Microsoft and Terradot use Brazilian data to train AI that cuts carbon sampling in half

A climate model became a carbon measurement tool

Enhanced rock weathering involves spreading crushed rock dust over farmland to speed up a chemical reaction that pulls CO₂ out of the atmosphere. The problem is that measuring how much carbon was actually removed requires going back to the field, drilling boreholes, collecting soil samples, and repeating lab analyses over time, because the result varies from one plot to another.

According to Microsoft, measuring and verifying this removal can account for a third to more than half of the technique's total cost. It's a problem of georeferenced data with a high cost of collection, the kind of bottleneck that's usually tackled with more sensing or with predictive models. Microsoft and the startup Terradot (terradot.earth) chose the second route and presented, this Thursday, November 12, TerraNova, a model that combines AI forecasting with physical sampling to estimate removal with fewer collections.

From weather forecasting to soil: how Aurora became TerraNova

TerraNova wasn't built from scratch. It starts from Aurora, Microsoft's Earth system model trained on more than 1 million hours of atmospheric data, originally designed for weather forecasting. The team reused what Aurora had already learned about temperature, humidity, and rainfall, and added soil and mineral data plus field measurements from Terradot to teach the model to predict how the rock behaves after it's applied.

Technically, the key point is the reuse of a latent representation (the summarized output Aurora generates about a location's climate dynamics) as input for a problem completely different from what motivated the original training. For those who work with geospatial models and deal with the cost of training from scratch for each new domain, this is the bet that a climate foundation model generalizes enough to save years of physical calibration.

Essas representações do tempo permitem generalizar para uma diversidade de locais, projetos e condições.

Felipe Oviedo, principal research scientist at Microsoft's AI for Good lab

The training data came from Terradot projects in São Paulo, Paraná, and Mato Grosso do Sul: according to Julia Sekula, Terradot's co-founder and CFO, the Brazilian farms provided thousands of paired soil measurements, the kind of data that makes it possible to track real change on the ground instead of relying only on physical simulation.

Nossa tese era que poderíamos aproveitar o que o Aurora já sabia sobre o tempo para aprender como o intemperismo das rochas acontece no solo.

Julia Sekula, co-founder and chief financial officer of Terradot

In one test, data from Brazilian fields helped the model generate predictions for a project in the Mississippi Delta, in the United States, without specific recalibration for that soil. It's this geographic leap, not the isolated sampling gain, that the team points to as the study's most relevant result.

The numbers behind the promise

The study, accepted at a NeurIPS 2026 conference workshop and co-authored by Sasankh Munukutla (Terradot), Felipe Oviedo, and three other Microsoft researchers, evaluated five different weather models using data from 264 fields in Brazil and the United States. An expanded version, with seven authors, was published on the Zenodo repository this Thursday; the full text will only be available in December.

The engineering reading of these numbers matters more than the "half the samples" headline. The gain shows up in two distinct decisions in a carbon project's workflow: before the rock dust is applied, the model helps prioritize which plots deserve a visit; afterward, it helps estimate removal with less physical collection. These are two different data products within the same pipeline, and each has its own point of risk.

What still hasn't been proven

All the numbers above come from simulations run on existing data, not from a new measurement cycle in production. The study itself acknowledges that, on completely new fields, the model alone still doesn't estimate average removal with sufficient reliability: physical samples remain part of the configuration that performed best.

For those who buy or certify carbon credits, the yardstick that decides whether sampling savings turn into cost savings per ton is the method's acceptance by the independent registries that today require physical collection and lab analysis for every ton reported.

O TerraNova não faz parte da forma como quantificamos nossos créditos hoje, e isso só vai mudar quando as evidências e os registros derem respaldo a esse uso.

Julia Sekula, co-founder and chief financial officer of Terradot

Oviedo argues that the most important result is generalization across distant regions, with different soils and weathering conditions.

Historicamente, acreditava-se que o intemperismo era variável demais para ser previsto com precisão em grandes áreas, e o TerraNova mostra que não é assim.

Felipe Oviedo, principal research scientist at Microsoft's AI for Good lab

Microsoft already has skin in the game on this result: the company's climate innovation fund invested in Terradot, and the company signed a contract in 2025 to buy 12,000 tons of CO₂ removal via weathering, with deliveries between 2026 and 2029. Terradot projects removing up to 100,000 tons with the dust already applied in the field and has a target of 1 million tons per year by 2035.

What changes for teams building environmental data pipelines in Brazil

For data and ML teams working with MRV (measurement, reporting, and verification) for carbon, agriculture, or ESG, the TerraNova case raises three concrete questions to bring to the next vendor or to your own research team:

• Does the model generalize outside the training set, or does it just repeat what it has already seen? TerraNova's leap from Brazilian farms to the Mississippi Delta is the kind of evidence that separates a useful model from one that just memorizes the calibration dataset.
• Who owns the paired data that becomes a competitive advantage? Terradot's "thousands of paired measurements" in SP, PR, and MS function as a proprietary data moat; any competitor wanting to reproduce the result needs the same volume of physical measurements, not just access to an open foundation model.
• Does the projected savings survive a carbon registry audit? A reduction in samples within a simulation is a research result; a reduction in auditable cost per ton, verified by an independent verifier, is a different level of evidence, and that's the bottleneck that still needs to be solved here.

The full study only comes out in December, and until then TerraNova remains a proof of concept, not an input for certification. But the pattern of reusing a foundation model trained for one domain (climate) on an adjacent problem (soil geochemistry) is an approach that any team dealing with geospatial data that's scarce and expensive to collect should watch closely, regardless of whether they're in the carbon market.