Experimental data for model calibration and evaluation was obtained from winter wheat dryland plot trials established at the South-Central Agricultural Laboratory (SCAL) research station located near Clay Center, NE, USA (40° 34′ 30.92″ N, 98° 8′ 16.1664″ W) during the 2020/21, 2021/22, and 2022/23 growing seasons. The 2022/23 trial was established but discontinued due to severe drought conditions combined with biotic stress (e.g., pest pressure), which resulted in minimal crop establishment and near-complete loss of aboveground biomass. The trial was therefore terminated because meaningful field measurements could not be collected. The climate at the site is humid continental (warm, rainy summers) with annual precipitation of 725 mm and a mean temperature of 12°C. Daily weather data were obtained from the Nebraska Mesonet (2024). During the calibration years, growing-season (planting to harvest) precipitation totaled 526 mm in 2020/21 and 381 mm in 2021/22 (Supplementary Table S1). The soil classification was Crete silt loam (fine, smectitic, mesic Pachic Udertic Argiustolls). More details about soil properties and weather are described in Supplementary Table S1.

Field experiments were set in a randomized complete block design with two cultivars differing in phenology, a medium-early (LCS Valiant; here on LCS) and a medium-late (WB4699; here on WB), four N rates (0, 56, 112, and 168 kg ha^-1), and three replications in a randomized complete block design. Experimental units were 3 m long and 2 m wide (6 m²), and the row spacing was 0.23 m. Cultivars with contrasting phenology were selected since they are typically exposed to different growing conditions during the critical period, potentially impacting response to N (Cossani and Sadras, 2021; Sadras et al., 2022). The N rates were applied during spring as broadcasted ammonium nitrate (34-0-0) at Feekes growth stage (GS) 4 (Large, 1954). Winter wheat was planted at 250 seeds m^-2 on 22 September 2020, 8 October 2021, and 10 October 2022 (soybean-wheat-corn rotation). Moreover, light tillage was conducted shortly before wheat planting.

External validation data were obtained from the University of Nebraska-Lincoln (UNL) Winter Wheat Variety Test Results (University of Nebraska-Lincoln, 2022), which include yield data collected from five Nebraska counties evaluated across six growing seasons (2017-2022), totaling 29 site-year combinations. These trials were used to benchmark the model’s accuracy across diverse growing conditions. To capture variability in seasonal water availability, site-years were categorized into three precipitation environments based on total precipitation during the growing season (October-June): low (< 250 mm), medium (250–550 mm), and high (> 550 mm).

Crop phenology was recorded based on the Feekes scale (Large, 1954), and plant stand was counted at early stages. Feekes stages were assessed visually using ten randomly selected plants per plot. A plot was recorded at a given stage when at least half of the sampled plants showed the corresponding morphological trait. Stages were determined independently each year based on observed plant development. These field-sampled Feekes stages correspond to key developmental transitions (stem elongation, jointing, booting, heading/flowering, grain filling) that align with the phenology stages simulated by APSIM-NG. Crop and soil samples were collected in-season on five dates, corresponding to Feekes GS 4, 6, 10, 10.5, and 11.2, to quantify aboveground biomass (by plant fraction: stem, leaves, spikes), N uptake, soil nitrate, and ammonium. Aboveground plant biomass was determined by clipping a 0.4 m² area from the middle of each plot. The biomass was first partitioned into different plant organs (stem, leaves, and spikes) before being oven-dried at 50°C for at least one week to constant weight. Soil moisture was measured during the growing seasons at depths of 30, 60, and 90 cm with tensiometers, with two replicates per depth installed in both the eastern and western sections of the field. These points were selected to represent the treatments of low N rate (0 kg N ha^-1) and high N rate (168 kg N ha^-1). Additionally, soil temperature sensors were placed at a 30 cm depth. All sensors were connected to a datalogger for continuous data collection.

Five soil cores were collected from each plot and homogenized to quantify soil texture, soil organic matter (SOM), soil nitrate (NO₃^-), and ammonium (NH₄⁺) content at 0-20, 20-40, and 40–60 cm depths, each year before sowing (Supplementary Table S1). Other nutrients, including phosphorus, sulfur, zinc, iron, manganese, copper, calcium, magnesium, sodium, and boron, were at sufficient levels across growing seasons (data not shown). Pest, weed, and disease management practices were implemented in alignment with the University of Nebraska’s recommendations for winter wheat, ensuring that these pressures did not limit the study. Detailed pesticide application information is available in Supplementary Table S1.

An AccuPAR LP-80 ceptometer (Decagon Devices, Pullman, WA) was used to measure leaf area index (LAI) non-destructively at Feekes GS 4, 7, 10, and 10.5. For each plot, four measurements were taken at the four corners of the plot to ensure a representative area of the wheat canopy. Grain yield, moisture content, test weight, and protein were determined at harvest with a Hege 140 self-propelled small plot combine (Wintersteiger, Salt Lake City, UT). Grain protein content was measured using near-infrared reflectance spectroscopy with a Perten DA 7250 (Perten Instruments Inc., Springfield, Illinois).

The Agricultural Production Systems Simulator Next Generation (APSIM-NG; Holzworth et al., 2018) is a comprehensive software suite that includes a variety of crop modules (Keating et al., 2003), such as the APSIM-Wheat (Zheng et al., 2015). In our study, this model was employed to simulate the growth and development of winter wheat. The APSIM-Wheat model has the potential to provide a detailed simulation of phenology, biomass, and grain yield, capturing the interactions between soil properties, climatic factors, and management (Holzworth et al., 2014). APSIM-Wheat conducts crop growth and development simulations on a daily time-step. It integrates environmental variables such as solar radiation, temperature, soil moisture, and soil N content alongside agronomic practices to predict growth outcomes (Zheng et al., 2015). Simulated soil moisture values at 30, 60, and 90 cm correspond to layer-weighted averages of APSIM-NG soil water predictions across the 0-30, 30-60, and 60–90 cm depth intervals, respectively. The model delineates the wheat life cycle into eleven distinct developmental phases divided by thermal time accumulation and influenced by vernalization, photoperiod, and N availability from emergence to the terminal spikelet phase. These phases are configured within the model configuration, with thermal time (tt) controlling the progression between phases (Holzworth et al., 2014). The phenological transitions simulated by APSIM-Wheat (e.g., stem elongation, head emergence, flowering, and grain filling) correspond to Feekes GS 4, 6, 10, 10.5, and 11.2 used for field sampling in this study, allowing direct alignment between observed Feekes stages and APSIM’s thermal-time-based developmental phases (Holzworth et al., 2014; Zheng et al., 2015).

A vital aspect of the APSIM-Wheat model is its calculation of daily biomass accumulation derived from the plant’s ability to intercept solar radiation. It includes calculations involving intercepted radiation (I), radiation use efficiency (RUE), diffuse factor (fd), stress factors (fs), and a carbon dioxide response factor (fc). The model calculates the amount of radiation intercepted using the leaf area index (LAI) and the extinction coefficient (k), according to principles described by Monsi and Saeki (2005). More descriptions of the wheat model can be found in the technical documentation (Zheng et al., 2015).

In addition, we utilized the following modules in APSIM-NG version 7270: Soil N for soil N and C cycling (Probert et al., 1998), SoilWat for soil water modeling, including fluctuating shallow groundwater tables and tile drainage (Probert et al., 1998; Cichota et al., 2021), SURFACEOM for residue management (Probert et al., 1998; Thorburn et al., 2001), soiltemp2 for simulating soil temperature dynamics (Campbell, 1985) and management rules for rotation, tillage, planting, fertilizer, and harvesting. The Soil N module was not calibrated because no independent measurements of soil C or N turnover were available; therefore, site-specific soil properties were used in conjunction with APSIM-NG’s default decomposition and mineralization parameters. Details about the calibrated soil profile characteristics by depth, including parameters such as water table depth, bare soil runoff curve number, hydraulic conductivity at the drained upper limit (DUL), and matric potential at DUL, are provided in Supplementary Table S2. In addition, soil hydrologic parameters included in the SoilWater module are provided in Supplementary Table S3.

Each simulation was initiated on January 1, 2014, with a spin-up period to ensure the fast-decomposing organic matter pools stabilized and reached equilibrium before the primary analysis began (Zhao et al., 2014). The simulation was conducted continuously to capture year-to-year carry-over effects, with daily time step outputs generated throughout the process. Soil organic matter values were initially derived from baseline measurements and SSURGO (SSURGO, 2024) up to a depth of 2.44 m for each growing season. Saturated hydraulic conductivity for each layer was estimated based on SOM and texture (Saxton and Rawls, 2006).

The Mesonet weather station was 1.8 km from the experimental site. Simulated management practices inputs were adjusted annually to reflect the actual practices within the winter wheat experiment, which was part of a corn-soybean-wheat rotation, with wheat typically following soybean. These adjustments included winter wheat planting dates, N fertilizer application timing and rates, and tillage (Supplementary Table S1). Both the soybean and corn phases of the rotation were also modeled using APSIM’s default cultivar settings for soybean (maturity group 2) and corn (100-day variety), as outlined in Supplementary Table S4.

The model calibration was carried out in the R environment (version 4.4, R Core Team, 2024) using the R package “apsimx” (version 2.7.72; Miguez, 2024). The calibration approach sought to minimize the objective function representing the deviation of the simulated data from the observed data (Equation 1; Wallach et al., 2001):

Where “Y_ijk” and “Y^*_ijk” respectively represent the observed and modeled values for the i^th data category (e.g., biomass, grain yield) at the j^th measurement and for the k^th replication. More specifically, the categories of observed data available for calibration were biomass, grain yield, leaf area index, N content by plant organ, soil moisture, and date of occurrence of emergence, flowering, maturity, and harvest. These APSIM-simulated stages correspond to the Feekes stages used for field sampling (GS 4 = stem elongation onset, GS 6 = jointing, GS 10-10.5 = boot to heading/flowering, GS 11.2 = grain filling), ensuring alignment between observed and simulated phenology. The inverse of the variance (1/ σi2) was used to weigh the errors for different data categories (Wallach et al., 2001).

The model calibration was carried out in a two-step procedure, first calibrating parameters related to phenology and then parameters related to biomass production and yield (Wallach et al., 2018; Archontoulis et al., 2020). The calibration of crop phenology was carried out by running an initial grid search over a wide range of values (Table 1) and evaluating the objective function (Equation 1) concerning the phenology data. The grid search explored a factorial combination of five equally spaced values within the range indicated in Table 1 of the parameters that control phenology, resulting in approximately 900,000 simulations. The parameters in Table 1 correspond to cultivar- and phenology-related settings in the APSIM-NG Wheat module (Holzworth et al., 2014; Zheng et al., 2015). The default parameter values reported in Table 1 are those provided within the APSIM-NG Wheat module used in this study. APSIM-NG undergoes continuous development, and default parameter values and model structure may vary across software versions. The parameter ranges used in the grid search were defined based on APSIM documentation and previously published model calibration studies (Wallach et al., 2001; Archontoulis et al., 2020). Subsequently, using the Nelder-Mead algorithm, the parameter value combination that minimized the objective function was used as starting values for unconstrained optimization (Nelder and Mead, 1965). Likewise, the calibration of the parameters that control growth and yield followed a procedure similar to phenology calibration. A grid search was conducted using five equally spaced values indicated in Table 1, resulting in 125 simulations. This smaller number was due to the calibration involving only three parameters, each with a narrower search range, reducing computational demands while maintaining optimization efficiency. For every simulation, the objective function (Equation 1) was evaluated concerning the biomass and yield data, and the parameter value combination that minimized the objective function was selected as starting values for the Nelder-Mead optimization (Nelder and Mead, 1965).

The experimental design enabled the evaluation of two winter wheat cultivars, LCS and WB, across two growing seasons (2020/21 and 2021/22) and four N rates. This approach generated a robust dataset to assess cultivar performance across varying environmental and management conditions and served as the basis for subsequent model calibration. Phenological observations indicated that both cultivars reached key growth stages at similar times across seasons, with no consistent differences (Supplementary Table S6).

Observations showed significant variability in shoot biomass, LAI, and grain yield across cultivars, N rates, and years. Grain yield was significantly affected by N rate in both years (p < 0.001), while cultivar effects were not significant in 2020/21 (p = 0.924) but became significant in 2021/22 (p < 0.05). No significant cultivar × N interaction was observed in either year (Supplementary Table S5).

On average, grain yield was 212.5 kg ha^-1 higher for LCS compared to WB, with observed means of 4561.0 and 4348.5 kg ha^-1, respectively (Table 2). Shoot biomass showed a similar pattern, with N rate showing a consistent effect across years (p < 0.01), and cultivar differences becoming significant only in 2021/22 (p < 0.01), as shown in Supplementary Table S5. At the final sampling date, shoot biomass averaged 5821.3 kg ha^-1 for LCS and 6093.8 kg ha^-1 for WB.

LAI followed a comparable trend, with significant effects of N rate in both years (p < 0.001). Cultivar effects were only significant in 2021/22 (p < 0.001), as shown in Supplementary Table S5. LAI averaged 2.76 for LCS and 2.68 for WB.

Observed winter wheat yield response to N application varied between cultivars (LCS and WB) and growing seasons (2020/21 and 2021/22; Figure 1). In the 2020/21 growing season, the observed EONR was similar between cultivars at approximately 100 kg N ha^-1 (11 kg N ha^-1 difference). However, the observed YEONR was slightly higher for LCS than WB (6565 kg ha^-1 and 6147 kg ha^-1, respectively; Figure 1). In contrast, the YN0 was lower for LCS than WB (3736 kg ha^-1 and 4852 kg ha^-1, respectively; Figure 1). For 2021/22, the observed EONR was lower for LCS compared to WB (126 kg N ha^-1 and 149 kg N ha^-1, respectively; Figure 1), with a 23 kg N ha^-1 difference. However, the YEONR (3893 kg ha^-1 and 3605 kg ha^-1, respectively) and YN0 (2580 kg ha^-1 and 2247 kg ha^-1, respectively) were slightly higher for LCS compared to the WB (Figure 1).

Calibration improved the model’s performance across all crop and soil parameters analyzed, reducing prediction errors and improving agreement between simulated and observed values (Figure 2). For phenology, the model reproduced key growth stages with high accuracy across years and cultivars. Average RMSE values were approximately 5 days in 2020/21 and 4 days in 2021/22, with corresponding RRMSE values of about 2.6-2.7% and 1.5-1.7%, respectively (Figure 2A, Supplementary Table S7). This performance reflects the effective parameterization of cultivar-specific thermal time and photoperiod sensitivity.

Simulated grain yield showed well to moderate agreement following calibration. The model captured yield trends across years and cultivars, with RRMSE values of 15% for LCS and 24% for WB (Figure 2D), and simulated cultivar differences consistent with observed means (Table 2). Yield predictions under high N application closely matched observations in 2021/22, although slight underestimation occurred for WB in 2020/21.

Shoot biomass predictions improved moderately after calibration, particularly in 2021/22. The model achieved RRMSE values of 38% for LCS and 46% for WB, with corresponding RMSEs of 1487 kg ha^-1 and 1961 kg ha^-1 (Figure 3B). Despite these improvements, agreement remained moderate to poor, indicating persistent challenges in capturing biomass dynamics under varying environmental conditions. Simulated aboveground biomass tended to be slightly overestimated for both cultivars (Table 2), and leaf and stem component predictions exhibited high variability (Supplementary Figure S1).

For LAI, calibration led to moderate improvements in model performance, but overall agreement remained poor. In 2020/21, RRMSE values were 63% for LCS and 75% for WB, while in 2021/22, agreement declined further to 75% and 87%, respectively (Figure 2C, Figure 3A). The model tended to underestimate peak canopy development, particularly in the later growing season, suggesting limitations in simulating cultivar-specific canopy dynamics or potential uncertainty in field measurements (Supplementary Figure S2A).

The high RRMSE values for LAI and biomass were primarily due to early-season measurements with very small observed values, which inflated the percent error despite small absolute differences. As shown in Figure 3, APSIM-NG captured canopy development well in 2020/21 but underestimated LAI in 2021/22 following a hail event that reduced leaf area in the field and was not represented in the model. These discrepancies did not affect soil moisture or nitrogen simulations, which are computed mechanistically by SoilWat and Soil N and are not driven by short-term variation in LAI or biomass.

Grain protein content was simulated with well agreement across both growing seasons and cultivars. After calibration, the model achieved RRMSE values below 10% for LCS and between 8% and 13% for WB (Figure 2E), indicating highly accurate predictions of protein concentration. These results reflect the model’s capacity to capture cultivar-specific N dynamics during the reproductive period. Additional details on organ-level N content, including leaves and stems, are provided in Supplementary Figure S3.

The APSIM-NG demonstrated variability in crop yield response to N, with model performance ranging from poor to well across treatments and showing clear differentiation between zero (N0) and high (N168) N rates for both cultivars (Supplementary Table S8). Under the 168 kg N ha^-1 treatment, the model performed well for LCS (RRMSE = 11%) and WB (RRMSE = 16%; Supplementary Table S8). However, under the 0 kg N ha^-1 treatment, performance was moderate for LCS (RRMSE = 30%) and poor for WB (RRMSE = 46%). This indicates that the model achieved better performance at higher N rates.

On average, the model predicted yield response to N well for LCS with a RRMSE for EONR of 18.9%. In contrast, the EONR agreement for WB was moderate, with an average RRMSE of 32.2% (Table 3), likely due to differences in phenology and N uptake timing not fully captured by the model. The model showed well agreement for YEONR across both cultivars and growing seasons, with RRMSE values below 10% (Table 3). For YN0, the model performed well in both growing seasons (2020/21 and 2021/22) for the LCS (RRMSE = 14.9% and RRMSE = 17.9%, respectively), while a moderate agreement (RRMSE = 38.5% and RRMSE = 31.8%, respectively) in both growing seasons for WB (Table 3).

For N uptake, APSIM-NG performance ranked Stem N< Leaves N < N grain, from 54% (stem) to 12% (grain), with model accuracy improving as N moved from structural to reproductive tissues (Supplementary Figure S3). This pattern highlights organ-specific differences in model performance. For instance, N uptake in the leaves showed moderate agreement for LCS (RRMSE = 27%) and WB (RRMSE = 33%; Supplementary Figure S3A). Stem N uptake has shown poor agreement for both cultivars (RRMSE = 51% for LCS and RRMSE = 58% for WB; Supplementary Figure S3B). In contrast, grain N uptake performed well for both cultivars (RRMSE = 11% for LCS and RRMSE = 13% for WB; Supplementary Figure S3C).

Soil NO₃^- and NH₄⁺ measurements were used to establish initial soil N conditions during model calibration in APSIM-NG. At the start of the calibration, measured soil NO₃^- values included 7.13 kg ha^-1 and 3.7 kg ha at 20–40 cm. For NH₄⁺, the model simulated 0.53 kg ha^-1 in the 0–20 cm depth and 0.05 kg ha^-1 in the 20–40 cm depth, providing a baseline for simulating N dynamics. During the first paired observed-simulated comparison, the model closely approximated observed values, with simulated NO₃^- of 2.4 kg ha^-1 in the 0–20 cm depth versus 2.7 kg ha^-1 observed, and simulated NH₄⁺ at 0.9 kg ha^-1 versus 0.85 kg ha^-1 observed (Supplementary Figures S2C, E). Beyond initialization, the model captured general seasonal trends in soil NO₃^- and NH₄⁺, including early-season increases and late-season depletion. However, it underestimated the magnitude of observed peaks, particularly in the 0–20 cm depth, and failed to reproduce full-season dynamics at 20–40 cm (Supplementary Figures S2C, E). However, due to the inherent variability and limited spatial and temporal resolution of soil sampling data, it is challenging to capture the full dynamics of soil N accurately. In contrast, crop performance metrics, such as yield and N uptake, integrate the cumulative effects of soil and environmental conditions over time, making them more reliable and comprehensive indicators of model accuracy.

The APSIM-NG winter wheat model captured yield variability across Nebraska environments with reasonable accuracy (Figure 4). For LCS, the model achieved an average error of 14% (RRMSE), with an overall RMSE of 750 kg ha^-1 and minimal bias (MBE = -38 kg ha^-1). The best predictions occurred in high-precipitation environments (> 550 mm cumulative rainfall), particularly in Saunders County in 2022, where error dropped below 1% (RRMSE = 0.9%). Nearly all LCS validation site-years (12 out of 13) were also classified as high precipitation, and accuracy across these cases ranged widely, from well agreement (e.g., Saunders 2022) to moderate agreement (e.g., Lancaster 2021; RRMSE = 25%). The single medium-precipitation case (Saunders 2018) did not yield improved performance (RRMSE = 23.3%; moderate agreement), although overall performance remained acceptable. Notably, model calibration was performed in medium-precipitation seasons (526 mm in 2020/21 and 381 mm in 2021/22; Supplementary Table S1), yet the best validation performance occurred in high-precipitation environments, indicating that accuracy was influenced by factors beyond precipitation regime alone.

For WB, the model showed greater variability, with a mean RRMSE of 19%, an overall RMSE of 1302 kg ha^-1, and a systematic underestimation of yields (MBE = -1068 kg ha^-1). As with LCS, the best-performing WB site-year occurred in a high-precipitation environment (>550 mm), such as Saunders in 2019 (RRMSE = 5%). Moderate errors were observed across medium-precipitation environments (e.g., Jefferson 2020; RRMSE = 22%), while some high-precipitation cases produced the largest errors (e.g., Saunders 2021; RRMSE = 29%).

Across all validation sites, the model performed well across precipitation regimes, with most county-environment combinations falling within the well-agreement range (16 out of 24 site-years; RRMSE < 20%). Moderate errors occurred in only a subset of cases (8 out of 25 site-years; 20-24% RRMSE), and no environment exceeded the poor-agreement threshold (>40% RRMSE). These results demonstrate that APSIM-NG can reliably reproduce winter wheat yield responses across diverse Nebraska environments with acceptable uncertainty, highlighting precipitation and site-specific factors as key drivers of prediction accuracy.

The APSIM-NG model effectively simulated key winter wheat processes, including phenology, grain yield, and grain protein content, with high accuracy following calibration. Calibrated RMSE for phenology was reduced to 4–5 days and RRMSE to < 2.2%, demonstrating the model’s strength in capturing cultivar development timing. These improvements are critical, as accurate phenology drives management decisions such as N application timing (Holzworth et al., 2018; Hu et al., 2021).

Trait-based calibration improved predictions across traits, especially grain yield and grain protein. For example, grain yield RRMSE declined from 33% to 15% for LCS and 29% to 24% for WB, while grain protein RRMSE decreased from 9.7% to 8.1% for LCS and 23% to 11% for WB. These gains reflect practical tuning of cultivar-specific parameters such as Phyllochron, GrainNumber, and MaximumNConc, consistent with prior studies on genotype calibration (e.g., Asseng et al., 2002; Kumar et al., 2023).

To enhance model transferability, we calibrated APSIM-NG using all available calibration years simultaneously, rather than treating each year independently. This approach allowed cultivar parameters to reflect stable physiological characteristics, while site-specific conditions were handled through weather and soil inputs. The model maintained reasonable accuracy across diverse conditions by prioritizing system-level fidelity over per-season fit, increasing confidence for broader applications. These results demonstrate the importance of integrating field-level management and environmental data into the model configuration. By using observed planting dates, N application timing, and cultivar-specific parameter sets, APSIM-NG could replicate phenological development and final grain yield across site-years. This alignment between observed management practices and simulation inputs is critical for models to generalize across years, particularly when environmental drivers (e.g., precipitation timing) are highly variable (Holzworth et al., 2014; Keating et al., 2003; Wallach et al., 2001). The ability to simulate year-to-year variation while maintaining cultivar consistency highlights the calibration strategy’s robustness. Yield in APSIM is determined by grain number and grain weight, rather than a fixed harvest index, allowing the model to compensate for variability in biomass allocation (Sadras and Lawson, 2011; Holzworth et al., 2018). This aligns with previous studies showing that water stress affects N dynamics and yield predictions (Brisson et al., 2002; Lobell and Asseng, 2017).

Our calibration of the APSIM-NG model for winter wheat in the Great Plains demonstrates its potential as a decision-support tool for CSA. Rather than indicating uniformly accurate performance, the model reproduced the overall yield response to N rates across cultivars, with higher accuracy at high N rates and reduced accuracy under zero-N conditions. The model still supports EONR estimation under varying environmental conditions, though uncertainty is greater when soil N availability is low. Such capability can reduce over-fertilization and greenhouse gas emissions (e.g., N₂Wang et al., 2023), supporting the CSA goals of improving productivity and resilience while lowering environmental impact (Lipper et al., 2014). The model’s ability to represent cultivar-specific traits also supports breeding and selection of genotypes better adapted to changing climates (Bai et al., 2022).

Shoot biomass and LAI predictions remained moderate to poor and were highly sensitive to seasonal variability and cultivar dynamics. In 2020/21, biomass was often underestimated, likely due to early-season N stress and delayed rainfall following spring fertilization (Loecke et al., 2004). Conversely, 2021/22 showed biomass overestimation, potentially related to hail damage and lower total precipitation. These results indicate the model’s sensitivity to environmental drivers, particularly water and N interactions, but highlight its limitations in simulating real-world stress events not represented in the model. This aligns with Zhao et al. (2014), who noted APSIM’s difficulty representing biomass under fluctuating N and water stress.

Biomass simulation challenges were partly due to limited translocation modeling during grain filling. Due to a simplified representation of remobilization processes, the model tends to overestimate vegetative biomass, particularly stems and leaves. This was evident in 2021/22, where overestimated biomass did not translate into overestimated yield. This mismatch supports prior findings that APSIM and other models may misrepresent sink-source dynamics during reproductive stages (e.g., Martre et al., 2003; Dreccer et al., 2009; Sinclair and Jamieson, 2006). Previous studies suggest that biomass is more sensitive to short-term environmental variation, while yield is often buffered by physiological mechanisms that stabilize grain production under moderate stress (Sinclair and Muchow, 1999; Asseng et al., 2002). This distinction reflects the physiological stability of grain sink strength under stress, which buffers yield even when canopy growth is variable (Goudriaan and Van Laar, 1994; Reynolds et al., 2005).

LAI predictions were similarly poor (RRMSE > 60%), with a consistent underestimation of peak canopy development. This could reflect both model structure and field measurement limitations. Ceptometer data may include noise, and APSIM’s treatment of LAI has not been as rigorously validated as its grain yield predictions (Ahmed et al., 2016; Pokovai and Fodor, 2019). LAI showed minimal value for yield prediction in this study, supporting prior work that stresses temporal canopy dynamics over peak values (Waldner et al., 2019; Thompson et al., 2024). Additionally, cultivar × environment interactions and canopy plasticity may contribute to LAI variability, posing challenges for general parameterization. These findings are consistent with prior evaluations of APSIM and similar models, which often perform well for yield but struggle with dynamic traits like LAI and biomass due to limitations in water-N interactions and structural assumptions (Hochman et al., 2009; Archontoulis et al., 2014; Asseng et al., 2015).

Lastly, poor prediction of NO₃^- and NH₄⁺ (Supplementary Figures S2C, E; RRMSE = 160.3% for NO₃^- and RRMSE = 131.21% for NH₄⁺) could affect the EONR uncertainty due to soil and crop dynamics (Archontoulis et al., 2020). These limitations reinforce the need to improve how APSIM-NG simulates yield response to N.

One possible reason for poor biomass and LAI predictions is the lack of explicit root growth and plasticity simulation. Root architecture, rooting depth, and water/nutrient uptake efficiency vary across cultivars and environments but are simplified in current APSIM configurations. Similarly, traits like tillering capacity and stay-green phenotype, which affect canopy dynamics, are not yet genotype-specific in APSIM-NG. These physiological features influence N uptake and canopy development but are often averaged in model parameterizations. Including cultivar-level physiological plasticity could improve LAI and biomass simulations, especially under variable N or water availability (MansChadi et al., 2006; Chenu et al., 2011; Giordano et al., 2024).

Low accuracy on soil N dynamics (NO₃^- and NH₄⁺) suggests focusing future improvements on refining the simulation of soil N transformations, including mineralization, immobilization, and leaching, particularly under fluctuating moisture conditions.

After calibration, the model’s performance improved significantly, as did EONR and yield response to N, YEONR, and YN0. In addition, the calibrated model generally underestimated EONR for both cultivars and growing seasons, with variability in performance across years.

While calibration of the APSIM-NG model improved its performance in simulating yield at EONR, some challenges still limit its predictive ability. A key issue is that the EONR estimation relies on small yield increments near the maximum N response, where even minor yield prediction errors can cause large deviations in EONR (Puntel et al., 2016). For example, Baum et al. (2023) found that a 10% error in yield predictions for corn can lead to a 34% error in EONR. Moreover, the dynamic nature of the N cycle, as influenced by management, soil properties, and weather interactions, exacerbates prediction errors, especially at lower N rates (Morris et al., 2018; Puntel et al., 2018; Correndo et al., 2021). Limited field data, particularly for low N, also hinders precise EONR estimation, as evidenced in the confidence intervals in Figure 4 (Miguez and Poffenbarger, 2022). This uncertainty is particularly true when the model fails to accurately capture the quadratic yield response to N, a critical pattern for EONR prediction (Baum et al., 2023). After calibration, the model performed well in simulating yields under zero N (YN0), particularly for WB, which showed consistently better agreement across both seasons. This suggests the model captures cultivar-specific responses under low N conditions, likely reflecting physiological differences between the cultivars. Similar trends have been noted in previous studies, where cultivar-specific traits such as N use efficiency and root architecture influenced yield predictions under limited N availability (Brisson et al., 2002; Hammer et al., 2010). These findings highlight the importance of cultivar-specific calibration (Asseng et al., 2015; Zhao et al., 2014) and the need to tailor N management strategies to differences in growth and uptake (Mirosavljević et al., 2024; MansChadi et al., 2006; Sadras and Lawson, 2011).

While we expected the model to capture N rate × environment × cultivar interactions, accurately simulating soil N remains challenging due to the complexity of N transformations under variable conditions (Butterbach-Bahl et al., 2011; Davidson and Kanter, 2014). This reinforces the need for improved soil inputs and root/N uptake calibration (Gabrielle et al., 2006; Zhang et al., 2015).

For example, real-time nitrate sensing technologies (Jiang et al., 2019; Baumbauer et al., 2022) could enhance model responsiveness to in-season conditions. These sensors allow dynamic soil N adjustment, improving realism over static pre-season inputs, similar to past improvements with soil moisture sensing (Patrignani and Ochsner, 2018). While other factors, such as water or phenology, may be better captured, poor soil N simulation limits APSIM’s utility in nutrient-focused applications.

A promising approach is integrating APSIM-NG with real-time sensors and remote sensing tools (Thompson et al., 2024), allowing dynamic simulation of N availability and crop response. Such tools could adjust N uptake in real time and improve in-season prediction accuracy (Asseng et al., 2002). This would expand APSIM-NG’s utility for precision N management and sustainability. Improving soil N and mineralization predictions could also enhance EONR estimation (Nowatzke et al., 2022; Thompson et al., 2024).