We Have Weather Forecasts. Why Not Soil Forecasts?
The Team-Soil project — BioSensor Solutions and AreaandDee LLC, working with NVIDIA — is building a way to forecast what living soil will do next. We're looking for partners to help us put sensors in the ground.
Every grower checks the weather. Almost nobody can check the biology.
Farmers can measure the physical side of soil: how wet it is, how warm, how sandy. What they can't see is the living side — the microbes that decide whether nitrogen gets released to the crop or locked away, whether carbon stays in the ground or escapes into the air. Today the only way to find out is to mail a sample to a lab and wait weeks for a report on a moment that has already passed. Decisions about fertilizer, water, and timing worth billions of dollars a year get made on that delay.
A soil forecast would change that. If a grower could see that the microbes were about to release nitrogen on their own, they could hold off on fertilizer. If the biology showed the soil was still holding water, they could skip an irrigation pass. The aim is fewer wasted inputs, lower costs, and over time better yields — decisions made on what the soil is about to do, not what it did last month.
What we've built so far - Two parts
The first is a sensor. BioSensor Solutions makes a printed, biodegradable sensor, developed in Dr. Gregory Whiting's lab at CU Boulder, that sits in the soil and continuously reports microbial activity, moisture, temperature, and how much CO₂ the soil is giving off. It has been through an independent field season at Clemson's Edisto Research and Education Center and is described in two peer-reviewed papers published this year. Ten units are about to go into the ground at the South Carolina Governor's School for Agriculture at John de la Howe.
The second is a model. AreaandDee has built software that takes weather conditions and basic soil properties and predicts how the soil's microbes will respond — essentially, how hard the soil is breathing. The work began at the NSF ASCEND Engine's first hackathon, where we were one of seven teams selected, with NVIDIA supporting the modeling. NVIDIA has since written it up on its Earth2Studio developer blog, where the technical details live.
One thing to say plainly: the weather data the model learned from is real, but the soil-biology data was simulated, because real sensor readings are only now starting to come in. Everything below should be read with that in mind.
What the testing showed
We tested the model the hard way — by hiding data from it and seeing what it could predict.
Hide a full year at a location it already knows, and it predicts about three-quarters of what actually happened that year. That's good.
Hide an entire region it has never seen, and it gets shakier. It does worst on the unusual soils: the richest in organic matter, the sandiest. Put simply, the model can't predict soil it has never met.
Why that's the plan, not the problem
The AI models behind chatbots learned from an internet anyone could copy. There is no internet of soil biology. That data has to be gathered the slow way, one sensor and one growing season at a time — which is exactly why it's worth so much once someone has it.
Our testing tells us which soils are missing. The sensor tells us how to fill the gap. Whoever gets sensors into enough of that ground first holds the dataset.
The timing is right. Weather-forecasting AI like NVIDIA's Earth-2 is now mature enough to build on. Carbon and water rules are pushing buyers to want measured numbers, not estimates. And people are already paying: BioSensor Solutions is running a paid pilot with the International Rice Research Institute, and S&P Global Energy is building Earth-2 forecasts into its risk analysis. The soil layer is the missing piece.
What we're looking for
Places to put sensors — growers, research farms, and agricultural organizations willing to host them, especially on unusual soils: very high organic matter, very sandy, flooded rice paddies. Those fields close the gap fastest.
Institutions with acreage — like our work with the SC Governor's School, but more of it.
Analytics and risk partners — companies running climate forecasts through risk models for lending, crop insurance, energy, or commodities. We're looking for the first one to plug soil into that pipeline.
Next, we connect the model to NVIDIA Earth-2's weeks-ahead weather forecasts. That's the step that turns it from a description of what soil did last week into a forecast of what it will do next. Every grower already checks the weather before deciding what to do in the field. We want the soil forecast sitting right next to it.
This will take growing seasons, not sprints, and the sensors that make it possible have to be in real fields on real farms. If you have ground, we'd like to talk.
The Team-Soil project is a collaboration between BioSensor Solutions, AreaandDee LLC, and NVIDIA. David Beitz and Sam Walker lead the biosensor work at BioSensor Solutions. Rich Loft leads the machine-learning work at AreaandDee. Sepideh Khajehei mentors the project from NVIDIA. The full technical write-up — architecture, GPU optimization, and complete cross-validation results — is on the Earth2Studio blog.