Pangaea: Geology, Data, AI, and Benchmarking

• Pangaea is a multifaceted term defining geological reconstructions, data repositories, model benchmarking, and AI supercontinent models, each with unique applications.
• Geological studies using Pangaea address continental misplacements by applying Earth radius corrections to paleomagnetic data, highlighting shifts in paleolatitude.
• The PANGAEA initiative advances Earth science by standardizing metadata, enhancing geospatial model evaluations, and integrating multi-agent systems for robust data analysis.

In recent technical literature, Pangaea or PANGAEA denotes several distinct objects: the Paleozoic landmass invoked in paleolatitude-based continental reconstructions; PANGAEA, a long-standing international data publisher for Earth and environmental sciences; PANGAEA, a standardized benchmark family for evaluating geospatial foundation models; and Pangaea, an “AI supercontinent” proposed to dissolve modality-specific “Intelligence Islands” into one unified model [(Herndon, 2011); (Pantiukhin et al., 7 Mar 2025); (Marsocci et al., 2024); (Chang et al., 22 Sep 2025)]. Across these usages, the term is associated with unification, but the underlying epistemic roles differ: geodynamic reconstruction, scientific data infrastructure, model benchmarking, and multimodal representation learning.

1. Geological reconstruction and paleolatitude

In geodynamic usage, Pangaea appears in the context of Pangaea-like reconstructions and paleoclimate interpretation. A 2011 paper argues that classical magnetic paleolatitude determinations may contain a systematic bias if rock magnetization was acquired when Earth’s radius was smaller than at present (Herndon, 2011). The classical dipole relation is written as

$\tan I = 2\tan \phi_{\mathrm{inf}},$

while the paper assumes that, at the time of acquisition on a smaller-radius globe $R_1$, the field was still essentially an axial dipole, so

$\tan I = 2\tan \phi_1.$

If the same physical arc distance $s$ is mapped from $R_1$ to the present Earth radius $R_0$, the modern equivalent latitude is

$\phi_0 = (R_1/R_0)\phi_1 = (1-\epsilon)\phi_1,$

with $\epsilon=(R_0-R_1)/R_0$. The resulting latitude shift is

$$\Delta \phi = \phi_{\mathrm{inf}}-\phi_0 = \epsilon \phi_1 = \epsilon\cdot \arctan[(\tan I)/2].$$

The same paper gives quantitative examples for two compressions, $\epsilon=10\%$ and $R_1$0. For $R_1$1, the inferred paleolatitude is $R_1$2, producing $R_1$3 and $R_1$4 in the two cases. For $R_1$5, the inferred paleolatitude is $R_1$6, producing $R_1$7 and $R_1$8. For $R_1$9, the inferred paleolatitude is $\tan I = 2\tan \phi_1.$0, producing $\tan I = 2\tan \phi_1.$1 and $\tan I = 2\tan \phi_1.$2 (Herndon, 2011).

The claimed implication is that continental blocks in a typical Pangaea-style fit would be placed too close to the pole by roughly $\tan I = 2\tan \phi_1.$3, compressing the reconstruction along the north–south direction and potentially shifting climatic belts. The same source concludes that classical paleolatitudes, and all derivative supercontinent fits and paleoclimate maps, should be treated with an additional systematic uncertainty of order $\tan I = 2\tan \phi_1.$4 if Earth’s radius history is not independently constrained (Herndon, 2011). This is presented as a potential source of error rather than as a universally adopted correction.

2. PANGAEA as an Earth and environmental science data publisher

In data-infrastructure usage, PANGAEA is a long-standing international data publisher for Earth and environmental sciences and one of the world’s largest open-access repositories in the field (Pantiukhin et al., 24 Feb 2026, Pantiukhin et al., 7 Mar 2025). As of mid-2023, it hosts over 400 000 discrete “datasets”, contributed by a global community of oceanographers, geologists, biologists, climatologists, and others, and spans more than 800 distinct geoscientific parameters (Pantiukhin et al., 7 Mar 2025, Pantiukhin et al., 24 Feb 2026). The holdings include tabular CSV or ASCII tables, multi-dimensional gridded arrays such as NetCDF and Zarr, unstructured image and point-cloud files, and ancillary metadata including campaign IDs, coordinate reference systems, and quality-control flags (Pantiukhin et al., 24 Feb 2026).

Metadata conventions are described as a loose, partly controlled-vocabulary model. Contributors supply minimal mandatory fields such as title, authors, date, geographic bounding box, temporal extent, and sampling method, together with variable-level fields, free-text abstracts, DOI registrations, and bibliographic citations (Pantiukhin et al., 7 Mar 2025). In practice, however, metadata are often non-uniform, incomplete, unprocessed, and heterogeneous in format. Examples given in the literature include different names for sea-surface temperature, missing units or depth references, raw logs without quality-control flags, and metadata embedded either in file headers or in separate XML or text files (Pantiukhin et al., 7 Mar 2025).

The main discovery and reuse obstacles are structural. Only ~15% of records have structured abstracts, creating a “long tail” of datasets that never get cited; heterogeneous formats require distinct toolchains and domain expertise; and ambiguous natural-language queries involving seasonality, geographic bounding boxes, or product-level intent are poorly handled by traditional keyword search or single-pass LLM wrappers (Pantiukhin et al., 24 Feb 2026). A plausible implication is that PANGAEA’s scientific value depends not only on archival scale but also on workflow systems capable of converting heterogeneous metadata and files into executable analytical pipelines.

3. Multi-agent systems built over PANGAEA

The repository’s scale and heterogeneity have made PANGAEA a testbed for multi-agent LLM systems. In one line of work, PANGAEA GPT is described as a centralized, hierarchical Multi-Agent System built around a Supervisor Agent that orchestrates retrieval, harmonization, analysis, validation, and reporting sub-agents over PANGAEA’s APIs, vector databases such as FAISS, and software sandboxes (Pantiukhin et al., 7 Mar 2025). The retrieval stage embeds a query with a domain-specialized encoder $\tan I = 2\tan \phi_1.$5 and ranks metadata by cosine similarity; the transformation stage parses NetCDF via xarray, CSV via pandas, and shapefiles via GDAL; the analysis stage computes quantities such as

$\tan I = 2\tan \phi_1.$6

and the validation stage uses checks such as

$\tan I = 2\tan \phi_1.$7

to flag outliers (Pantiukhin et al., 7 Mar 2025).

A later implementation, PANGAEA-GPT, formalizes a Supervisor–Worker topology with five specialists: Oceanographer, Ecologist, DataFrame, Visualization, and Writer (Pantiukhin et al., 24 Feb 2026). Routing is explicitly data-type-aware,

$\tan I = 2\tan \phi_1.$8

where the dataset feature vector determines the assigned agent. Code is executed in a persistent Jupyter kernel isolated by a UUID’d filesystem sandbox, and self-correction is expressed as

$\tan I = 2\tan \phi_1.$9

where $s$0 is the captured traceback from execution (Pantiukhin et al., 24 Feb 2026). Generated figures enter a reflexive visual quality-control loop in which a vision-enabled LLM critiques the image against a 10-point rubric and requests regeneration until a threshold is met or a retry cap is reached (Pantiukhin et al., 24 Feb 2026).

Quantitative evaluations in the two papers are explicit. On a benchmark of 1 000 diverse queries, PANGAEA GPT improved metadata interpretation accuracy from 68% to 91%, increased retrieval precision@10 from 0.62 to 0.84, reduced average end-to-end latency from 12 s to 6.5 s per complex query, and increased overall throughput by 2.3× in parallel sessions (Pantiukhin et al., 7 Mar 2025). On a separate benchmark of 100 real-world queries across six domains and five metrics, baseline Elasticsearch BM25 scored 2.87/10, a simple one-pass LLM-to-Lucene system scored 5.46/10, and Agentic Search scored 8.14/10, with the largest gains in Precision and Parameter Coverage (Pantiukhin et al., 24 Feb 2026). Across four concrete scenarios, the system encountered five runtime errors and resolved all without human help, while visual QC performed 11 critique-refinement cycles and improved final plot scores to $s$1 (Pantiukhin et al., 24 Feb 2026).

4. PANGAEA as a benchmark for geospatial foundation models

A distinct usage of the name refers to PANGAEA, a benchmark introduced to standardize the evaluation of geospatial foundation models (Marsocci et al., 2024). The original paper expands the acronym as “PAtch-level aNalysis and Global Evaluation of gEospatial foundation mAchines” and motivates it as a response to fragmented, geographically biased, and often too narrow downstream evaluations (Marsocci et al., 2024). A later paper glosses the acronym as “PAthways for iNclusive and Generalizable Evaluation of geospAtiAl foundation models” (Rodriguez et al., 27 May 2026). This suggests that the benchmark name has become associated with a broader evaluation program rather than a single immutable expansion.

The original benchmark comprises eleven downstream datasets covering wildfire, marine pollution, agriculture, flood detection, humanitarian assistance and disaster response, urban land cover, urban land-cover change, crop mapping in South Sudan, urban change, smallholder agriculture, and forest biomass estimation (Marsocci et al., 2024). It spans multiple spatial resolutions (1.5 m–30 m), sensor modalities (optical, multi-spectral, SAR), and temporal configurations (single, bi-, multi-temporal), and it standardizes preprocessing, band adaptation, decoders, data splits, and metrics (Marsocci et al., 2024). Preprocessing includes per-band normalization

$s$2

band matching with zero-padding of missing bands, and either resize or crop to match model input (Marsocci et al., 2024). For dense outputs, the protocol freezes the encoder and attaches a standard UPerNet decoder; for supervised baselines, UNet or ViT + UPerNet is trained from scratch (Marsocci et al., 2024).

PANGAEA also formalizes label-scarcity regimes. The original release uses full train/validation/test splits together with stratified 10% and 50% training subsets (Marsocci et al., 2024). Later work describes a three-way 10%, 50%, and 100% split design for nine segmentation tasks, explicitly using the 10% split to stress data-efficient adaptation and expose decoders that overfit when labels are scarce (Forgaard et al., 22 Jan 2026). Metrics are standard: IoU and mIoU for segmentation and change detection, RMSE for regression, and, in later benchmark variants, OA and F1 for classification (Marsocci et al., 2024, Rodriguez et al., 27 May 2026).

An important benchmark result is negative rather than celebratory: the original study reports that geospatial foundation models do not consistently outperform supervised models, especially on simpler or heavily imbalanced tasks, and that temporal modeling, spectral generalization, and multimodal fusion remain unresolved weaknesses (Marsocci et al., 2024). This finding remains a useful corrective to the assumption that pretraining alone guarantees universal dominance.

5. Subsequent PANGAEA benchmark variants and model results

Subsequent papers use the PANGAEA name for several tightly controlled evaluation slices. One branch emphasizes nine semantic-segmentation challenges including BurnSr, MADOS, PASTIS, Sen1Floods11, FBP, DEN, CTM-SS, SN7, and AI4Farms, with Avg. mIoU as the principal ranking metric (Jakubik et al., 15 Apr 2025). Another evaluates scene classification, semantic segmentation, and continuous-value regression under a frozen-encoder protocol across EuroSAT-MS, six segmentation datasets, and two regression tasks (Rodriguez et al., 27 May 2026). A third treats PANGAEA-bench as a unified suite spanning semantic segmentation, change detection, regression, and multilabel classification over MADOS, HLSBurnScars, SpaceNet7, BioMassters, and BigEarthNet v2 (Bernuzzi et al., 26 Oct 2025).

These variants have become a focal point for model comparison. TerraMind, described as the first any-to-any generative, multimodal foundation model for Earth observation, uses dual-scale early fusion and Thinking-in-Modalities (TiM) (Jakubik et al., 15 Apr 2025). On the nine-task segmentation configuration, TerraMindv1-L reports Avg. mIoU = 59.6 and Avg. Rank = 2.6, while TerraMindv1-B reports 58.4 and 2.9. The same paper states that TerraMindv1-B beats every other foundation model on all nine benchmarks and that the large model gains an additional $s$3 pp overall (Jakubik et al., 15 Apr 2025). TiM is illustrated on Sen1Floods11, where generating an artificial LULC map boosts water-IoU from 82.3 to 84.8, a +2.5 pp gain (Jakubik et al., 15 Apr 2025).

FLORO evaluates on six segmentation benchmarks, EuroSAT-MS classification, and two regression tasks (Rodriguez et al., 27 May 2026). It reports OA = 0.910 and F1 = 0.909 on EuroSAT-MS, an Average mIoU = 65.89 and Average rank = 4.00 across six segmentation datasets, BioMassters performance of $s$4 and RMSE = 45.61 t·ha$s$5 at pixel level and $s$6 and RMSE = 11.13 t·ha$s$7 at chip level, and Espeletia performance of $s$8 and RMSE = 0.263 m at pixel level (Rodriguez et al., 27 May 2026). The same source emphasizes that this was achieved with a pretraining corpus roughly 100× smaller than that of TerraMind-L.

THOR focuses on the 10% split of nine dense segmentation tasks and reports the best aggregate ranking in that regime (Forgaard et al., 22 Jan 2026). In the reported table, THOR-Base achieves Avg Rank = 3.78, compared with 4.00 for TerraMind-B, and is state of the art on Sen1Floods11, CropMap, and SN7 (Forgaard et al., 22 Jan 2026). A key ablation shows that, with a frozen THOR encoder on Sen1Floods11 and only 10% labels, shrinking the patch size from 16×16 to 4×4 px boosts test mIoU from 61.9 to 81.1, a +19.2 mIoU gain without retraining the encoder (Forgaard et al., 22 Jan 2026).

WaveMAE isolates the effect of wavelet-based masked autoencoding and geo-conditioned positional encoding under a uniform pretraining dataset, fMoW-S2, with 882 779 patches (Bernuzzi et al., 26 Oct 2025). On its PANGAEA configuration, WaveMAE-Base reports MADOS mIoU = 64.6, HLSBurn mIoU = 85.60, BioMas RMSE = 106.38, SpNet7 mIoU = 52.30, and BENet mAP = 45.28 (Bernuzzi et al., 26 Oct 2025). The paper claims gains of +26.2% on MADOS, +2.35% on HLSBurnScars, –2.5 mRMSE on BioMassters, and +0.61% mAP on BigEarthNet, while noting that SpaceNet7 remains led by SatMAE++ (Bernuzzi et al., 26 Oct 2025).

Taken together, these studies show both continuity and fragmentation. The continuity lies in shared concerns—standardized heads, frozen encoders, limited-label regimes, and cross-sensor transfer. The fragmentation lies in the fact that “PANGAEA” can refer to different task subsets and protocol variants. A plausible implication is that benchmark comparisons are most informative when the exact PANGAEA configuration is made explicit.

6. Pangaea as an “AI supercontinent”

A further, conceptually separate usage appears in “AI Pangaea: Unifying Intelligence Islands for Adapting Myriad Tasks”, which explicitly borrows the geological metaphor (Chang et al., 22 Sep 2025). The paper defines isolated task-specific models across text, vision, time-series, graphs, tables, and related modalities as “Intelligence Islands”. Their reported drawbacks are lack of universality, poor knowledge accumulation, and under-utilization of data. Pangaea is proposed as an AI supercontinent that can ingest any data type, accumulate “universal knowledge,” and adapt to downstream tasks, including modalities not seen during pre-training (Chang et al., 22 Sep 2025).

The central mechanism is a modality-agnostic weighted-graph representation. Every datum is viewed as

$s$9

where nodes are numerical values and edges capture topological relations, producing a triplet set $R_1$0 (Chang et al., 22 Sep 2025). In practice, a raw object $R_1$1 is split into numeric parts $R_1$2, a local topology descriptor $R_1$3, and a global topology descriptor $R_1$4. Each triplet is mapped to a 512-dimensional triplet token by

$R_1$5

where $R_1$6 is a shared linear layer, $R_1$7 is a learnable embedding table, and $R_1$8 is a rotary positional embedding (Chang et al., 22 Sep 2025). The core model is an 8-layer Transformer with bidirectional attention, hidden dimension 256, intermediate dimension 512, and 8 heads (Chang et al., 22 Sep 2025).

Pre-training uses a parallel reconstruction objective over 296 datasets spanning five modalities, with audio and point-cloud held out until fine-tuning (Chang et al., 22 Sep 2025). The paper enumerates 243 tabular datasets with $R_1$9 M samples, 48 time-series datasets with $R_0$0 M segments, ImageNet with 1.28 M images, English Wikipedia with $R_0$1 M text segments of 512 tokens, and three large graph benchmarks with $R_0$2 M nodes (Chang et al., 22 Sep 2025). Training uses $R_0$3 M parameters, 64 × Ascend D910B NPUs, 55 h, 240 K steps, batch size 4 096, and AdamW with lr = 2e-4, warmup 3%, and cosine restarts (Chang et al., 22 Sep 2025).

The downstream claims are broad. Pangaea is fine-tuned without structural changes on 45 general tasks spanning 7 modalities—including tabular, time-series, vision, audio, graph, point-cloud, and text—and the paper states that all 45 specialized baselines were outperformed by a single Pangaea model (Chang et al., 22 Sep 2025). Reported average improvement is +7.5% absolute for classification/F1/AUC and –MSE by 7.5% for regression (Chang et al., 22 Sep 2025). On 15 scientific tasks, Pangaea reports, among other numbers, 79.0% ACC on prostate cancer grading versus 66.3% for MAE-ViT-B, 94.1% AUC on BBBP versus 93.0% for MTL-BERT, 2.45 RMSE on ERA5 global temperature forecasting versus 3.20 for PatchTST, 161.6 MAE on JARVIS-DFT band-gap prediction versus 253.0 for CartNet, and 27.8% ACC on MMLU versus 26.2% for BERT-Tiny and $R_0$4 for RoBERTa (Chang et al., 22 Sep 2025).

The paper also advances a compact scaling model for cross-modal knowledge accumulation. If each pre-training modality contributes an independent Bernoulli trial with success probability $R_0$5, downstream accuracy follows

$R_0$6

The reported fit is

$R_0$7

equivalently $R_0$8 and $R_0$9, which the paper identifies as the CDF of a geometric distribution (Chang et al., 22 Sep 2025). The same source lists limitations: the model remains a black box; some modality combinations display an affinity phenomenon whose theory remains open; and the geometric CDF is described as an intuitive fit lacking a rigorous proof of universality (Chang et al., 22 Sep 2025). These caveats are integral to the proposal, not incidental to it.