Microsoft and Terradot’s TerraNova AI Could Reduce Carbon Removal Sampling, but Field Tests Remain Essential

Why powdered rock needs sophisticated measurement​

Enhanced rock weathering accelerates chemical reactions that naturally consume carbon dioxide. Crushing suitable rock exposes more surface area, allowing reactions to proceed faster when the material is spread over agricultural soil. But outcomes depend on rock composition, soil chemistry and weather, making extensive sampling and laboratory analysis necessary.

That creates two connected problems: deciding where to deploy rock and establishing what happened afterward. A promising field is not automatically a proven carbon-removal project.

Microsoft Source reports that TerraNova combines Terradot’s field data with Microsoft’s Aurora foundation model. In simulations, researchers reduced sampling requirements at the largest sites by 75% while maintaining the same accuracy. That is a reported simulation result, not a demonstrated 75% reduction in operating costs.

What the operational study actually establishes​

A related October 2026 study, From Weather to Enhanced Weathering, provides a more precise picture of the evidence. Researchers evaluated five weather foundation models using 5,192 measurement sequences from 264 agricultural fields in Brazil and the United States. Each sequence paired measurements from different times. Their strongest configuration, TerraNova-ERW, used Aurora 1.5.

The evaluation withheld fields from training to test performance on unfamiliar locations. The authors also grouped nearby fields to reduce the risk of geographic information leaking between training and evaluation. That matters: predicting another measurement from familiar surroundings is a less demanding test than transferring useful knowledge to a different site.

The authors report three practical findings:

• With two samples per field, the approach recovered more than 60% of the highest-removal fields.
• It roughly halved the samples needed to estimate a field’s average removal.
• Sampling reductions on the densest fields ranged from 28% to 98%.

Those figures describe different evaluation outcomes, not interchangeable promises for every farm.

Most importantly, the researchers explicitly say predicting an unfamiliar field’s average removal without field-specific measurements is not yet feasible. The supported approach combines a model with a small number of physical samples.

Why a weather model belongs in this experiment​

Aurora is an Earth-system foundation model, not a language model asked to improvise soil chemistry. Microsoft Research says Aurora 1.5 expanded its atmospheric coverage with 22 additional weather variables, hourly temporal resolution and ensemble forecasting.

For TerraNova-ERW, the researchers kept the weather models fixed and trained smaller prediction components using their atmospheric representations alongside site information. In plain English, they reused patterns learned from weather data rather than training an entire environmental model from scratch.

For software developers, the interesting lesson is transfer learning: a specialized foundation model can supply useful inputs to a different scientific task. But the field-specific limitation is equally instructive. Reusable representations do not remove the need for relevant, well-measured local data.

Promising efficiency, not automatic carbon certification​

Microsoft is both a research participant and an investor in Terradot through its Climate Innovation Fund, according to Microsoft Source. Its account therefore needs to be read as a participant’s report, not independent validation. The company also acknowledges that further development and extensive validation remain necessary.

The practical interpretation is narrower—and more useful—than “AI solves carbon removal.” If these results hold up more broadly, better predictions could help scientists spend their sampling budgets where measurements are most valuable. That would make AI an efficiency tool inside a scientific verification process, rather than a substitute for one. The study’s own distinction between sample-assisted estimation and unsupported sample-free accounting is the boundary worth watching.