From climate hazards to yield losses: AI surrogate impact modelling

We present the Surrogate Engine for Crop Simulations for Maize (SECS4M), a deep-learning emulator designed to replicate the process-based ECroPS crop growth model for grain maize in Europe while enabling computationally efficient, large-scale applications in climate services. SECS4M is built on a nested Long Short-Term Memory architecture capturing short- and long-term weather–crop interactions, while it ingests only three daily meteorological inputs, minimum and maximum temperature and total precipitation, thus minimizing the uncertainty that follows the use of a much wider input stream as in ECroPS. Trained on ERA5-forced yield outputs, SECS4M accurately reproduces crop growth trajectories, harvest timing, and yield distributions. Computational requirements are reduced from ~70s to ~0.008s per grid-cell–year, a four-order-of-magnitude speed-up that enables ensemble-scale, operational use.

Forced with bias-adjusted SEAS5.1 forecasts, SECS4M reproduces observed 2022 impacts and supports probabilistic identification of Areas of Concern (AoC) based on tercile-based yield anomalies. Under CMIP6 scenarios SSP3-7.0 and SSP5-8.5 to 2050, the emulator highlights specific regions as persistent hotspots of yield risk, while others exhibit mixed signals. SECS4M thus provides a scalable, digital twins enabled and data-efficient framework for seasonal forecasting, AoC mapping, and scenario analysis. Finally, the methodology can be extended to other crops and can be tested for its potential on other regions.