)
)
Most of the weather we care about doesn’t happen where the sensors are. It happens in between them: on a busy stretch of a road, over a wind farm, across a city block where no one is measuring anything in real time. For businesses, those gaps matter: they’re where assets sit exposed, road conditions affect driving safety, and energy consumption is shaped.
We’re now working to deliver accurate forecasts for that space in between, and we’re doing it with our next-generation deep learning forecasting model. Its goal is, simply put, to predict ground truth where it does not exist. An uneasy task, but we’re seeing some incredible results.
The space in between, aka the missing piece
Most accurate weather observations come from ground stations: precise, but sparse. Each station tells you exactly what’s happening at that point in space, but not what’s going on ten, twenty, or fifty kilometers away.
To build a truly useful forecast, you need to fill in those blanks. We’re now working on exactly that. Instead of treating sensors as isolated dots, the model learns how the atmosphere behaves between them and generates a contiguous picture: a dense forecast field over an entire area based on sparse, real-world‑ measurements.
Think of it as creating a virtual sensor for every point on the map.
“Fooling” the model for better forecasts
Normally, in weather forecasting, the input format matches the output format. With ground-truth sensor data, we can use time series historical data, and our predictions come out as a time series, too. But this time we’re shaking things up and blending different input formats to produce very accurate and dense forecasts by combining sparse ground truth with radar data, satellite imagery, and km-scale local NWP model.
)
What happens as a result is something truly remarkable; the output is a data-rich image: a grid where each pixel represents a forecast at that location. Using deep learning, the model meshes all context into a single, coherent scenario. It goes beyond station data, showing what the weather parameters are like between observation points. While precipitation forecasting with this approach has already been in production for a while, we can predict any number of variables measured by the sensors: temperature, wind speed, wind gusts, and more.
The model is trained on up to two years of historical data. During training, we know the “answer” at the sensor locations, so we can teach the model how patterns in satellite, radar, and NWP forecasts translate into what would have been measured on the ground. We then evaluate the forecasts by comparing them against ground-truth observations not used in the training dataset. The results are quite impressive: we already outperform high-resolution NWP forecasts.
Building a better map with fewer footsteps
This approach is like drawing a highly detailed map of a territory your foot never touched, guided only by a handful of trusted landmarks and a lot of learned experience.
For weather sensitive‑ businesses, that means:
No blind spots between stations
More confidence in local conditions that affect safety, uptime, and logistics
A forecast that doesn’t just simulate the atmosphere, but also understands what sensors would be telling you if they could
We’re now in the midst of the work—testing, benchmarking, and fine-tuning. But the direction is clear: dense, accurate forecasts built from sparse, real-world measurements are not just possible; they’re‑ already outperforming the status quo.
)
)
)