Pith. sign in

REVIEW 3 major objections 4 minor 79 references

TerraNova learns one representation across 1,024 gridded and national variables, coupling the two geometries through a population-weighted contrastive objective.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 05:16 UTC pith:DWTZONLE

load-bearing objection A serious, unusually honest systems paper whose headline cross-geometry claim is partly a product of its own training objective; worth a careful referee, not a desk reject. the 3 major comments →

arxiv 2607.29527 v1 pith:DWTZONLE submitted 2026-07-31 cs.LG cs.AIcs.CYecon.EMstat.ML

TerraNova: A Foundation Model for the Anthropocene

Carlos Rodriguez-Pardo , Massimo Tavoni This is my paper
classification cs.LG cs.AIcs.CYecon.EMstat.ML
keywords foundation modelrepresentation learningclimate changeEarth systemuncertainty quantificationmulti-geometry alignmentdownscalingcontrastive learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper's central claim is that the barrier to modeling the physical Earth and human societies as one coupled system is geometric: physical measurements are continuous fields that cross borders, while societal statistics are reported for administrative units. TerraNova is a foundation model trained on 1,024 variables—512 gridded Earth-system fields and 512 national indicators—in their native geometries within a single backbone. Two contrastive objectives couple those geometries: one aligns each country with its territory through a population-weighted embedding average, and the other anchors locations to pretrained geospatial embeddings. If the approach works, a single frozen model can encode both geometries, support cross-geometry retrieval and national-to-gridded downscaling, reconstruct dense fields from sparse observations, and return calibrated predictive distributions for every query—a joint observational layer for climate-economy research that also spans time, oceans, and uncertainty.

Core claim

TerraNova is, to the authors' knowledge, the first model to learn one representation across the observational breadth of the Anthropocene—1,024 physical and societal variables in their native geometries within a single backbone. The model maps a query (task, spatial unit, year) to the parameters of a Normal-Inverse-Gamma predictive distribution, using dedicated encoders for location, country, time, and task, cross-modal transformers that fuse them into a shared spatiotemporal state, and a hypernetwork that generates a per-query decoder. Two contrastive objectives couple the geometries: a population-weighted alignment between each country and coordinates sampled within its territory, and an a

What carries the argument

The central mechanism is the inter-geometry coupling achieved by two InfoNCE contrastive losses. The country–location alignment (the load-bearing one) aligns each country's spatiotemporal embedding to the mean embedding of coordinates sampled inside its borders, with sampling density proportional to log(1 + population), making the national summary concentrate where people live. The second loss aligns location embeddings to pretrained geospatial embeddings, distilling image-derived semantics without requiring imagery at inference. Around these, the architecture supplies a shared spatiotemporal state (a 256-dimensional 'common bus'), a hypernetwork that generates per-query decoders, and an evi

Load-bearing premise

The model's cross-geometry capabilities rest on the assumption that a country's statistics are well summarized by the population-weighted average of the locations inside its borders, so for variables whose signal is institutional or historical rather than spatial, that coupling could be an artifact.

What would settle it

Train a version with the country–location alignment replaced by a uniform spatial mean (no population weighting) and compare cross-geometry retrieval and downscaling on a mix of demographic and institutional targets; if institutional targets drop to chance while demographic targets hold, the coupling is a population-geography artifact. Equivalently, check whether the model's recovered development axis disappears when the population-weighting is removed.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single trained backbone can be reused to attach new variables with lightweight adapters (~4.2e5 parameters), reaching roughly 0.95 R² on gridded tasks from few labels in under a minute on a laptop.
  • National statistics can be downscaled to 0.25° gridded fields (e.g., terrestrial GPP, tree cover), recovering within-country structure that national means miss, with per-cell uncertainty.
  • Dense fields can be reconstructed from sparse lattices (down to 1/4096 of cells) better than classical interpolation, with predictive width that widens as data thins.
  • Cross-geometry retrieval works in both directions: a territorial embedding names its country at high recall, and downscaling writes a national quantity back onto the grid.
  • Every query returns a predictive distribution that, after per-task recalibration, achieves nominal coverage and flags extrapolation (wider intervals at forecast years).

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the country–location coupling generalizes as claimed, the same contrastive scheme could be applied to any hierarchical partition—regions, watersheds, or land-use zones—providing a template for coupling unit-based statistics with continuous fields beyond nations.
  • The representation could serve as a learned prior for climate-econometric models, e.g., by initializing damage-function regressions with the field–indicator couplings, even though the authors explicitly disclaim causal interpretation.
  • Because the geospatial alignment teachers are land-only, the ocean capability documented in the paper may be partly a placebo; an explicit marine-imagery alignment would likely improve ocean performance and is directly testable.
  • The 'emergent development axis' may reflect the population-weighted alignment prior rather than independent structure; an ablation that replaces population weighting with uniform territorial sampling would help separate these, though the paper does not run it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. TerraNova is a foundation model trained jointly on 512 gridded Earth-system fields and 512 national socioeconomic indicators, using dedicated location, country, time, and task encoders, cross-modal transformers, a hypernetwork-generated per-query decoder, and a Normal–Inverse–Gamma evidential head. Two contrastive objectives couple the geometries: a population-weighted country–location InfoNCE alignment (Lcl) and an alignment to pretrained geospatial embeddings (Lgeo). The paper reports evaluations on held-out variables, countries, and years: static prediction against geospatial encoders, temporal interpolation/nowcasting, sparse-field reconstruction, country-level extrapolation, national-to-gridded downscaling, uncertainty calibration, and computational cost. The central claim is that TerraNova is the first model to learn one representation across the observational breadth of the Anthropocene in a single backbone.

Significance. If the cross-geometry coupling is genuine, TerraNova would be a valuable resource for climate–society research: a single frozen backbone supporting retrieval, downscaling, sparse reconstruction, and calibrated predictive distributions would lower the cost of many downstream studies. The paper has notable strengths: held-out splits on variables, countries, and years; multiple seeds; a crossed ablation with a natural ocean placebo; honest reporting of cases where TerraNova loses or is not capacity-matched; and explicit caveats about calibration and the non-causal nature of the representation. These are real assets. However, the evidence for the central 'one representation' claim is weakened by circularity in the main coupling demonstration and by the narrow scope of the downscaling test, so the paper needs substantial additional experiments before the headline claim is established.

major comments (3)
  1. [§5.2 and §4.2, Eq. (9)] The cross-geometry retrieval test re-tests exactly the country–location InfoNCE objective Lcl used in training. The query is the mean embedding of population-weighted coordinates in a country's territory, and the correct answer is the country embedding — precisely the positive pair optimized by Eq. (9). Recall@1 of 0.875 is therefore a sanity check, not an independent demonstration that the two geometries share a generalizable latent space. The ablation in Fig. 9a shows Lcl is necessary, but does not show that the coupling transfers beyond the training objective. Please add a held-out cross-geometry task that was not directly optimized: for example, predict a held-out national indicator from territorial coordinates of a variable not used in training, or compare against a baseline in which the country embedding is the mean of its territory's location embeddings without any learned alignme
  2. [§6.5, Fig. 16] The downscaling demonstration uses only four biosphere targets (GPP, fAPAR, LAI, tree cover), all of which are gridded pretraining targets with strong climate/vegetation spatial structure. The positive within-country correlations may reflect the model's pretrained knowledge of these specific fields rather than a general societal–physical coupling. No held-out societal variable is downscaled, and §7 itself states that downscaling governance variables is 'not always meaningful.' The label-shuffle null is weak; please include a flat-national-mean null and a spatial interpolation baseline, and test at least one held-out gridded target or one societal indicator whose national values are known and which has an independent gridded reference, to support the claim that the coupling supports national-to-gridded downscaling as a general capability.
  3. [§6.1, Fig. 11 and §7] The headline benchmark compares TerraNova with a 416k-parameter MiSS adapter and generated decoder against frozen baselines read out by a linear probe. This is not capacity-matched, as the paper acknowledges in §7. The like-for-like comparison (TN-probe) ranks 7th of 9, so the reported lead is attributable to the adapter/decoder rather than to the frozen shared representation. Please add a matched-capacity baseline (e.g., an MLP probe or the same MiSS adapter applied to frozen baseline embeddings) or, failing that, revise the abstract and contribution claims to state that the frozen backbone plus cheap adapter is competitive. Also, the text says the adapted read-out receives coordinates only and no time; the ocean advantage is then attributed to time/uncertainty axes the baselines 'cannot express' (§6.1), which is confusing — please clarify what actually drives the marine-target margin.
minor comments (4)
  1. [Title/author line] Spacing errors: 'TerraNov a' and 'Ta voni' appear in the header. Please fix.
  2. [§4.2, Eq. (9) and text] The sampling density log(1 + P_y(x)) is mentioned, but P_y(x) is not explicitly defined in the main text; please define it as the population density for year y at location x.
  3. [§6.5 and Fig. 9b] The text in §6.5 says the null is a label shuffle, while the caption of Fig. 9b says '0 = flat national value.' Please reconcile the two descriptions so the null is unambiguous.
  4. [§6.1] The claim that the margin on marine targets is due to axes 'they do not represent' (time, uncertainty) is not supported by the setup, since the adapted read-out receives only coordinates and no time. Please reword or supply the missing analysis.

Circularity Check

1 steps flagged

Cross-geometry retrieval re-tests the same country–location InfoNCE objective used in training, so part of the coupling evidence is self-definitional; most other evaluations are genuine transfer.

specific steps
  1. self definitional [§4.2 (country–location alignment) and §5.2 (cross-geometry retrieval)]
    "Country–location alignment(Lcl) acts within TerraNova’s own shared space: it aligns the country’s spatiotemporal embedding with the mean embedding of coordinates sampled inside its territory, against other countries as negatives. ... we take the mean embedding of the coordinates inside a country and ask which of a held-out pool of 80 countries it names. ... what this measures is whether a country’s own code sits where its territory sits."

    The retrieval query is the same object as the Lcl positive pair: a territorial mean of coordinate embeddings queried against the country embedding, with other countries as negatives. InfoNCE training directly maximizes similarity between each country code and its territorial mean, so the reported recall@1 0.875 is the training objective evaluated on its own construction, not an independent emergent-coupling test. The matched ablation that drops Lcl only demonstrates that the objective does what it was defined to do. This makes one of the headline coupling demonstrations self-definitional, while the paper's other held-out-variable, held-out-year and held-out-country benchmarks remain genuine transfer.

full rationale

Most of TerraNova's evaluation is externally meaningful: §6.1 uses held-out environmental targets against geospatial baselines, §6.2 holds out years, §6.3 reconstructs unseen fields from sparse measurements, §6.4 holds out countries, and §6.5 uses biosphere targets. These do not reduce to the training losses. The one clear reduction is §5.2: the cross-geometry retrieval metric is the country–location InfoNCE objective of §4.2 read back out, so high recall is expected by construction and should not be cited as evidence that the coupling emerged beyond its training signal. The paper is transparent that compatibility is 'built by explicit alignment rather than emerging from data alone' (Discussion), but it still presents the retrieval as validation of the coupling. The self-citations (WorldTensor, neural conditional transport maps) are data/architecture references and not load-bearing for the central claim. Overall the core derivation is not circular: the model genuinely learns a shared representation from 1,024 reconstruction tasks with two auxiliary alignment losses, and most capabilities are tested on held-out data. Score reflects the one self-definitional validation result, not the central claim.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

No new physical entities are postulated. The model introduces a learned latent state, two alignment objectives, and the CountryTensor dataset, but these are model artifacts or data assets rather than hypotheses with independent falsifiable handles. The main burden is carried by the six domain assumptions above, especially the shared-statistical-structure premise and the population-weighted territorial summary.

free parameters (6)
  • Contrastive/reconstruction loss weights lambda_rec, lambda_tr, lambda_geo, lambda_cl = not reported in main text
    Fixed weights in Eq. (8) balance reconstruction and the two alignment objectives; values are chosen by hand and are not stated, affecting how much coupling vs reconstruction shapes the representation.
  • Task-sampling floors and bounded boosts = not reported
    §4.3 per-domain floors and bounded task boosts control how the 512 gridded vs 512 country tasks are sampled; without values, the training mixture cannot be reproduced.
  • Active-resampling multiplier bound = not reported
    §4.3 active task resampling reweights tasks by predictive variance; the bound and smoothing schedule are hand-chosen and affect final calibration/performance.
  • Conformal recalibration constant per task = implied by q=0.18 in Fig. 12 for one task; learned per task
    §5.3: raw NIG intervals are too wide and are corrected with one recalibration constant per task; reported 90% intervals depend on this fitted constant.
  • MiSS adapter rank = rank 4
    §4.4/§6.1: rank-4 MiSS residuals (~416k parameters) are used for all adaptation experiments; rank is chosen by hand and is part of the performance claims.
  • Population-sampling log transform = log(1+P_y(x))
    §4.2: sampling density follows log(1+P_y(x)); the log compression is a design choice shaping the country–territory summary.
axioms (6)
  • domain assumption The 512 gridded fields and 512 national indicators share a statistical structure that can be decoded into accurate predictions across both geometries.
    Stated in §3.1: 'We assume the two data types share a statistical structure...' This is the premise of the whole multi-geometry model; if false, no single latent space exists.
  • domain assumption Country-level indicators selected from the Quality-of-Government ecosystem are sufficiently reliable and commensurable after per-task standardization.
    §4.1: CountryTensor relies on the QoG ecosystem; governance/institutional indices are constructed, normative, and unevenly measured. The paper itself notes this in Broader Impact.
  • domain assumption Pretrained geospatial embeddings (SatCLIP, GeoCLIP, Copernicus-Embed) are meaningful semantic anchors for location, despite being land-biased.
    §4.2 external alignment distills these banks; the paper acknowledges the banks are land-only and uses ocean as placebo.
  • ad hoc to paper Population-weighted mean coordinate embeddings adequately summarize a country's territory for the country–location alignment.
    §4.2: sampling density log(1+P_y(x)); collapses within-country spatial heterogeneity to a single summary.
  • domain assumption NIG predictive distribution with per-task standardisation provides meaningful aleatoric/epistemic proxies; raw scale is uncalibrated but ordering is informative.
    §3.4/§5.3; authors explicitly say the component split is not separately identifiable and raw intervals require recalibration.
  • domain assumption Held-out variables absent from the 1,024 training set are sufficiently related to training tasks for a fresh task-row/MiSS adapter to transfer.
    §4.4/§6.3-6.4; adaptation claims are based on this transfer assumption.

pith-pipeline@v1.3.0-daily-deepseek · 23557 in / 19794 out tokens · 212469 ms · 2026-08-03T05:16:40.107467+00:00 · methodology

0 comments
read the original abstract

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. We introduce TerraNova, a foundation model trained on 1,024 physical and societal records in their native geometries: 512 gridded Earth-system fields and 512 national indicators. Dedicated encoders represent location, country, time and task, cross-modal transformers fuse them into a shared spatiotemporal state, and a hypernetwork generates a per-query decoder whose evidential head returns a predictive distribution. Two contrastive objectives couple the representation: a population-weighted alignment between each country and coordinates in its territory, and one to pretrained geospatial embeddings carrying image-derived semantics. Read out through that decoder, the representation is competitive with purpose-built geospatial encoders while spanning axes they do not represent (time, oceans and uncertainty) and supporting country-level capabilities. The frozen backbone reconstructs dense fields from sparse observations and adapts to unseen variables in minutes on consumer hardware.

Figures

Figures reproduced from arXiv: 2607.29527 by Carlos Rodriguez-Pardo, Massimo Tavoni.

Figure 1
Figure 1. Figure 1: The physical Earth is observed as continuous fields, whereas societal indicators are [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: TerraNova architecture. Dedicated encoders for location, country, time and task feed geometry routing and temporal fusion; a task-conditioned hypernetwork then generates the local decoder, whose NIG head returns a predictive distribution. 3.2 Encoders Four purpose-built encoders map location, country, time and task to unit-radius embeddings (Eq. 2). Location. The location encoder serves two regimes at once… view at source ↗
Figure 3
Figure 3. Figure 3: Training corpus. 1,024 variables in their native geometries: 512 gridded Earth￾system fields and 512 country-level indicators. The two terms are useful read-outs: the first tracks observation and data variability, the second uncertainty in the predicted mean; but they are not perfectly separately identifiable from the marginal likelihood without additional assumptions (Meinert et al., 2023). We therefore t… view at source ↗
Figure 4
Figure 4. Figure 4: (a) The location embedding on a Robinson projection (PCA, then UMAP McInnes et al. (2018) to three dimensions, mapped into CIELAB Luo (2023)). (b) The same at a local zoom. (c) The static country embeddings (ICA to three dimensions, then CIELAB). Full task embeddings (1024 tasks) - tsne_cosine_p30 modality country gridded C culture/society C development/inequality C economy/finance C environment/energy C h… view at source ↗
Figure 5
Figure 5. Figure 5: (a) A t-SNE (van der Maaten and Hinton, 2008) of the 1,024 task embeddings. (b) A UMAP of the year embeddings, 1950 to 2020. 4.4 Adaptation to unseen tasks The task encoder reserves rows for variables absent from pretraining. For a new target τ ⋆ , the native adaptation operator jointly optimises a fresh task row and rank-4 MiSS residuals (Kang and Yin, 2026) in the task–spatiotemporal fusion: the input fa… view at source ↗
Figure 6
Figure 6. Figure 6: Cosine distance between each place’s spatiotemporal representation in 2020 and in [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: An emergent development axis. The leading principal component of the country representation against the real Human Development Index, which enters neither decomposition. (a) From the decoded indicator panel. (b) From the raw embedding vectors. decomposition. The sign of a principal component is arbitrary, and the figure orients the axis so that it increases with development. Three indicators that remain in… view at source ↗
Figure 8
Figure 8. Figure 8: Embedding arithmetic across the model spaces: linear semantic axes, a composed [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: A crossed pair of leave-one-out models, identical except for which alignment [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Per-domain maps of predictive uncertainty, each standardised within its own [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison with existing geospatial embeddings, on unseen data. a,b, Two targets at 8,192 labels: truth, training sample, TerraNova’s reconstruction and the strongest baseline. c, Mean rank against label budget, where TN-adapt is the decoder read-out and TN-probe the frozen-embedding read-out of the same model. d, Mean held-out R2 , and e, the same split into land and ocean. f, Per-target results against … view at source ↗
Figure 12
Figure 12. Figure 12: Reconstructing and nowcasting. Filled circles are training years, open circles held-out years inside the window and open squares the held-out blocks at either end; the line is the predicted mean and the band the conformal 90% interval. In this experiment, we aim to assess the temporal capabilities of TerraNova. Each national record (unseen during training) is split by holding out its earliest and latest y… view at source ↗
Figure 13
Figure 13. Figure 13: The fit of Fig [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: a, A dense 0.25◦ field reconstructed from a sparse lattice of measured cells (measured | truth | reconstruction | predictive width σ). b, Held-out R2 (circles, left axis) and PSNR (triangles, right axis) against measurement sparsity, for the best adaptation read-out and the best classical interpolator. c, The predictive width as a fraction of the target’s standard deviation. meat supply per capita a groun… view at source ↗
Figure 15
Figure 15. Figure 15: Few nations in, all nations out. Two indicators reconstructed from 32 reporting countries: truth, the revealed countries, the predicted national level and the predictive width. a, Meat supply per capita. b, Tourist arrivals. Maps show one seed; the text quotes three-seed means. it fills in, widening from 0.16–0.45 of the target’s standard deviation as the lattice thins (Fig. 14c), and it stays calibrated … view at source ↗
Figure 16
Figure 16. Figure 16: From national panels to gridded fields. a, Terrestrial GPP downscaled from national values: the national input, TerraNova’s gridded reconstruction and the gridded truth. b, National panels against gridded point labels at matched budgets. c, Within-country correlation per country; grey marks countries with no sub-national variance. reveal the value for K of the rest, fit a fresh task row with rank-4 MiSS r… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

79 extracted references · 19 canonical work pages

  1. [1]

    Robinson

    Daron Acemoglu, Simon Johnson, and James A. Robinson. The colonial origins of comparative development: An empirical investigation. American Economic Review, 91 0 (5): 0 1369--1401, 2001. doi:10.1257/aer.91.5.1369

    work page doi:10.1257/aer.91.5.1369 2001

  2. [2]

    General geospatial inference with a population dynamics foundation model

    Mohit Agarwal, Mimi Sun, Chaitanya Kamath, Arbaaz Muslim, Prithul Sarker, Joydeep Paul, Hector Yee, Marcin Sieniek, Kim Jablonski, et al. General geospatial inference with a population dynamics foundation model. arXiv preprint arXiv:2411.07207, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Bruinsma, Tom R

    Anna Allen, Stratis Markou, Will Tebbutt, James Requeima, Wessel P. Bruinsma, Tom R. Andersson, Michael Herzog, Nicholas D. Lane, Matthew Chantry, J. Scott Hosking, and Richard E. Turner. End-to-end data-driven weather prediction. Nature, 641 0 (8065): 0 1172--1179, 2025. doi:10.1038/s41586-025-08897-0

    doi:10.1038/s41586-025-08897-0 2025

  4. [4]

    Word embeddings: A survey

    Felipe Almeida and Geraldo Xex \'e o. Word embeddings: A survey. arXiv preprint arXiv:1901.09069, 2019

    Pith/arXiv arXiv 1901

  5. [5]

    Deep evidential regression

    Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, pages 14927--14937, 2020

    2020

  6. [6]

    Policies, projections, and the social cost of carbon: Results from the DICE -2023 model

    Lint Barrage and William D Nordhaus. Policies, projections, and the social cost of carbon: Results from the DICE -2023 model. Proceedings of the National Academy of Sciences, 121 0 (13): 0 e2312030121, 2024. doi:10.1073/pnas.2312030121

    work page doi:10.1073/pnas.2312030121 2023

  7. [7]

    Accurate medium-range global weather forecasting with 3D neural networks

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3D neural networks. Nature, 619 0 (7970): 0 533--538, 2023. doi:10.1038/s41586-023-06185-3

    doi:10.1038/s41586-023-06185-3 2023

  8. [8]

    Bruinsma, Ana Lucic, Megan Stanley, Anna Allen, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan A

    Cristian Bodnar, Wessel P. Bruinsma, Ana Lucic, Megan Stanley, Anna Allen, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan A. Weyn, Haiyu Dong, et al. A foundation model for the E arth system. Nature, 641 0 (8065): 0 1180--1187, 2025. doi:10.1038/s41586-025-09005-y

    doi:10.1038/s41586-025-09005-y 2025

  9. [9]

    Brown, Michal R

    Christopher F. Brown, Michal R. Kazmierski, Valerie J. Pasquarella, William J. Rucklidge, Masha Samsikova, Chenhui Zhang, Evan Shelhamer, Estefania Lahera, Olivia Wiles, Simon Ilyushchenko, Noel Gorelick, Lihui Lydia Zhang, Sophia Alj, Emily Schechter, Sean Askay, Oliver Guinan, Rebecca Moore, Alexis Boukouvalas, and Pushmeet Kohli. AlphaEarth Foundations...

    Pith/arXiv arXiv 2025

  10. [10]

    Global non-linear effect of temperature on economic production

    Marshall Burke, Solomon M Hsiang, and Edward Miguel. Global non-linear effect of temperature on economic production. Nature, 527 0 (7577): 0 235--239, 2015. doi:10.1038/nature15725

    doi:10.1038/nature15725 2015

  11. [11]

    Global sensitivity analysis of integrated assessment models with multivariate outputs

    Leonardo Chiani, Emanuele Borgonovo, Elmar Plischke, and Massimo Tavoni. Global sensitivity analysis of integrated assessment models with multivariate outputs. Risk Analysis, 45 0 (8): 0 2132--2156, 2025. doi:10.1111/risa.70002

    work page doi:10.1111/risa.70002 2025

  12. [12]

    Clay Foundation Model : An open source AI model for E arth, 2024

    Clay Foundation . Clay Foundation Model : An open source AI model for E arth, 2024. URL https://github.com/Clay-foundation/model. Version 1.5

    2024

  13. [13]

    mlco2/codecarbon: v2.4.1, 2024

    Beno \^i t Courty, Victor Schmidt, Sasha Luccioni, Goyal-Kamal , Marion Coord, Boris Feld, et al. mlco2/codecarbon: v2.4.1, 2024. URL https://doi.org/10.5281/zenodo.11171501

    doi:10.5281/zenodo.11171501 2024

  14. [14]

    Atlas of AI : Power, Politics, and the Planetary Costs of Artificial Intelligence

    Kate Crawford. Atlas of AI : Power, Politics, and the Planetary Costs of Artificial Intelligence . Yale University Press, New Haven, CT, 2021. ISBN 978-0-300-20957-0. doi:10.12987/9780300252392

    doi:10.12987/9780300252392 2021

  15. [15]

    Geology of mankind

    Paul J Crutzen. Geology of mankind. Nature, 415 0 (6867): 0 23, 2002. doi:10.1038/415023a

    work page doi:10.1038/415023a 2002

  16. [16]

    RANGE : Retrieval augmented neural fields for multi-resolution geo-embeddings

    Aayush Dhakal, Srikumar Sastry, Subash Khanal, Adeel Ahmad, Eric Xing, and Nathan Jacobs. RANGE : Retrieval augmented neural fields for multi-resolution geo-embeddings. In Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition ( CVPR ) , pages 24680--24689, 2025. doi:10.1109/CVPR52734.2025.02298

    arXiv 2025

  17. [17]

    Global warming has increased global economic inequality

    Noah S Diffenbaugh and Marshall Burke. Global warming has increased global economic inequality. Proceedings of the National Academy of Sciences, 116 0 (20): 0 9808--9813, 2019. doi:10.1073/pnas.1816020116

    doi:10.1073/pnas.1816020116 2019

  18. [18]

    Wendy Nelson Espeland and Mitchell L. Stevens. Commensuration as a social process. Annual Review of Sociology, 24: 0 313--343, 1998. doi:10.1146/annurev.soc.24.1.313

    work page doi:10.1146/annurev.soc.24.1.313 1998

  19. [19]

    Meehl, Catherine A

    Veronika Eyring, Sandrine Bony, Gerald A. Meehl, Catherine A. Senior, Bjorn Stevens, Ronald J. Stouffer, and Karl E. Taylor. Overview of the C oupled M odel I ntercomparison P roject P hase 6 ( CMIP6 ) experimental design and organization. Geoscientific Model Development, 9 0 (5): 0 1937--1958, 2016. doi:10.5194/gmd-9-1937-2016

    doi:10.5194/gmd-9-1937-2016 1937

  20. [20]

    A review of criticisms of integrated assessment models and proposed approaches to address these, through the lens of BECCS

    Ajay Gambhir, Isabela Butnar, Pei-Hao Li, Pete Smith, and Neil Strachan. A review of criticisms of integrated assessment models and proposed approaches to address these, through the lens of BECCS . Energies, 12 0 (9): 0 1747, 2019. doi:10.3390/en12091747

    work page doi:10.3390/en12091747 2019

  21. [21]

    ImageBind : One embedding space to bind them all

    Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. ImageBind : One embedding space to bind them all. In Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition ( CVPR ) , pages 15180--15190, 2023

    2023

  22. [22]

    Dai, and Quoc V

    David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. In International Conference on Learning Representations ( ICLR ) , 2017

    2017

  23. [23]

    The potential to narrow uncertainty in regional climate predictions

    Ed Hawkins and Rowan Sutton. The potential to narrow uncertainty in regional climate predictions. Bulletin of the American Meteorological Society, 90 0 (8): 0 1095--1108, 2009. doi:10.1175/2009BAMS2607.1

    doi:10.1175/2009bams2607.1 2009

  24. [24]

    The potential to narrow uncertainty in projections of regional precipitation change

    Ed Hawkins and Rowan Sutton. The potential to narrow uncertainty in projections of regional precipitation change. Climate Dynamics, 37 0 (1-2): 0 407--418, 2011. doi:10.1007/s00382-010-0810-6

    work page doi:10.1007/s00382-010-0810-6 2011

  25. [25]

    Solomon Hsiang, Robert Kopp, Amir Jina, James Rising, Michael Delgado, Shashank Mohan, D. J. Rasmussen, Robert Muir-Wood, Paul Wilson, Michael Oppenheimer, Kate Larsen, and Trevor Houser. Estimating economic damage from climate change in the United States . Science, 356 0 (6345): 0 1362--1369, 2017. doi:10.1126/science.aal4369

    doi:10.1126/science.aal4369 2017

  26. [26]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA : Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. arXiv:2106.09685

    Pith/arXiv arXiv 2022

  27. [27]

    Position: The P latonic R epresentation H ypothesis

    Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. Position: The P latonic R epresentation H ypothesis. In Proceedings of the 41st International Conference on Machine Learning ( ICML ) , volume 235 of Proceedings of Machine Learning Research, pages 20617--20642. PMLR, 2024

    2024

  28. [28]

    H \'e naff, Matthew M

    Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier J. H \'e naff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, and Jo \ a o Carreira. Perceiver IO : A general architecture for structured inputs and outputs. In International Confe...

    2022

  29. [29]

    TerraMind : Large-scale generative multimodality for E arth observation

    Johannes Jakubik, Felix Yang, Benedikt Blumenstiel, Erik Scheurer, Rocco Sedona, Stefano Maurogiovanni, Jente Bosmans, Nikolaos Dionelis, Valerio Marsocci, Niklas Kopp, et al. TerraMind : Large-scale generative multimodality for E arth observation. In Proceedings of the IEEE / CVF International Conference on Computer Vision ( ICCV ) , pages 7383--7394, 20...

    arXiv 2025

  30. [30]

    MiSS : Revisiting the trade-off in LoRA with an efficient shard-sharing structure

    Jiale Kang and Qingyu Yin. MiSS : Revisiting the trade-off in LoRA with an efficient shard-sharing structure. In International Conference on Learning Representations ( ICLR ) , 2026. URL https://openreview.net/forum?id=gohmWoUSoS. Formerly titled ``Bone: Block Affine Transformation as Parameter Efficient Fine-tuning Methods for Large Language Models''

    2026

  31. [31]

    SatCLIP : Global, general-purpose location embeddings with satellite imagery

    Konstantin Klemmer, Esther Rolf, Caleb Robinson, Lester Mackey, and Marc Ru wurm. SatCLIP : Global, general-purpose location embeddings with satellite imagery. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 39, pages 4347--4355, 2025 a . doi:10.1609/aaai.v39i4.32457

    work page doi:10.1609/aaai.v39i4.32457 2025

  32. [32]

    E arth embeddings: Towards AI -centric representations of our planet

    Konstantin Klemmer, Esther Rolf, Marc Ru wurm, Gustau Camps-Valls, Mikolaj Czerkawski, Stefano Ermon, Alistair Francis, Nathan Jacobs, Hannah Rae Kerner, Lester Mackey, Gengchen Mai, Oisin Mac Aodha, Markus Reichstein, Caleb Robinson, David Rolnick, Evan Shelhamer, Vincent Sitzmann, Devis Tuia, and Xiaoxiang Zhu. E arth embeddings: Towards AI -centric rep...

    doi:10.31223/x5hx9s 2025

  33. [33]

    Kopiczko, Tijmen Blankevoort, and Yuki M

    Dawid J. Kopiczko, Tijmen Blankevoort, and Yuki M. Asano. VeRA : Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. arXiv:2310.11454

    Pith/arXiv arXiv 2024

  34. [34]

    Matti Kummu, Maija Taka, and Joseph H. A. Guillaume. Gridded global datasets for Gross Domestic Product and Human Development Index over 1990--2015. Scientific Data, 5 0 (1): 0 180004, 2018. doi:10.1038/sdata.2018.4

    work page doi:10.1038/sdata.2018.4 1990

  35. [35]

    Learning skillful medium-range global weather forecasting

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, Alexander Merose, Stephan Hoyer, George Holland, Oriol Vinyals, Jacklynn Stott, Alexander Pritzel, Shakir Mohamed, and Peter Battaglia. Learning skillful medium-range global weather forecasting. Scien...

    doi:10.1126/science.adi2336 2023

  36. [36]

    SGDR : Stochastic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. SGDR : Stochastic gradient descent with warm restarts. In International Conference on Learning Representations (ICLR), 2017

    2017

  37. [37]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations ( ICLR ) , 2019. arXiv:1711.05101

    Pith/arXiv arXiv 2019

  38. [38]

    Ming Ronnier Luo. CIELAB . In Encyclopedia of Color Science and Technology, pages 251--257. Springer International Publishing, 2023. doi:10.1007/978-3-030-89862-5_11

    work page doi:10.1007/978-3-030-89862-5_11 2023

  39. [39]

    Mastrandrea, Katharine J

    Michael D. Mastrandrea, Katharine J. Mach, Gian-Kasper Plattner, Ottmar Edenhofer, Thomas F. Stocker, Christopher B. Field, Kristie L. Ebi, and Patrick R. Matschoss. The IPCC AR5 guidance note on consistent treatment of uncertainties: A common approach across the working groups. Climatic Change, 108 0 (4): 0 675--691, 2011. doi:10.1007/s10584-011-0178-6

    doi:10.1007/s10584-011-0178-6 2011

  40. [40]

    UMAP : Uniform manifold approximation and projection for dimension reduction

    Leland McInnes, John Healy, and James Melville. UMAP : Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. doi:10.48550/arXiv.1802.03426

    Pith/arXiv arXiv doi:10.48550/arxiv.1802.03426 2018

  41. [41]

    The unreasonable effectiveness of deep evidential regression

    Nis Meinert, Jakob Gawlikowski, and Alexander Lavin. The unreasonable effectiveness of deep evidential regression. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 9134--9142, 2023. doi:10.1609/aaai.v37i8.26096

    doi:10.1609/aaai.v37i8.26096 2023

  42. [42]

    Linguistic regularities in continuous space word representations

    Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. Linguistic regularities in continuous space word representations. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies ( NAACL-HLT ) , pages 746--751. Association for Computational Linguistics, 2013

    2013

  43. [43]

    Decolonial AI : Decolonial theory as sociotechnical foresight in artificial intelligence

    Shakir Mohamed, Marie-Therese Png, and William Isaac. Decolonial AI : Decolonial theory as sociotechnical foresight in artificial intelligence. Philosophy & Technology, 33 0 (4): 0 659--684, 2020. doi:10.1007/s13347-020-00405-8

    doi:10.1007/s13347-020-00405-8 2020

  44. [44]

    Moore, Moritz A

    Frances C. Moore, Moritz A. Drupp, James Rising, Simon Dietz, Ivan Rudik, and Gernot Wagner. Synthesis of evidence yields high social cost of carbon due to structural model variation and uncertainties. Proceedings of the National Academy of Sciences, 121 0 (52): 0 e2410733121, December 2024. doi:10.1073/pnas.2410733121. URL https://www.pnas.org/doi/10.107...

    work page doi:10.1073/pnas.2410733121 2024

  45. [45]

    Instant neural graphics primitives with a multiresolution hash encoding

    Thomas M \"u ller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics ( TOG ) , 41 0 (4): 0 102:1--102:15, 2022. doi:10.1145/3528223.3530127

    arXiv 2022

  46. [46]

    ClimaX : A foundation model for weather and climate

    Tung Nguyen, Johannes Brandstetter, Ashish Kapoor, Jayesh K Gupta, and Aditya Grover. ClimaX : A foundation model for weather and climate. In International Conference on Machine Learning, pages 25904--25938. PMLR, 2023

    2023

  47. [47]

    Revisiting the social cost of carbon

    William D Nordhaus. Revisiting the social cost of carbon. Proceedings of the National Academy of Sciences, 114 0 (7): 0 1518--1523, 2017. doi:10.1073/pnas.1609244114

    doi:10.1073/pnas.1609244114 2017

  48. [48]

    A new scenario framework for climate change research: the concept of shared socioeconomic pathways

    Brian C O'Neill, Elmar Kriegler, Keywan Riahi, Kristie L Ebi, St \'e phane Hallegatte, Timothy R Carter, Ritu Mathur, and Detlef P van Vuuren. A new scenario framework for climate change research: the concept of shared socioeconomic pathways. Climatic Change, 122 0 (3): 0 387--400, 2014. doi:10.1007/s10584-013-0905-2

    work page doi:10.1007/s10584-013-0905-2 2014

  49. [49]

    Artificial intelligence to support cross-disciplinary climate change research

    Yang Ou et al. Artificial intelligence to support cross-disciplinary climate change research. Nature Climate Change, 16 0 (5): 0 505--507, 2026. doi:10.1038/s41558-026-02624-x

    work page doi:10.1038/s41558-026-02624-x 2026

  50. [50]

    FourCastNet : A global data-driven high-resolution weather model using adaptive F ourier neural operators

    Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, Pedram Hassanzadeh, Karthik Kashinath, and Animashree Anandkumar. FourCastNet : A global data-driven high-resolution weather model using adaptive F ourier neural operators. arXiv prepri...

    Pith/arXiv arXiv 2022

  51. [51]

    Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam, and Matthew Willson

    Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R. Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam, and Matthew Willson. Probabilistic weather forecasting with machine learning. Nature, 637 0 (8044): 0 84--90, 2025. doi:10.1038/s41586-024-08252-9

    doi:10.1038/s41586-024-08252-9 2025

  52. [52]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning ( ICML ) , volume 139 of Pro...

    2021

  53. [53]

    Measuring the intrinsic dimension of E arth representations

    Arjun Rao, Marc Ru wurm, Konstantin Klemmer, and Esther Rolf. Measuring the intrinsic dimension of E arth representations. In International Conference on Learning Representations ( ICLR ) , 2026. URL https://openreview.net/forum?id=gQPD83DrGp

    2026

  54. [54]

    The S hared S ocioeconomic P athways and their energy, land use, and greenhouse gas emissions implications: An overview

    Keywan Riahi, Detlef P van Vuuren, Elmar Kriegler, Jae Edmonds, Brian C O'Neill, Shinichiro Fujimori, Nico Bauer, Katherine Calvin, Rob Dellink, Oliver Fricko, et al. The S hared S ocioeconomic P athways and their energy, land use, and greenhouse gas emissions implications: An overview. Global Environmental Change, 42: 0 153--168, 2017. doi:10.1016/j.gloe...

    doi:10.1016/j.gloenvcha.2016.05.009 2017

  55. [55]

    Earth beyond six of nine planetary boundaries

    Katherine Richardson, Will Steffen, Wolfgang Lucht, J rgen Bendtsen, Sarah E Cornell, Jonathan F Donges, Markus Dr \"u ke, Ingo Fetzer, Govindasamy Bala, Werner von Bloh, et al. Earth beyond six of nine planetary boundaries. Science Advances, 9 0 (37): 0 eadh2458, 2023. doi:10.1126/sciadv.adh2458

    doi:10.1126/sciadv.adh2458 2023

  56. [56]

    A harmonised dataset for E arth system foundation models

    Carlos Rodriguez-Pardo and Massimo Tavoni. A harmonised dataset for E arth system foundation models. Scientific Data, 2026. ISSN 2052-4463. doi:10.1038/s41597-026-07913-w. URL https://doi.org/10.1038/s41597-026-07913-w

    work page doi:10.1038/s41597-026-07913-w 2026

  57. [57]

    NeuBTF : Neural fields for BTF encoding and transfer

    Carlos Rodriguez-Pardo, Konstantinos Kazatzis, Jorge Lopez-Moreno, and Elena Garces. NeuBTF : Neural fields for BTF encoding and transfer. Computers & Graphics, 114: 0 239--246, 2023. doi:10.1016/j.cag.2023.06.018

    work page doi:10.1016/j.cag.2023.06.018 2023

  58. [58]

    Neural conditional transport maps

    Carlos Rodriguez-Pardo, Leonardo Chiani, Emanuele Borgonovo, and Massimo Tavoni. Neural conditional transport maps. Transactions on Machine Learning Research, 2026. arXiv:2505.15808

    arXiv 2026

  59. [59]

    Geographic location encoding with spherical harmonics and sinusoidal representation networks

    Marc Ru wurm, Konstantin Klemmer, Esther Rolf, Robin Zbinden, and Devis Tuia. Geographic location encoding with spherical harmonics and sinusoidal representation networks. In The Twelfth International Conference on Learning Representations, 2024

    2024

  60. [60]

    Baraniuk

    Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G. Baraniuk. WIRE : Wavelet implicit neural representations. In Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition ( CVPR ) , pages 18507--18516, 2023

    2023

  61. [61]

    James C. Scott. Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed. Yale University Press, New Haven, CT, 1998. ISBN 978-0-300-07016-3

    1998

  62. [62]

    Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems ( NeurIPS ) , volume 33, pages 7462--7473, 2020

    2020

  63. [63]

    Participation is not a design fix for machine learning

    Mona Sloane, Emanuel Moss, Olaitan Awomolo, and Laura Forlano. Participation is not a design fix for machine learning. In Proceedings of the 2nd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO '22), pages 1--6. ACM, 2022. doi:10.1145/3551624.3555285

    arXiv 2022

  64. [64]

    Cornell, Ingo Fetzer, Elena M

    Will Steffen, Katherine Richardson, Johan Rockstr \"o m, Sarah E. Cornell, Ingo Fetzer, Elena M. Bennett, Reinette Biggs, Stephen R. Carpenter, Wim de Vries, Cynthia A. de Wit, et al. Planetary boundaries: Guiding human development on a changing planet. Science, 347 0 (6223): 0 1259855, 2015. doi:10.1126/science.1259855

    doi:10.1126/science.1259855 2015

  65. [65]

    Integrated Assessment of Global Environmental Change with IMAGE 3.0: Model description and policy applications

    Elke Stehfest, Detlef van Vuuren, Tom Kram, Lex Bouwman, Rob Alkemade, Michel Bakkenes, Hester Biemans, Arno Bouwman, Michel den Elzen, Jan Janse, et al. Integrated Assessment of Global Environmental Change with IMAGE 3.0: Model description and policy applications. Netherlands Environmental Assessment Agency, 2014

    2014

  66. [66]

    James Purser

    Richard Swinbank and R. James Purser. Fibonacci grids: A novel approach to global modelling. Quarterly Journal of the Royal Meteorological Society, 132 0 (619): 0 1769--1793, 2006. doi:10.1256/qj.05.227

    doi:10.1256/qj.05.227 2006

  67. [67]

    Prithvi- EO -2.0: A versatile multitemporal foundation model for E arth observation applications

    Daniela Szwarcman, Sujit Roy, Paolo Fraccaro, orsteinn El \' G \' slason, Benedikt Blumenstiel, Rinki Ghosal, Pedro Henrique de Oliveira, Joao Lucas de Sousa Almeida, et al. Prithvi- EO -2.0: A versatile multitemporal foundation model for E arth observation applications. IEEE Transactions on Geoscience and Remote Sensing , 64: 0 1--20, 2026. doi:10.1109/T...

    arXiv 2026

  68. [68]

    Fourier features let networks learn high frequency functions in low dimensional domains

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33: 0 7537--7547, 2020

    2020

  69. [69]

    Uncertainty in Integrated Assessment Modeling of Climate Change

    Massimo Tavoni and Giovanni Valente. Uncertainty in Integrated Assessment Modeling of Climate Change . Perspectives on Science, 30 0 (2): 0 321--351, 2022. doi:10.1162/posc_a_00417. URL https://direct.mit.edu/posc/article-abstract/30/2/321/109169/Uncertainty-in-Integrated-Assessment-Modeling-of?redirectedFrom=fulltext

    work page doi:10.1162/posc_a_00417 2022

  70. [70]

    Past and projected climate impacts on human development

    Massimo Tavoni, Marta Mastropietro, and Jonathan Spinoni. Past and projected climate impacts on human development. Research Square preprint, July 2025. Preprint, not peer reviewed. doi:10.21203/rs.3.rs-6923904/v1

    work page doi:10.21203/rs.3.rs-6923904/v1 2025

  71. [71]

    o m, S \

    Jan Teorell, Aksel Sundstr \"o m, S \"o ren Holmberg, Bo Rothstein, Natalia Alvarado Pachon, and Cem Mert Dalli. The Quality of Government Standard Dataset , version Jan21 . University of Gothenburg: The Quality of Government Institute, 2021. doi:10.18157/qogstdjan21

    work page doi:10.18157/qogstdjan21 2021

  72. [72]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  73. [73]

    Visualizing data using t-SNE

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE . Journal of Machine Learning Research, 9 0 (86): 0 2579--2605, 2008

    2008

  74. [74]

    Gomez, ukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems ( NeurIPS ) , volume 30, pages 5998--6008, 2017

    2017

  75. [75]

    GeoCLIP : CLIP -inspired alignment between locations and images for effective worldwide geo-localization

    Vicente Vivanco Cepeda, Gaurav Kumar Nayak, and Mubarak Shah. GeoCLIP : CLIP -inspired alignment between locations and images for effective worldwide geo-localization. In Advances in Neural Information Processing Systems ( NeurIPS ) , volume 36, pages 8690--8701, 2023. doi:10.52202/075280-0379

    doi:10.52202/075280-0379 2023

  76. [76]

    Stewart, Thomas Dujardin, Nikolaos Ioannis Bountos, Angelos Zavras, Franziska Gerken, Ioannis Papoutsis, Laura Leal-Taix \'e , and Xiao Xiang Zhu

    Yi Wang, Zhitong Xiong, Chenying Liu, Adam J. Stewart, Thomas Dujardin, Nikolaos Ioannis Bountos, Angelos Zavras, Franziska Gerken, Ioannis Papoutsis, Laura Leal-Taix \'e , and Xiao Xiang Zhu. Towards a unified C opernicus foundation model for E arth vision. In Proceedings of the IEEE / CVF International Conference on Computer Vision ( ICCV ) , pages 9888...

    2025

  77. [77]

    Some Contributions of Integrated Assessment Models of Global Climate Change

    John Weyant. Some Contributions of Integrated Assessment Models of Global Climate Change . Review of Environmental Economics and Policy, 11 0 (1): 0 115--137, January 2017. doi:10.1093/reep/rew018

    work page doi:10.1093/reep/rew018 2017

  78. [78]

    Do artifacts have politics? Daedalus, 109 0 (1): 0 121--136, 1980

    Langdon Winner. Do artifacts have politics? Daedalus, 109 0 (1): 0 121--136, 1980. Modern Technology: Problem or Opportunity?

    1980

  79. [79]

    Attentional multi-graph convolutional network for regional economy prediction with open migration data

    Fengli Xu, Yong Li, and Shusheng Xu. Attentional multi-graph convolutional network for regional economy prediction with open migration data. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2225--2233. ACM, 2020. doi:10.1145/3394486.3403273

    arXiv 2020