v1

Tectonic reconstruction snapshots v1

Present-day geometry templates and finite rotations for the math.gl tectonic time machine and other vis.gl examples. The files are derived from pinned scientific datasets and are CC-BY-4.0, not MIT. See each model’s attribution and manifest, and the adjacent license text.

Model	Source revision	Files	Total size	Ages before present	Frame
Cao et al. (2024)	2.4	tectonic.parquet	1,590,691 bytes	0–1800 Ma	Paleomagnetic
Müller et al. (2022)	1.2.4	geometry.parquet + rotations.parquet	1,927,854 bytes	0–1000 Ma	Optimised mantle

The three files total 3,518,545 bytes (3.52 MB), with ZSTD level 6 compression. Git LFS stores the Parquet files following this repository’s .gitattributes policy. The manifest records sizes, SHA-256 hashes, source archive/file hashes, schemas, source revisions, licenses and conversion tool versions.

Downloading

The URLs below pin the data snapshot commit ba62eb5e331b46693b88ebf85c4b464ec1149027. Use GitHub’s media endpoint for Parquet:

https://media.githubusercontent.com/media/visgl/deck.gl-data/ba62eb5e331b46693b88ebf85c4b464ec1149027/earth/tectonic-movements/v1/cao2024/tectonic.parquet
https://media.githubusercontent.com/media/visgl/deck.gl-data/ba62eb5e331b46693b88ebf85c4b464ec1149027/earth/tectonic-movements/v1/muller2022/geometry.parquet
https://media.githubusercontent.com/media/visgl/deck.gl-data/ba62eb5e331b46693b88ebf85c4b464ec1149027/earth/tectonic-movements/v1/muller2022/rotations.parquet

Use raw.githubusercontent.com/visgl/deck.gl-data/ba62eb5e331b46693b88ebf85c4b464ec1149027/... for manifests and attribution. Raw URLs for LFS-backed Parquet may return a small LFS pointer instead of the data. With Git, install Git LFS and run git lfs pull with an include filter for earth/tectonic-movements/v1/**/*.parquet.

Data conventions

Geometry records are GeoParquet 1.1.0, containing WKB Polygon geometry in longitude/latitude degrees (implicit OGC:CRS84) with spherical edges. Each row has featureId, polygonIndex, plateId, name, beginAge, endAge and geometry. Source vertices and holes are retained; ring closure is added when absent. Source winding is retained without an orientation assertion. These are unreconstructed present-day templates, not reconstructed paleoshorelines or evolving plate topology.

Rotation records contain age, plateId, available, w, x, y and z. Ages are millions of years before present, sampled every 10 Ma. Quaternions use w,x,y,z order with anchor plate 0, and Cartesian X at lon=0/lat=0, Y at lon=90/lat=0, Z at the North Pole. Keep each model’s geometry, rotations and reference frame paired. A missing rotation has available=false and null quaternion components; it is never replaced with an identity rotation. Filter geometry by its valid age range and handle unavailable rotations explicitly. Interpolate between sampled rotations for playback; no future reconstruction scenarios are included.

Cao’s tectonic.parquet is a tagged table: recordType is geometry or rotation. Both record types share plateId. Geometry fields are null on rotation rows; rotation fields are null on geometry rows. The geometry column is nullable WKB. Templates are stored once, without duplicating them for each age. Its first row group contains all 877 geometry records; the following 18 groups contain rotations. Müller retains separate files, with ten geometry groups and ten rotation groups.

Rotation groups cover 100 Ma windows, sorted youngest first. Windows are half-open, for example 0–90 and 100–190 Ma at 10 Ma sampling. The final window also includes the model’s final endpoint (1700–1800 Ma for Cao, 900–1000 for Müller). All plates for an age stay together. Cao’s two source rotation files are complementary 0–1000 and 1000–1800 Ma time slices, combined into this one snapshot. Müller uses only optimisation/1000_0_rotfile_MantleOpt.rot; its paleomagnetic file covers the same interval in an alternative frame and must not be merged with it.

Each file also embeds math.gl.tectonic.manifest: source and archive checksums, model revision, authors/license, reference frame, coordinate conventions, sampling, polygon/plate counts, conversion tools and verification details. Its file section contains the current file’s schema and row-group inventory. Consumers can recover these from the footer without the sidecar. The external manifest additionally records finished-file sizes and SHA-256 hashes; those cannot be embedded in the same file without changing the bytes they describe. It also exposes the full multi-file inventory before a reader fetches any Parquet footer.

Each file’s footer contains a JSON math.gl.tectonic.rowGroups index with rowGroup, recordType, rows, minAge and maxAge. The same index appears in the manifest, so consumers can choose groups before fetching the footer. Geometry ranges are null. Parquet column statistics also include rotation age min/max and, for Cao, record type. Readers can skip unwanted age windows and decode only needed columns. HTTP range access requires support from both the reader and the host; row groups alone do not make an HTTP client stream ranges automatically.

For example, select Cao’s 500–590 Ma window from the manifest, then decode it in batches without reading names or geometry:

import json
import pyarrow.parquet as pq

manifest = json.load(open('cao2024/manifest.json'))
file = manifest['files'][0]
groups = [entry['rowGroup'] for entry in file['rowGroupIndex']
          if entry['recordType'] == 'rotation'
          and entry['minAge'] <= 550 <= entry['maxAge']]
parquet = pq.ParquetFile('cao2024/tectonic.parquet')
for batch in parquet.iter_batches(row_groups=groups, batch_size=4096,
        columns=['age', 'plateId', 'available', 'w', 'x', 'y', 'z']):
    # Process each Arrow batch; apply an exact-age filter if needed.
    pass

Compression alone is not a browser-startup benchmark; measure transfer, decoding and first usable pose.

Reproducing and validating

The original MIT conversion and validation scripts are in scripts. Use a separate environment. pyGPlates is an external GPL-2.0 offline tool; no GPlates implementation or runtime dependency is included in these assets or the browser application. The converted scientific data retains CC-BY-4.0.

From this directory:

python -m venv /tmp/tectonic-converter
/tmp/tectonic-converter/bin/pip install -r scripts/requirements.txt
curl --fail --location https://zenodo.org/api/records/13628813/files/1.8Ga_model_GSF.zip/content -o /tmp/cao2024-v2.4.zip
curl --fail --location https://zenodo.org/api/records/13636799/files/Muller_etal_2022_SE_v1.2.4.zip/content -o /tmp/muller2022-v1.2.4.zip
/tmp/tectonic-converter/bin/python scripts/create-parquet.py --cao /tmp/cao2024-v2.4.zip --muller /tmp/muller2022-v1.2.4.zip --out /tmp/tectonic-parquet-regenerated
/tmp/tectonic-converter/bin/python scripts/validate-parquet.py --root /tmp/tectonic-parquet-regenerated
/tmp/tectonic-converter/bin/python scripts/validate-parquet.py
/tmp/tectonic-converter/bin/python scripts/validate-source-rotations.py --cao /tmp/cao2024-v2.4.zip --muller /tmp/muller2022-v1.2.4.zip

Validation reads every file and checks checksums, schemas, embedded provenance, actual column compression, GeoParquet metadata, polygon coordinates/rings, age statistics, complete plate sets, unit quaternions, streaming column reads, indexed 100 Ma windows and missing-value conventions. The independent source validator checks every rotation sample against explicitly selected canonical source files without importing the converter configuration. Conversion additionally checks full Arrow/Parquet round trips and reconstructed test points against pyGPlates. Compression level is recorded by the converter; the Parquet footer reports the codec, not its level.