Kyle Heyblom focuses on improving the accuracy and global equity of weather forecasts by integrating artificial intelligence with novel Earth observations. Many regions, particularly in the tropics, face increasing exposure to extreme weather as the climate warms, yet they continue to receive the least reliable forecasts. Kyle aims to reduce this gap by expanding the physical information available to AI-based prediction systems.
Modern forecasts rely heavily on numerical weather prediction models, which solve large systems of physical equations. These models have been essential, but they are computationally expensive and must approximate many important atmospheric processes. Recent AI weather models offer a powerful complement. They can learn predictive relationships directly from extensive archives of observations and simulations and often achieve comparable or superior accuracy at far lower computational cost. Because current AI models are trained primarily on numerical model output, however, their skill is constrained by the structural biases they inherit from those numerical systems.
Kyle develops AI methods that learn from sparse remote sensing measurements and incorporates this information into state-of-the-art AI forecasting systems. He uses targeted and novel data streams to create weather prediction models that are faster, more physically grounded, and better suited to regions with limited forecasting resources.