AgriPINN: A process-informed neural network for interpretable and scalable crop biomass prediction under water stress

Accurate estimation of crop above-ground dry biomass (AGB) under water stress is important for assessing agricultural production, managing irrigation, and supporting water-security-oriented decision-making. Data-driven models enable efficient large-scale prediction but often provide limited physiological interpretation and can be sensitive to domain shifts, whereas process-based crop models require extensive calibration and detailed management inputs. We propose AgriPINN, a process-informed neural network that embeds a differentiable biomass-growth equation derived from LINTUL5 into neural-network training. AGB is the sole supervised target. The network simultaneously estimates leaf area index (LAI), photosynthetically active radiation (PAR), radiation-use efficiency (RUE), and the water-stress factor ( F W ); their predicted values enter the process residual, whereas the corresponding field measurements are held out from optimisation and used only for post hoc evaluation. AgriPINN is pretrained on 65 years of SIMPLACE simulations covering 397 German NUTS-3 regions and fine-tuned on three years of wheat and maize field experiments under sheltered, rainfed, and irrigated treatments. On the main benchmark, AgriPINN reduces RMSE by 29.7% for winter wheat and 22.8% for maize relative to CNN–Transformer. Relative to the strongest non-AgriPINN baseline for each crop, the corresponding reductions are 9.0% and 2.2%. Additional evaluations include four representative neural backbones, sensitivity analysis of the process-loss weight, an ablation without pretraining, and perturbation tests on simulated labels. The transferability demonstrated in this study is limited to the independent datasets and the German temperate agro-ecological setting evaluated here.