Microsoft and Terradot’s AI cuts soil sampling by up to 75% in carbon-removal simulations

Measuring carbon removal from powdered rock could require far fewer trips to collect soil. Microsoft and Terradot have developed TerraNova, an AI model that reduced sampling needs by up to 75% at the largest sites in simulations, Microsoft disclosed November 12. The approach targets a costly measurement problem without eliminating physical checks.

Why powdered rock needs so much measurement

The technique, called enhanced rock weathering, accelerates a natural chemical process. Certain rocks react with rainwater and carbon dioxide, removing CO₂ from the atmosphere. Crushing them into powder exposes more surface area and speeds up reactions that would otherwise take thousands of years. Terradot spreads the powder over farmland while crops continue to grow.

But spreading rock is not the same as knowing how much carbon it removes. Results depend on the rock’s minerals, soil chemistry, rainfall, humidity and temperature. Those conditions can differ between neighboring fields. Scientists collect samples before and after application, then use specialized laboratory equipment to examine changes in the soil’s chemical composition.

Terradot has collected tens of thousands of soil samples from fields in Brazil and the United States. Its scientists have also drilled wells to track water movement and taken other chemical measurements. That work helps establish where the process works best, while providing evidence for companies and governments paying for carbon removal.

From weather forecasts to soil chemistry

TerraNova builds on Aurora, Microsoft’s AI model for predicting weather and other environmental conditions. Terradot co-founder Sasankh Munukutla worked with Microsoft researchers and its AI for Good Lab to adapt Aurora to predict chemical changes in soil, using scientific data from fields where Terradot works with farmers.

The model considers soil composition, minerals in crushed rock from nearby quarries, local weather and global patterns such as La Niña. It uses those inputs to predict carbon removal under different conditions and identify promising locations. Microsoft researcher Felipe Oviedo said training on Brazilian field data enabled predictions for fields in Mississippi.

For measurement, the approach pairs model predictions with a smaller set of soil samples. How much rock reacts in the soil is one key input for estimating carbon removal. In simulations using agricultural-field data, researchers found that fewer samples could deliver the same level of accuracy at the largest sites, according to Microsoft.

The potential savings depend on validation

Measuring and verifying carbon removal can account for a third to more than half the cost of enhanced rock weathering, Microsoft says. Reducing sampling could therefore ease a substantial expense if further research confirms the results. Choosing effective sites also matters: assessing conditions above and below ground is complicated, and trial and error is expensive.

The research is awaiting publication as a preprint and has not been peer-reviewed. Earlier versions were presented at an ICLR 2026 AI-for-science workshop and accepted for a climate-focused NeurIPS 2026 workshop. Munukutla says continued development and extensive validation are needed to use models like TerraNova responsibly; Terradot plans further work with the scientific community.