NeuralCrop: Combining physics and machine learning for improved crop modelling
Authors: Yunan Lin, Sebastian Bathiany, Maha Badri, Maximilian Gelbrecht, Philipp Hess, Brian Groenke, Jens Heinke, Christoph Müller, Niklas Boers
Organizations: Munich Climate Center and Earth System Modelling Group, Department of Aerospace and Geodesy, School of Engineering and Design, Technical University of Munich, Munich, Germany. · Potsdam Institute for Climate Impact Research, Potsdam, Germany. · University of Exeter, Exeter, UK.
Global gridded crop models (GGCMs) are crucial for projecting climate impacts on agricultural productivity and assessing associated risks to food security. However, state-of-the-art GGCMs are subject to substantial uncertainties arising from reduced representations of complex processes. Projected increases in the frequency and severity of climate extremes require more reliable agricultural yield simulations. Recent machine learning approaches trained on observational data provide alternatives to GGCMs, but remain under-constrained due to sparse observations and show poor out-of-distribution generalization. Here we introduce NeuralCrop, a differentiable hybrid GGCM that combines the strengths of an advanced process-based GGCM (LPJmL) with data-driven components. NeuralCrop is first trained to reproduce its underlying fully process-based version, before it is fine-tuned on observational flux data of physiological variables. Evaluation against held-out flux-tower observations from European wheat and US maize sites shows 32.6-45.2% reductions in mean normalized root-mean-square error for daily carbon and water fluxes. Although fine-tuned only on physiological variables, NeuralCrop improves temporal correlations with independent yield records across 71.4% of the cultivated area for European wheat and 84.0% for the US Corn Belt during 1981-2019. It performs highly competitively compared to GGCMI Phase~3 models. NeuralCrop reproduces yield anomalies more accurately across a range of soil-moisture conditions and especially for extreme drought conditions. For large-scale, long-term simulations, NeuralCrop runs nearly two orders of magnitude faster than the CPU-based LPJmL baseline model, mainly thanks to GPU acceleration. Our results highlight end-to-end hybrid crop modelling as a promising avenue for improving crop simulations and informing food security assessments under climate change.
Figures & tables
Figure 1: Schematic of the NeuralCrop rollout and two-stage training framework. a, At each time step t , NeuralCrop receives forcings xt and the previous model state st−1 . The forcings include weather, soil properties, and field management, such as crop calendars, tillage, and fertilizer. Orange outlines identify the carbon and water processes that contain neural network components. Nitrogen and energy retain fully process-based representations. These coupled processes produce outputs y^t and an updated state st , which is carried forward to the next time step. Training minimizes L(y^t,yt∗) , where yt∗ denotes the target for the corresponding training stage, and loss gradients propagate through the differentiable process calculations to update the neural-network parameters. b, During regional-scale pre-training, outputs generated by NeuralCrop in its fully process-based version provide the targets for Lpre , yielding parameters θpre . c, During site-scale fine-tuning, the neural-network parameters are initialized with θpre and further optimized against eddy-covariance observations of gross primary productivity (GPP), ecosystem respiration (RECO), and evapotranspiration (ET) using Lfine . No simulated or observed crop yield data are used for training at any stage. The maps show nine wheat and seven maize flux-tower sites. The two Pennsylvania maize sites, which are included in the analysis but lie outside the main Corn Belt map, are shown in an inset. Held-out flux-tower observations and independent subnational yield statistics are used for evaluation.
Figure 2: Held-out evaluation of daily carbon and water fluxes at flux-tower sites. a,b , Agreement between simulations and observations across 50 crop seasons held out from training. a , Pearson correlation, and b , normalized root mean square error (nRMSE) for gross primary productivity (GPP), ecosystem respiration (RECO), and evapotranspiration (ET). Each point in a–b represents one held-out season. Boxes show the median and interquartile range, and whiskers extend to 1.5 times the interquartile range. Annotations above a and b show the change in mean correlation and the percentage change in mean nRMSE, respectively, for the hybrid relative to the default simulations. Brackets show 95% intervals from 10,000 bootstrap resamples of site groups. c,d , Examples of observed and simulated daily GPP, RECO, and ET during the held-out 2017 winter wheat season at DE-RuS in western Germany ( c ) and the 2019 maize season at US-UC2 in central Pennsylvania, USA ( d ). The default simulations use the fully process-based configuration, with all neural network components disabled. The hybrid model version reproduces the observations much more accurately than the default version.
Figure 3: Evaluation of yield simulations and comparison with AgMIP models. a,b , Administrative-unit temporal correlation ( r ) between reported and simulated European wheat yields for the default and hybrid simulations over 1981–2019. c , Difference between the correlations obtained with the hybrid and fully process-based (default) model versions, respectively ( Δr ). Positive values indicate improvement by the hybrid version and occur across 71.4% of mapped wheat area. d,e , Corresponding county-level default and hybrid simulations for US Corn Belt maize over 1981–2019. f , Corresponding Δr map. Positive values occur across 84.0% of mapped maize area. g , Crop area-weighted mean correlations for the default and hybrid simulations and the AgMIP GGCM distributions over the available 1981–2016 period. Points within the AgMIP boxplots denote individual GGCMs.
Figure 4: Sensitivity of regional yield skill to the individual neural network components. The left column shows European wheat and the right column US maize. a,b , Crop area-weighted temporal correlation between detrended simulated and reported annual yields for the full hybrid version and three variants evaluated during inference. In the GPP ( λ , Vmax ) variant, process-based calculations replace the neural network predictions of these variables. In the respiration and transpiration variants, the neural corrections to crop respiration and to root-water supply in crop transpiration, respectively, are removed. c,d , Corresponding crop area-weighted normalized RMSE (nRMSE). Higher correlation and lower nRMSE indicate better agreement. e,f , Percentage of cultivated area in four administrative-unit temporal-correlation classes ( r<0 , 0≤r<0.3 , 0.3≤r<0.5 , and r≥0.5 ) for the four versions.
Figure 5: Yield anomalies across soil moisture conditions. a,b , Crop area-weighted mean absolute error (MAE) of standardized yield anomalies for European wheat and US maize, grouped by the growing season ERA5-Land root zone soil-moisture index of each administrative-unit year. The index is the crop area-weighted mean of grid-cell soil-moisture percentile ranks relative to 1981–2010. Anomalies are standardized separately for each yield series within each administrative unit. Points show the estimates and bars show 95% intervals from 5,000 country- or state-cluster bootstrap resamples. Default and hybrid simulations are evaluated over 1981–2019, whereas AgMIP ensemble median is evaluated over 1981–2016. c,d , Crop area-weighted regional mean yield anomalies from reported statistics, default, hybrid, and AgMIP ensemble median. Insets show the ratio of simulated to observed interannual standard deviation. Annotations in the top left of each panel give the temporal correlation ( r ) and RMSE between the reported regional mean yield time series and the corresponding series from the default and hybrid simulations over 1981–2019 (default → hybrid). For European wheat, correlation (RMSE) is thus improved by 80.6 (44.1)%, while for US maize correlation (RMSE) is improved by 25.6 (32.5)% by NeuralCrop compared to the fully process-based baseline. e,f , Crop area-weighted mean yield anomaly in 10-point soil-moisture index bins. Shading shows 95% country- or state-cluster bootstrap intervals. The n labels in a,b,e,f give the number of distinct calendar years contributing at least one administrative-unit record to each moisture bin in the evaluation over 1981–2019.
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1: Schematic of data flow through the trainable components of NeuralCrop. The schematic also includes optional components that are not activated in the main-text simulations.
Figure S2: Comparison of soil water content (SWC) simulated by hybrid and pure machine learning models under in-sample and out-of-sample scenarios, with LPJmL-simulated SWC used as baseline. Panels a and b show in-sample performance for the hybrid and pure machine learning models, respectively, while panels c and d present their out-of-sample performance.
Variable
Symbol
Dimension
Units
Time
Inter-cellular to atmospheric CO 2 ratio
λ
1
-
daily
Maximum Rubisco capacity
Vmax
1
gCm−2
daily
Gross primary productivity
GPP
1
gCm−2
daily
Crop carbon pools
Ci
4
gCm−2
daily
Crop respiration
Rcrop
1
gCm−2
daily
Crop transpiration
Tcrop
1
mm
daily
Appendix
Table S1: Training data generated by the fully process-based simulations.
Type
Variable
Units
Time
Climate
Temperature
∘C
daily
Climate
Precipitation
mm
daily
Climate
Net longwave radiation
W/m2
daily
Climate
Downward shortwave radiation
W/m2
daily
Climate
CO 2
ppmv
annual
Soil
Soil texture
-
static
Appendix
Table S2: Model inputs.
Type
Variable
Symbol
Units
Time
Feature
Air temperature
TAF
∘C
daily
Feature
Precipitation
PF
mm
daily
Feature
Downward shortwave radiation
SWIN_F
W/m2
daily
Feature
Downward longwave radiation
LWIN_F
W/m2
daily
Feature
Upward longwave radiation
LWOUT
W/m2
daily
Feature
CO 2
CO2F_MDS
μmolmol−1
daily
Appendix
Table S3: Flux network observational data.
Figure S3: NeuralCrop models with identical model parameters trained on rollouts of 1, 30, 60, 180, 365, and 720 days (a-f).
Grid cells
LPJmL
NeuralCrop
Times (speed up)
1
0.50 s
46.01 s
0.011x
1573
833.99 s
51.24 s
16.27x
7865
4169.94 s
70.02 s
59.55x
14157
7505.89 s
91.16 s
82.33x
Appendix
Table S4: Inference time of GPU-accelerated NeuralCrop and CPU-based LPJmL simulating for a 0.5∘×0.5∘ spatial resolution with daily time steps over a 20-year simulation period (7300 days) for different numbers of grid cells. NeuralCrop was running on one single NVIDIA H100 GPU with 80GB of memory, and LPJmL was running on a CPU node equipped with two AMD EPYC 9554 processors (128 cores in total) and 768 GB of memory.
Figure S4: Inference time of GPU-accelerated NeuralCrop and CPU-based LPJmL simulating for a 0.5∘×0.5∘ spatial resolution with daily time steps over a 20-year simulation period (7300 days) for different numbers of grid cells. NeuralCrop was running on one single NVIDIA H100 GPU with 80GB of memory, and LPJmL was running on a CPU node equipped with two AMD EPYC 9554 processors (128 cores in total) and 768 GB of memory.
Figure S5: Distribution of flux tower sites used in the grouped site held-out experiment. a , Global locations of the independent site groups. Orange and blue markers denote European wheat and US maize groups, respectively. Marker size indicates the number of physical sites in each group. b,c , Enlarged the physical wheat and maize tower sites. Markers for nearby sites in the same group are displaced for visibility. Site names and season counts are given in Table S5 . The evaluation contains 10 independent site groups, 16 physical sites, and 50 crop seasons.
Crop
Group
Sites
Grouping basis
Seasons
Wheat
W01
DE-Geb
standalone site
4
Wheat
W02
DE-RuM, DE-RuS, DE-Seh
nearby sites
11
Wheat
W03
FR-Aur, FR-Lam
nearby sites
8
Wheat
W04
FR-EM2
standalone site
2
Wheat
W05
FR-Gri
standalone site
2
Wheat
W06
IT-CA2
standalone site
2
Appendix
Table S5: Flux tower sites and independent site groups used for the holdout evaluation. Season counts are the crop seasons remaining after quality control.
Crop
Target/metric
Default
Hybrid
Improvement
95% CI
Wheat
GPP r
0.716
0.896
+0.180
[0.102, 0.293]
Wheat
RECO r
0.763
0.877
+0.114
[0.084, 0.143]
Wheat
ET r
0.681
0.875
+0.194
[0.163, 0.226]
Wheat
GPP nRMSE
0.995
0.488
0.507
[0.353, 0.648]
Wheat
RECO nRMSE
0.677
0.390
0.287
[0.163, 0.410]
Wheat
ET nRMSE
0.778
0.528
0.249
[0.181, 0.314]
Appendix
Table S6: Performance at sites excluded from training. The table reports mean performance metrics for the default and hybrid simulations, assigning equal weight to each site group, together with the paired mean improvement and its 95% bootstrap interval. Positive improvement indicates higher correlation or lower nRMSE for the hybrid simulations than for the default simulations.
Figure S6: Training loss trajectories and sensitivity to random seed for the grouped site holdout evaluation. a,b , Normalized training objective over 150 epochs for wheat and maize. Thin curves show individual fold and seed runs, solid curves show the median at each epoch, and shaded bands show the interquartile range. c,d , Differences in correlation and nRMSE between the default and hybrid simulations for GPP, RECO, and ET in seasons excluded from training. Colors denote the three random seeds, and points represent individual seasons.
Figure S7: Spatial distribution of grid-cell temporal correlations between the default and pre-trained simulations for European wheat regions at 0.5 ∘ resolution over 1981–2019. a–c , Grid-cell Pearson temporal correlation calculated directly from the annual crop-season GPP, ET, and end-of-season yield series, respectively. Rain-fed and irrigated outputs are combined using fixed-2015 crop area. Panel annotations give the median across valid grid cells.
Figure S8: Spatial distribution of grid-cell temporal correlations between the default and pre-trained simulations for US Corn Belt at 0.5 ∘ resolution over 1981–2019. a–c , Grid-cell Pearson temporal correlation calculated directly from the annual crop-season GPP, ET, and end-of-season yield series, respectively. Rain-fed and irrigated outputs are combined using fixed-2015 crop area. Panel annotations give the median across valid grid cells.
Figure S9: Evaluation of yield simulations and comparison with AgMIP models. a,b , Administrative-unit normalized RMSE (nRMSE) between reported and simulated European wheat yields for the default and hybrid simulations over 1981–2019. c , Difference between the nRMSE obtained with the hybrid and fully process-based (default) model versions, respectively ( ΔnRMSE ). Negative values indicate improvement by the hybrid version and occur across 98.2% of mapped wheat area. d,e , Corresponding county-level default and hybrid simulations for US Corn Belt maize over 1981–2019. f , Corresponding ΔnRMSE map. Negative values occur across 98.4% of mapped maize area. g , Crop area-weighted nRMSE for the default and hybrid simulations and the AgMIP GGCM distributions over the available 1981–2016 period. Points within the AgMIP boxplots denote individual GGCMs.
Figure S10: Regional yield-correlation gains by radial distance from flux-tower training sites over 1981–2019. Geodesic distances are measured from administrative-unit polygon centroids to the nearest crop-specific training-site model grid-cell center. Each distance class contains approximately one-quarter of the mapped crop area. a,b , Crop area-weighted mean change in administrative-unit temporal correlation ( Δr ) for European wheat and US maize. c,d , Mapped crop area with positive Δr in the same distance classes. Points show the estimates. Bars show 95% intervals from 5,000 country-cluster (wheat) or state-cluster (maize) bootstrap resamples. The distance classes differ between crops because their training-site distributions and mapped crop areas differ.
Figure S11: Regional yield sensitivity to random seed over 1981–2019. a,b , Distributions of administrative-unit temporal correlation and nRMSE between simulated and reported European wheat yields for three training seeds. c,d , Corresponding distributions for US Corn Belt maize. Points denote administrative units. Boxes show the median and interquartile range and whiskers extend to 1.5 times the interquartile range. nRMSE is displayed on a logarithmic axis.
Figure S12: Association between changes in seasonal process–yield correlations. a,b , Gross primary productivity (GPP), c,d , crop respiration, and e,f , crop transpiration. Left panels show European wheat and right panels show US Corn Belt maize over 1981–2019. Each point represents an administrative unit. Marker size increases with mapped crop area. Horizontal and vertical axes show changes from the default to the hybrid simulation in temporal correlation with reported yield for the simulated process variable and simulated yield, respectively. Annual process, simulated yield, and reported yield series are linearly detrended separately before correlations are calculated. Crop respiration and transpiration are summed over the active crop season. Solid orange lines show crop area-weighted linear fits. Dotted lines show the 1:1 relationship. Quadrant percentages give mapped crop-area shares. The 95% confidence intervals for the area-weighted correlations use 5,000 spatial-cluster bootstrap resamples of countries for wheat and states for maize.
European wheat
US maize
Index interval
n
Leading year (%)
Top three (%)
n
Leading year (%)
Top three (%)
20-point index classes
<20
35
2011 (11.2)
29.2
37
1988 (13.7)
35.7
20–40
38
2015 (8.0)
23.2
39
1994 (8.3)
20.5
40–60
39
2008 (8.0)
20.8
39
2002 (6.2)
16.5
60–80
38
2001 (10.3)
22.8
39
1984 (6.5)
18.5
Appendix
Table S7: Calendar-year support and concentration of crop-area weights in the soil-moisture index bins. The count n is the number of distinct calendar years with at least one administrative-unit year in the bin. The leading and top-three year shares are percentages of the total crop-area weight summed over administrative-unit years in that bin. Counts and shares use records over 1981–2019.
Figure S13: Sensitivity of the soil moisture response to anomaly scaling. Yield anomalies are expressed as percentages of the fitted linear trend for each administrative unit. Crop area-weighted means are then calculated in ten 10-point bins of the growing-season root-zone soil-moisture index. Panels a and b show European wheat and US maize, respectively, with curves for observations and the default and hybrid simulations. The shading shows 95% intervals from 5,000 country- or state-cluster bootstrap resamples.
Department of Mathematics and Department of Computer Science, Iowa State University, Ames, IA, USA. · Department of Horticultural Sciences, University of Florida, Gainesville, FL, USA. · Department of Mechanical Engineering, and Translational AI Center, Iowa State University, Ames, IA, USA.
Artificial Intelligence Group, Wageningen University and Research, Wageningen, the Netherlands · Earth Observation for Agroecosystems, Agroscope, Zurich, Switzerland · Potsdam Institute for Climate Impact Research (PIK), Germany +4