A team from the Technical University of Munich, the Potsdam Institute for Climate Impact Research, and partner institutions describes NeuralCrop, a differentiable hybrid of the LPJmL crop model and several small neural networks. In historical evaluations of European wheat and US maize, the authors report stronger yield estimates than the original process-based model and competitive results against data-only baselines and a multi-model ensemble. The preprint focuses on climate-impact analysis and crop-risk simulation; it does not validate an operational next-season forecasting system.

Why harvest simulations diverge

Global gridded crop models, known as GGCMs, simulate photosynthesis, respiration, soil water, transpiration, and plant growth across the planet, commonly at a resolution near 0.5 degrees, or roughly 50 kilometers. Governments and researchers use their output to estimate how crop yields may respond to drought, warming, changing rainfall, and management choices.

Within AgMIP, the Agricultural Model Intercomparison and Improvement Project, 14 modeling groups have contributed historical simulations for maize, wheat, rice, and soybean. Their models show mixed skill when reproducing year-to-year yield changes and geographic yield patterns. Those disagreements propagate into estimates of drought losses, adaptation benefits, and future food production.

Process-based models encode biological mechanisms through simplified equations and calibrated parameters, leaving some responses to heat and water stress poorly represented. Data-only yield models can fit historical relationships, but future climates may fall outside their training distribution. NeuralCrop keeps LPJmL’s process structure while assigning selected, uncertain calculations to trainable functions.

Where the neural pieces go

NeuralCrop starts from LPJmL, a dynamic global vegetation model that represents natural vegetation, managed cropland, grassland, and terrestrial hydrology in one simulation. The researchers rewrote relevant components so the model supports automatic differentiation, which tracks how each calculation contributes to the final error and allows training signals to pass through an entire growing season.