TerraNova

A defining problem of the Anthropocene is to model the physical Earth and human societies as one coupled system, yet no learned representation spans their observational breadth. We argue the obstacle is geometric: the physical Earth is measured as continuous fields that ignore political borders, whereas societies are reported for administrative units. Earth-system foundation models serve the first geometry; coupling it to the second has required lossy averaging over borders. TerraNova is trained on 1,024 physical and societal records in their native geometries: 512 gridded Earth-system fields and 512 national indicators.

• Paper: https://arxiv.org/abs/2607.29527
• Code: https://github.com/crp94/terranova-model
• Project page: https://carlosrodriguezpardo.es/projects/TerraNova/

Model summary

TerraNova encodes location, country, time and task with dedicated encoders, fuses them with cross-modal transformers into a shared spatiotemporal state, and generates a per-query decoder with a hypernetwork whose evidential (Normal-Inverse-Gamma) head returns a predictive distribution. The released checkpoint is the frozen backbone: 501 tensors, 363,431,610 parameters, fp32 safetensors, 1.45 GB. It was trained on 512 gridded Earth-system fields and 512 national indicators (1,025 trained task rows including one internal alignment row, plus 256 rows reserved for adaptation) across 247 ISO3 countries (including the synthetic OCN open-ocean row) on the 0.25-degree (721 x 1440) WorldTensor grid.

Files

• model.safetensors: the frozen backbone weights (fp32).
• config.json: architecture configuration (ArchConfig) and export provenance.
• tasks.json: the 1,025 trained task rows: id, index, modality (gridded/country/ internal), and bare key.
• countries.json: the 247 ISO3 country rows: iso3, index, and a synthetic flag (set for OCN).

How to use

pip install git+https://github.com/crp94/terranova-model

from terranova import TerraNova

model = TerraNova.from_pretrained("crp94/terranova")
p = model.predict("t2m_mean", coords=[[12.5, 41.9]], year=2015)   # standardised units
print(p.mean, p.sigma)

E = model.embed(coords=[[12.5, 41.9]], year=2015, space="spatiotemporal")   # [1, 256]

See the code repository for terranova.Adapter (fit a new variable on the frozen backbone with rank-4 MiSS in minutes on a laptop) and worked examples.

Conventions

• Coordinates are [lon, lat] in degrees.
• Years are integers in [1900, 2035]. The observed training record ends in 2025; later years are extrapolation.
• ISO3 codes must be in the checkpoint's registry (model.countries); nothing is auto-registered. Use Natural Earth's ISO_A3_EH column, never the raw ISO_A3, which carries the -99 sentinel for France and Norway.
• Outputs on trained tasks are in standardised (z-score) units, not physical units, because the per-task normalisation statistics are not shipped with this release. Adapted tasks return physical units.
• The WorldTensor grid is 721 x 1440 nodes at 0.25 degrees, latitude -90..90, longitude -180..179.75 ascending.

Provenance

• Run: evidential_full_512_drop015_tr2048_s42.
• Checkpoint: EMA weights, epoch 126, global step 4,195,121.
• Export date: 2026-09-09.
• Source checkpoint SHA-256 (from config.json's provenance.source_sha256): e2a24956a8963d50b050e2de94c340b3a72f3b75a7b01991c8f7d89dfe4e7dee.

Intended use and limitations

TerraNova is intended for research use: embedding locations, countries and time; predicting the 1,024 trained variables; and adapting the frozen backbone to new variables with a small amount of labelled data. It is not a causal model and its predictions should not be used as the sole basis for decisions affecting people or ecosystems without independent validation.

• Trained-task predictions are standardised units; physical units require the training statistics, which are not released in this version.
• Years beyond 2025 are extrapolation; the encoder's range ends at 2035.
• Downscaling recovers only within-country structure and only for variables that vary within countries; institutional indicators have no meaningful sub-national field.
• The OCN pseudo-country is a training artefact for open ocean and is flagged, not hidden.

Citation

@article{rodriguezpardo2026terranova,
  author        = {Rodriguez-Pardo, Carlos and Tavoni, Massimo},
  title         = {TerraNova: A Foundation Model for the Anthropocene},
  year          = {2026},
  eprint        = {2607.29527},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2607.29527}
}

Licence

CC BY 4.0. See LICENSE-WEIGHTS.md in the code repository. Attribution: cite arXiv:2607.29527.