A machine learning-derived metamodel of BEEHAVE predicts how honey yield depends on weather and region across Germany

Honeybees are vital pollinators but face growing stress from weather, land-use change, and parasites. Detailed simulation models like BEEHAVE help explore these impacts but are slow, limiting large-scale applications. To address this, we developed machine learning metamodels that emulate BEEHAVE outputs. We ran BEEHAVE simulations using a new, faster Go implementation and generated millions of synthetic weather scenarios with our SynHr weather generator. Using these data, we trained two metamodels, a Neural Network and an XGBoost model, providing a comparison between a slower training method and a faster one. Applied to historical weather data across Germany, both metamodels accurately reproduce BEEHAVE’s annual honey yield predictions. Our spatial and temporal simulations confirmed a positive linear relationship between foraging hours and honey production that saturates at high foraging hours. The worker bee population peaked at intermediate foraging levels and declined beyond that. This work demonstrates how weather influences colony performance and shows that metamodeling can effectively complement mechanistic models, enabling scalable digital twin applications for environmental research.