Substantial Differences in Crop Yield Sensitivities Between Models Call for Functionality‐Based Model Evaluation

Crop models are often used to project future crop yield under climate and global change and typically show a broad range of outcomes. To understand differences in modeled responses, we analyzed modeled crop yield response types using impact response surfaces along four drivers of crop yield: carbon dioxide (C), temperature (T), water (W), and nitrogen (N). Crop yield response types help to understand differences in simulated responses per driver and their combinations rather than aggregated changes in yields as the result of simultaneous changes in various drivers. We find that models' sensitivities to the individual drivers are substantially different and often more different across models than across regions. There is some agreement across models with respect to the spatial patterns of response types but strong differences in the distribution of response types across models and their configurations suggests that models need to undergo further scrutiny. We suggest establishing standards in model evaluation based on emergent functionality not only against historical yield observations but also against dedicated experiments across different drivers to analyze emergent functional patterns of crop models. Plain Language Summary Crop models are widely used to compute crop yields under future climate change. Yields are determined by many interacting processes. Simulated future crop yields often show a broad uncertainty range. We investigate the sensitivity of nine different crop models to individual model inputs (carbon dioxide, temperature, water, nitrogen) in a very large simulation data set and find that there are substantial differences. We conclude that crop model evaluation needs to include analyses of functional properties to avoid that very diverse model responses to drivers are not tracked if interacting processes cancel out in the historical evaluation period but not in future scenarios, leading to large differences between models. Key Points Crop models show strong differences in input sensitivities Standardized modeling experiments reveal differences in emergent functional relationships New standards in model evaluation are needed