP the decisive signal is the random-init control collapsing to IoU₁ 0.31 — pretraining is essential in that regime.

Qualitative generation. Chained generation confirms the intended cross-modal correspondences: slope and aspect from DTM at both scales; reflectance whose illumination pattern follows the supplied metadata (shadows align with the sun direction implied by sub-solar coordinates) even where it doesn't match pixel-for-pixel; and coarse spatial pattern across the full co-registered set. Absolute values drift (see Out of scope). These are qualitative sanity checks, not a systematic evaluation.

Limitations

Ablations not yet isolated. Separating the contribution of geometry tokenization and mixed-resolution training from lunar pretraining as a whole remains to be run; the ice-prospectivity random-init control is the one place architecture and pretraining are teased apart.
NAC pretraining is site-limited by construction — restricted to NAC PHO sites and frames with co-registered 3 m stereo DTMs (1,095 frames): globally distributed, not globally dense.
No geodetic reference frame; ice-prospectivity target is a model rather than a measurement (see Out of scope).

Citations

@article{fraccaro2026lfm,
  title  = {Multimodal-Multiresolution Foundation Model for Lunar Remote Sensing},
  author = {Fraccaro, Paolo and Nyirjesy, Gabby and Szwarcman, Daniela and Patil, Himanshu
            and Gaur, Vishal and Lal, Rohit and Slank, Rachel A. and Dawson, Geoffrey
            and Debary, Hiyam and Dionelis, Nikolaos and Barker, Michael K. and Annex, Andrew
            and Viswanathan, Vishnu and Morse, Zachary and Schaefer, Ethan I. and Kumar, Ankur
            and Watson, Campbell D. and Dawson-Rigas, Rebekah I. and Maskey, Manil
            and Roy, Sujit and Ramachandran, Rahul and Bernab\'e-Moreno, Juan},
  year   = {2026}
}

@misc{sombench2026collection,
  author      = {Patil, Himanshu and Nyirjesy, Gabby and Slank, Rachel A. and Gaur, Vishal
                  and Szwarcman, Daniela and Fraccaro, Paolo and Dionelis, Nikolaos and Barker, Michael K.
                  and Annex, Andrew and Viswanathan, Vishnu and Morse, Zachary and Schaefer, Ethan I.
                  and Debary, Hiyam and Kumar, Ankur and Lal, Rohit and Dawson, Geoffrey
                  and Watson, Campbell and Dawson-Rigas, Rebekah I. and Maskey, Manil
                  and Bernab\'e-Moreno, Juan and Ramachandran, Rahul and Roy, Sujit},
  title        = {{SomBench}: Benchmark Dataset for Advancing Machine Learning in Lunar Science},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/collections/nasa-ibm-ai4science/lunar-fm-ml-ready-benchmark-dataset-sombench}}
}

Acknowledgments

Supported by NASA under Award No. 80MSFC25M0084. Data products courtesy of LROC, LOLA, Diviner, Mini-RF, Kaguya/SELENE, GRAIL, Lunar Prospector, and USGS.

Contact: Sujit Roy (sujit.roy@nasa.gov) · Paolo Fraccaro (paolo.fraccaro@ibm.com)