Crushed rock, climate and a weather model
Spreading powdered rock on farmland sounds low-tech, yet it is one of the more promising ways to pull carbon dioxide out of the air. On November 12, 2026, Microsoft’s Source feature described TerraNova, a new AI model from carbon-removal company Terradot and Microsoft that aims to make the technique cheaper to verify and easier to scale.
How enhanced rock weathering works
Rainwater slowly breaks down certain rocks through chemical reactions that trap carbon dioxide, a natural climate regulator that normally takes thousands of years. Enhanced rock weathering speeds it up by crushing rock into powder, which exposes more surface area. Terradot works with farmers to spread the powder over agricultural fields while crops keep growing. Microsoft notes an added benefit: some rocks can reduce soil acidity, similar to limestone, which may improve nutrient availability and yields.
The measurement bottleneck
The catch is proof. How much carbon is removed depends on the rock type, soil chemistry, rainfall, humidity and temperature, and conditions vary from field to field. Scientists must collect many soil samples before and after spreading, then analyze them with specialized lab equipment. Terradot’s team of roughly two dozen PhD-level scientists has gathered tens of thousands of samples from fields in Brazil and the U.S. Microsoft says measuring and verifying removal can account for a third to more than half of the cost. Site selection is a second challenge, since trial and error is expensive.
What TerraNova does
TerraNova builds on Aurora, Microsoft’s foundation model for weather and environmental conditions. Terradot co-founder Sasankh Munukutla worked with the Microsoft Climate Innovation Fund, AI for Good Lab and Microsoft Research to adapt Aurora to predict chemical changes in soil, using data from fields where Terradot works with farmers. The model considers a field’s soil composition, minerals in rock from nearby quarries, regional humidity, rainfall, temperature and global patterns such as La Niña to predict carbon removal under different conditions, helping identify where the process is likely to work best.
In one example, training on data from Brazilian fields let it make predictions for fields in Mississippi.
Munukutla said he first saw the idea at a chocolate maker on Hawaii’s Big Island in 2024, where a display explained that tropical soils are “fast-forwards” of soils elsewhere because warm, wet climates accelerate natural weathering. That made him wonder whether Aurora, trained on more than a million hours of atmospheric data, could learn geochemistry too. Microsoft’s Climate Innovation Fund, an early Terradot investor, called the joint effort a “moonshot” aimed at solving a problem holding back the whole industry’s growth.
The early numbers, and the caveats
TerraNova doesn’t replace physical measurement. It combines predictions with a smaller number of soil samples. In simulations using field data, researchers found it could cut the samples needed at the largest sites by 75% while keeping the same accuracy. Microsoft is careful to say that is conditional on further research confirming the results. The work will be described in a forthcoming preprint ahead of peer review, and earlier versions were accepted for a NeurIPS 2026 climate workshop and presented at an ICLR 2026 AI-for-science workshop.
Terradot says rock powder spread in Brazil and the U.S. is expected to remove up to 100,000 tons of CO2 over time, and it aims to reach one million tons a year by 2035. Munukutla stresses that extensive validation is still needed to use such models responsibly.
What it means
Carbon removal buyers and governments want credible evidence, and credibility is expensive. If AI can shrink the sampling burden without eroding trust, verification costs could fall and more projects become viable. The broader lesson is that a model built for weather turned out to be adaptable to soil chemistry, a sign that foundation models may carry over to other planetary-scale problems. Readers should treat the 75% figure as a simulation result, not a field-proven outcome.
Conclusion
TerraNova is early, but it targets a real barrier: proving how much carbon the ground has absorbed. Peer review and field validation will decide whether it lives up to the promise.