›
foundation modelsfor the earth's crust.
› click anywhere on the crust · drill
lat 40.712800° / lon -74.005200° / depth 3214.7 m
training · tacc/horizon
seismic record · live · channels 01—06sample rate 1 khz · gain 12 db
§ 01thesislat 19.39° · lon -138.30° · depth 0317m┄ ┄ ┄ ┄
§ 01
thesis
beneath the noise of every measurement lies a lower-dimensional manifold of latent geological processes. we are training the model that finds it.
natural hydrogen — a zero-carbon fuel produced continuously by the serpentinization of ultramafic rocks — may be a multi-trillion-ton resource. it replenishes. it costs roughly $1/kg.
finding it requires reading the crust. petabytes of borehole and seismic data already exist, scattered across incompatible silos, 5–10% complete. classical methods cannot interpolate. foundation models can.
§ 02the modellat 26.78° · lon -96.60° · depth 0534m┄ ┄ ┄ ┄
§ 02
the model
jennifer-h2 — a multi-modal foundation model for the subsurface.
joint embedding neural network interpolation for earth resources · h2
inputs
gravity · magnetics · crustal thickness · bathymetry · 2D/3D seismic · borehole logs
core
masked joint-embedding predictive transformer. self-supervised, cross-modal, trained on a unified corpus of real-world heterogeneous data.
output
full posterior over subsurface properties. probabilistic zero-shot inversion in minutes, not weeks.
§ 03evidencelat 34.17° · lon -54.90° · depth 0751m┄ ┄ ┄ ┄
§ 03
evidence
validation — 2 PB ocean borehole corpus, neurips 2026.
corpus coverage · equirectangular10 active basins · 25.0 PB
0+ PB
subsurface corpus
200,000+ boreholes · all major basins · unified
0.00
R² · data reconstruction
vs. lasso baseline 0.24 — 3× better
0.00
AUC · lithology classification
12 classes · cross-basin
ζ
zero-shot generalization
generalizes to unseen cores without retraining
§ 04teamlat 41.56° · lon -13.20° · depth 0968m┄ ┄ ┄ ┄