Code, figures and derived data for
Müller, R. D., Fichtner, A., & Schiller, C. J. (2026). Mantle plume conduits are most prominent beneath the mid-mantle viscosity increase. Geochemistry, Geophysics, Geosystems.
Whether global tomography resolves mantle plume conduits was disputed for two decades; the full-waveform models REVEAL and RevealLO now show slow columns beneath many hotspots. We trace least-cost descending paths beneath 49 hotspots and from 300 ambient locations, calibrate them against injected synthetic conduits, and analyse their properties by depth against the ambient population. A connected slow column reaches the base of the mantle beneath 22-27% of hotspots against 4-7% of random sites, a contrast other tomography models render weakly or not at all, but many plume columns are intermittent. Hotspot conduits are more prominent beneath the lower-mantle viscosity increase inferred across 800-1200 km than above it, and in RevealLO, our preferred model, hotspot paths are more inclined than ambient paths at 1000-1500 km and nowhere else. Deflections show no horizon at the viscosity increase, and corridor width grows smoothly with depth as the resolving length. Traced endpoints lie within low-velocity provinces 2.6 times more often than matched ambient endpoints, and 33 of 49 hotspots have a second column within 3% of cost ending a median 960 km away: tomography alone cannot always uniquely determine a plume's source.
A conduit is represented as the cheapest descending path through a tomographic shear-velocity model. A step of physical length L through material whose anomaly is a per cent costs
w = L * exp(s * a)
so slow material is cheap to traverse and lateral wandering is penalised. Paths must descend monotonically - that is what separates a conduit from a tour of the low-velocity network - but are otherwise free to move sideways, so the lateral offset of a conduit is recovered from the data rather than assumed. The search runs in reverse, as a multi-source sweep upward from the deep mantle by min-plus dynamic programming, so one sweep yields the cost from every cell in the model and a hotspot or a null location is a lookup.
Two things make a result out of that. What the search can see is fixed by requiring it to recover synthetic conduits of known amplitude, which puts a stated sensitivity floor on every non-detection. What counts as a detection is fixed by comparing each hotspot with random locations put through the identical procedure. Neither is a free choice.
This repository holds the analysis code and the finished figures. The tomographic models are not redistributed (see Model availability); the derived tables in data/ and the recovered conduit geometries that the figures are drawn from are archived on Zenodo (see Citation).
src/ analysis and figure scripts (Python 3), and the shell drivers below
run_all.sh the whole workflow: trace, calibrate, classify, tabulate
refigure.sh redraw every figure from what an earlier run_all.sh already wrote
path_cost.py the least-cost path search
tomo_io.py model loading, depth cropping, file screening
contrast.py the local-contrast channel
inject.py synthetic conduits, for calibration
classify.py calibrate on a model, then classify on it
paths.py recover conduit geometry and section planes
null_snapshot.py the null distribution behind the classification
compare.py merge the models, agreement, correlations
table.py the ranked table
track_depth.py the depth ladder, and the geometry of each connection
site_profile.py strongest anomaly beneath a hotspot, against depth
province.py membership of the lowermost-mantle slow provinces
longevity_analysis.py / longevity.py hotspot-track longevity
sections_figure.py the annular cross-section atlas (Figures 8-11, S3a-c)
fig*.py the remaining main-text and supplementary figures
manuscript_stats.py every statistic the paper quotes, in one file
figcheck.py / figstyle.py figure-legibility and label-collision checks
data/ small inputs carried with the repository (see Inputs)
figures/ the eleven main-text and five supplementary figures, PDF and PNG
out/ (derived products of every step) and results/ are not tracked; the archived copy on Zenodo lets the figures and tables rebuild from them without the tomography models. Running src/run_all.sh regenerates both from scratch.
None of the tomographic model files used here are redistributed in this repository.
Every model file used in this study was provided by the Seismology and Wave Physics group at ETH Zürich. RevealLO has not been released or formally published. The other four are published models, obtainable from their authors at the references below; the copies used here are ETH-prepared and will not be byte-identical to the authors' own distributions, so a run from an independently obtained copy may differ in detail from the numbers in the paper.
| model | as used here | reference |
|---|---|---|
| RevealLO | 0.5°, 10 km depth, cropped at 2880 km | not yet released (ETH Zürich, Seismology and Wave Physics) |
| REVEAL | 0.5° | Thrastarson et al. (2024), BSSA 114, 1392-1406 |
| GLAD-M35 | 1°, every 2nd shell | Cui et al. (2024), GJI 239, 478-502 |
| SPiRaL | as distributed | Simmons et al. (2021), GJI 227, 1366-1391 |
| SEMUCB-WM1 | as distributed | French & Romanowicz (2014), GJI 199, 1303-1327 |
src/tomo_io.load_anomaly absorbs most differences in file layout on its own (variable and coordinate names matched case-insensitively, depth in metres or kilometres, longitude on -180..180 or 0..360, dimensions in any order). Two things it cannot infer, and that must be set per model in the MODELS array at the top of src/run_all.sh: --var, whether the file carries vsv/vsh to Voigt-average or a single pre-averaged variable, and --depth-max, which has to sit above the core-mantle boundary at 2891 km. The driver skips any model whose file it cannot find, so run_all.sh runs correctly on whatever subset is available.
Carried in data/:
hotspots_courtillot2003.csv: the 49-hotspot list of Courtillot et al. (2003), EPSL 205, 295-308, redistributed here unmodified.oib_hotspot_names.csv,hotspot_basin.csv: hotspot naming and basin-membership lookups used by the classification and figure scripts.bbs2007_figure14_ranking.csv: the hotspot-longevity ranking of Boschi, Becker & Steinberger (2007), used bysrc/longevity_analysis.py.ridges_presentday_zahirovic2022.csvandS2_Shapefiles_Extinct_and_active_ridge_segments/: present-day and extinct/active ridge geometries following Zahirovic et al. (2022) and MacLeod et al. (2017), used by the ridge-proximity tests.
Obtained from their sources when the analysis is rerun from scratch: the five tomographic models above, and the ocean island basalt geochemical database of Hardardottir & Jackson (2025) and the raw Pacific radiometric age table of Chase & Wessel (2021) (https://doi.org/10.5281/zenodo.5576466, CC-BY-4.0), for the two steps that compare against them - each is skipped when its file is absent.
python3 -m pip install -r requirements.txt
pdfplumber, needed only by figcheck.py, wants a Pillow newer than several other packages accept, so it gets its own environment rather than installing in front of whatever else is there:
python3 -m venv .venv-figcheck
.venv-figcheck/bin/pip install -q pdfplumberEdit the MODELS array at the top of src/run_all.sh to point at your copies of the tomographic models, then
cd src
./run_all.sh # everything, skipping models already done
FORCE=1 ./run_all.sh # re-run models even where output exists
./run_all.sh SEMUCB-WM1 # a single modelModels are independent, so the per-model loop splits across machines; the aggregation steps need whatever set you want compared. refigure.sh redraws every figure from what run_all.sh already wrote, without recomputing anything, and finishes by checking them against the width they are placed at in the paper:
cd src && ./refigure.sh # or ./refigure.sh GLADM35 for one model
./checkfigs.sh # the same check on its ownCost. The expensive step is calibration: one cost field per configuration per injected conduit, 72 configurations and 32 conduits, so roughly 1200 sweeps per model. A sweep is a few seconds at 1 degree and about four times that at 0.5 degrees. Budget one to two hours for a half-degree model at ~100 depth shells, four to six for RevealLO at its native 10 km sampling.
Depth decimation is safe here. --every N thins the depth axis, and the answer does not depend on shell spacing, because the search prices a path by physical length and amplitude rather than by counting connected cells - unlike a percolation or connected-component analysis, where the threshold moves by a factor of thirty between REVEAL's 97 shells and RevealLO's 10 km spacing.
python3 src/fig1_map.py # Figure 1: hotspots and traced paths
python3 src/fig_workflow.py # Figure 2: workflow
python3 src/fig_corridor.py # Figure 3: near-optimal corridor width
python3 src/fig_lean.py # Figure 4: directional tests of path lean
python3 src/fig_bands.py # Figure 5: where conduits differ from ambient mantle, by depth
python3 src/fig_deflection.py # Figure 6: distribution of path-deflection depths
python3 src/fig_calibration.py # Figure 7: synthetic calibration of corridor width
python3 src/sections_figure.py --all # Figures 8-11: the 49-hotspot cross-section atlas
python3 src/fig_si_sweeps.py # Figure S1: calibration sweeps
python3 src/fig_si_profiles.py # Figure S2: corridor profiles
python3 src/sections_figure.py --top 9 # Figures S3a-c: best-constrained sections, other models
sections_figure.py needs --file/--tag/--var for the model being drawn; refigure.sh shows the full invocations, including the paper model's --fix-azimuths pass that derives the section planes every other model is then drawn on. figures/figure_map.txt is the authority for which file is which figure number.
Whole-radius depth axes. RevealLO is distributed from the surface to the centre of the Earth, so shells straddling the core-mantle boundary at 2891 km carry velocities where shear collapses and read as anomalies near -97%, which would make a path to the deep mantle almost free. --depth-max crops above the boundary.
Files assembled from several depth ranges. Joins between depth ranges can average across velocities belonging to different depths and produce spurious extreme anomalies. tomo_io.check_shells refuses to run on any model in which a shell's 99.9th-percentile absolute anomaly exceeds four times the median over all shells below 300 km.
Naming and layout conventions. Variables appear as vsv/VSV, depth in metres or kilometres, longitude on -180..180 or 0..360, dimensions in either order, and a repeated wrap meridian at +-180 that would let a path cross the seam for nothing. tomo_io.load_anomaly normalises all of these; a depth axis is taken to be in metres if its maximum exceeds 2e4, since a depth in kilometres cannot exceed the Earth's radius.
Figures carry no titles and no text below 8 pt at the printed width; checkfigs.sh enforces this on the PDFs. sections_figure.py and the other GMT-drawn figures quote every size through a pt() helper as the size it will print at, scaled to the drawing canvas, so changing a panel's pitch cannot quietly shrink the type.
The code in src/ is released under the MIT licence (LICENSE). The figures in figures/ are released under CC BY 4.0. Third-party data in data/ retain the licences of their sources, named above. The tomographic models are the property of their authors and are not redistributed here.
Müller, R. D., Fichtner, A., & Schiller, C. J. (2026). Mantle plume conduits are most prominent beneath the mid-mantle viscosity increase. Geochemistry, Geophysics, Geosystems.
CITATION.cff carries the same in machine-readable form. Please also cite the tomographic models you use, listed above.
Boschi, L., Becker, T. W., & Steinberger, B. (2007). Mantle plumes: Dynamic models and seismic images. Geochemistry, Geophysics, Geosystems, 8, Q10006. https://doi.org/10.1029/2007GC001733 Chase, C. G., & Wessel, P. (2021). PHT2021: Pacific hotspot track ages. Zenodo. https://doi.org/10.5281/zenodo.5576466 Courtillot, V., Davaille, A., Besse, J., & Stock, J. (2003). Three distinct types of hotspots in the Earth's mantle. Earth and Planetary Science Letters, 205, 295-308. https://doi.org/10.1016/S0012-821X(02)01048-8 Cui, C., et al. (2024). GLAD-M35. Geophysical Journal International, 239, 478-502. https://doi.org/10.1093/gji/ggae270 French, S. W., & Romanowicz, B. A. (2014). SEMUCB-WM1. Geophysical Journal International, 199, 1303-1327. https://doi.org/10.1093/gji/ggu334 MacLeod, C. J., et al. (2017). Extinct and active spreading ridge segments. Supplementary dataset. Simmons, N. A., et al. (2021). SPiRaL. Geophysical Journal International, 227, 1366-1391. Thrastarson, S., et al. (2024). REVEAL. Bulletin of the Seismological Society of America, 114, 1392-1406. Zahirovic, S., et al. (2022). Plate model reconstruction.