Circular Economy
Microsoft and Terradot Claim 75% Sampling Cut in Rock Weathering Model
Microsoft and Terradot report a 75% sampling cut from their TerraNova model across 264 fields. The figure is a simulation result, and certified net removal remains a separate, audited process.

Waypoints
Microsoft and Terradot described TerraNova on November 12, 2026, an AI model for estimating enhanced rock weathering.
The working paper reports 5,192 measurements from 264 fields in the US and Brazil, with cross-validation across 64 sites.
Reported sampling reduction is 75% at a 0.75 tCO₂e/ha/yr error target and 63% at 0.5 tonnes.
The paper's files are embargoed on Zenodo until December 12, 2026; findings are described as simulations ahead of peer review.
Isometric protocol v1.3 requires independent validation and verification before any net-removal claim.
A 75% reduction in soil sampling is the headline number Microsoft and Terradot attached to TerraNova, the AI model they described on November 12, 2026. The figure deserves an asterisk, and the companies themselves supplied it: the gains come from preliminary simulations, and they do not establish independently verified carbon removals. Terradot lists the report under the same date.
The model matters for the enhanced rock weathering stream, where measurement cost shapes project economics. If a foundation model can cut the number of physical soil samples a developer must collect and analyse per hectare, the cost of producing a defensible removal estimate falls with it.
What TerraNova actually does
According to Microsoft's announcement, TerraNova adapts Aurora — the company's atmospheric foundation model — to predict chemical changes in soil, using data from fields where Terradot works with farmers. The approach pairs model predictions with physical soil samples rather than replacing measurement outright. Microsoft says further development and extensive validation are needed.
Aurora's underlying research appeared in Nature in May 2025 and describes a foundation model trained on diverse geophysical data, adaptable to environmental forecasting tasks. TerraNova extends that research direction from weather prediction into geochemistry.
The stakes are material. Basalt and similar rock amendments only count as carbon removal if the resulting weathering can be measured, modelled and audited. Sampling is the expensive part.
What the sampling results actually describe
The authors' Zenodo working-paper abstract reports training and evaluation on 5,192 enhanced-rock-weathering measurements drawn from 264 fields in the United States and Brazil, with grouped cross-validation across 64 sites. These are the authors' reported results, not an independent replication.
The headline percentages vary with the error target:
- Across the seven largest sites, a 75% sampling reduction at a target error of 0.75 tonnes of CO₂ equivalent per hectare per year.
- At the tighter 0.5-tonne target, the reported reduction falls to 63%.
The headline number therefore depends on how much error a project will tolerate. It is not a uniform reduction established for every field.
Microsoft describes the findings as simulations and the research as a forthcoming preprint ahead of peer review. The Zenodo record is dated November 12 and exposes only an abstract; the paper's files are embargoed until December 12, 2026.
Predicted weathering is not verified removal
Carbon accounting adds a layer beyond predicting rock reactions. The Isometric Enhanced Weathering in Agriculture protocol, version 1.3, accounts for gross removal alongside lifecycle greenhouse-gas emissions to arrive at a net-removal figure, and its requirements include evidence verified by independent third parties.
Two sections do the structural work. Section 6.1 calls for a project design document describing the relevant geochemical models and a comprehensive sampling plan. Section 6.2 requires independent project validation and verification of the net-removal statement.
Those requirements explain why a reduction in the samples needed for an estimate does not, by itself, demonstrate certified net removals. The TerraNova figures are an accounting benchmark, not evidence that the model has received registry approval.
Why the report matters beyond Terradot
For university researchers, the development raises a concrete question of external validity: how reliably can a model carry information between fields operating under different conditions? For research offices communicating such work, the useful distinction runs between a predictive result, its supporting field measurements, and independently verified net carbon removal.
TerraNova's reported gains make that distinction more consequential as AI models enter environmental measurement research. A model that lowers sampling cost also lowers the barrier to claiming removals, which places more weight on the verification steps that follow.
The milestone to watch is December 12, 2026, when the embargo lifts and the preprint — still ahead of peer review — becomes available for scrutiny.
via casrai.org (Source)
More from Olivia Hart
Show full bio
Staff writer covering regulation, policy and their impact on Circular economy and the environmental industry at Circular Wire.
509 articles