AI-Ready Benchmark Dataset

• AI-Ready Benchmark Dataset is a standardized, quality-controlled collection that integrates simulated data and observations for reproducible machine learning research.
• It harmonizes multi-source inputs like physics-based simulations and flux tower measurements, ensuring consistent spatial-temporal coverage and rigorous preprocessing.
• The dataset supports advanced modeling techniques including transfer learning and adversarial domain alignment, improving metrics such as R² and RMSE in environmental flux prediction.

An AI-ready benchmark dataset is an integrated, standardized, and quality-controlled collection of data specifically curated to facilitate the development, evaluation, and comparison of AI methodologies with respect to defined scientific tasks. Such datasets address limitations inherent in conventional approaches—such as data sparsity, heterogeneity, and process complexity—by providing structured protocols, metadata, and rigorous splits that enable reproducible machine learning workflows. The AgroFlux benchmark exemplifies an AI-ready dataset tailored for spatial-temporal modeling of greenhouse gas (GHG) fluxes within agroecosystems, combining physics-based simulations and observational data to support the trustworthy evaluation of sequential deep learning models and transfer learning schemes in this domain (Cheng et al., 2 Feb 2026).

1. Data Sources, Integration, and Protocol Architecture

The foundation of the AgroFlux AI-ready benchmark lies in harmonizing multi-source data representing key carbon and nitrogen fluxes:

• Physics-based Simulations:
  • Ecosys (Grant, 2001; Zhou et al., 2021): Daily mechanistic simulation, ~200 driver variables (weather, soil, management), generating outputs in carbon, nitrogen, water, and thermal domains.
  • DayCent (Del Grosso et al., 2001; EPA, 2021): Similar scope and variable schema, with simulation outputs not exactly matching Ecosys but adhering to analogous spatiotemporal structure.
• Observational Data:
  • Eddy-covariance flux towers: 11 sites across the US Midwest (IL/IA/MI/NE/MN), 2000–2020, with daily CO₂ flux and GPP readings, coincident meteorological and soil measurements, and discrete crop management metadata.
  • Controlled-environment chambers: Single corn–soy rotation site, 2016–2018, measuring N₂O flux, soil mineral nitrogen, moisture profiles, and precipitation treatments on an hourly-to-daily basis.

The integration protocol enforces a consistent driver schema $X \in \mathbb{R}^{n \times T \times D_{in}}$ across all sources. Preprocessing includes 1st–99th percentile clipping, standardization, annual segmentation (365-day sequences), and boolean masking of missing observations. Master “ML-ready” archives are distributed as CSV or NumPy, with CF-compliant metadata (JSON), and complete folder structures (/Ecosys/, /DayCent/, /Observations/) ensure data provenance and auditability.

2. Spatial, Temporal, and Variable Coverage

Spatially, Ecosys simulations cover 99 counties in Iowa, Illinois, and Indiana from 2000–2018, while DayCent offers daily records for 2,562 Midwest sites, each under 42 distinct management regimes. Observational towers provide 5–19 years of coverage per site; six soil chambers contribute experimental N₂O fluxes (2016–2018).

Temporally, all simulations and observations operate at daily resolution, formatted into 365-day annual sequences. Variables—drivers and outputs—are meticulously documented. Input “driver” variables ($D_{in}=12$) enumerate meteorological (TMAX, TMIN, PREC, RADN, HUMIDITY, WIND), edaphic (TBKDS, TCSAND, TCSILT, TPH, TSOC), and management features (FERTZR_N, PDOY, PLANTT). Outputs ($D_{out}$ variable by data source/model) include:

• Carbon: Reco, NEE ($\mathrm{NEE} = F_{auto} + F_{het} + F_{ch4}$), GPP, yield, $\Delta$SOC, LAI.
• Nitrogen: N₂O flux, $[\mathrm{NH}_4^+]$, $[\mathrm{NO}_3^-]$ at depth.
• Water/Thermal: SWC, ET, soil temperatures.
• Observations: Direct CO₂, GPP, N₂O fluxes consistent with modeled units.

Data access supports CSV, NumPy, and NetCDF (optional), with all splits and folders preconfigured for standardized ML experimentation. Metadata adheres to consistent conventions: units, descriptions, coordinates, and time-bounds.

3. Processing, Quality Control, and Benchmark Curation

Outlier clipping at the 1st/99th percentiles and Z-score normalization (training split statistics) precede the annual sequence chopping. Missing values are strictly masked in both training and evaluation via indicator masks $M \in \{0,1\}$. No explicit spatial interpolation is conducted beyond the assigned county/site resolution. All curated scenarios are accompanied by protocol sheets detailing data origins, driver-output mappings, and masking policies.

Benchmark splits utilize five-fold spatial cross-validation with fixed random seeds. Simulation and observational datasets are merged via the unified driver schema and calendar alignment, supporting the comparison of algorithms on realistic generalization tasks (temporal and spatial extrapolation).

4. Baseline Modeling, Loss Functions, and Evaluation Metrics

The benchmark includes sequential deep-learning baselines trained over sequence inputs $X \in \mathbb{R}^{B \times L \times D_{in}}$ to predict outputs $Y \in \mathbb{R}^{B \times L \times D_{out}}$ ($D_{in}=12$0). Models comprise:

• LSTM / EA-LSTM: Stacked 3-layer, hidden size 50, dropout 0.2; EA-LSTM incorporates input feature attention per timestep.
• Temporal CNN (TCN): 3 residual blocks, kernel=5, dilations=[1,2,4], hidden=50.
• Transformer variants: Standard (3 encoder+1 decoder, 5 heads, model dim 50); iTransformer (feature-attention), Pyraformer (pyramid attention for multi-resolution).

Optimization uses masked mean-squared error (MSE) loss, Adam optimizer (lr=1e–3, batch=256, iTransformer batch=10), and convergence on validation R². The loss and metrics are:

• Masked MSE:

$D_{in}=12$1

• Testing Metrics:
  • Root mean squared error (RMSE):

  $D_{in}=12$2 - Mean absolute error (MAE):

  $D_{in}=12$3 - Coefficient of determination (R²):

  $D_{in}=12$4

Performance results, reflecting the difficulty gradient from simulated to real-world data, illustrate robust predictive performance for temporal and spatial extrapolation in simulation (Ecosys CO₂ flux, R² = 0.938–0.953), with limited generalization for real-world observations (tower CO₂ R² = 0.784 temporal, 0.563 spatial; N₂O R² = 0.433–0.883 depending on metric/task).

5. Transfer Learning Schemes and Domain Alignment

Transfer learning protocols are formalized to exploit large simulated datasets (Ecosys or DayCent) for pretraining, followed by fine-tuning on observational targets:

1. Pretrain on simulated data (temporal/spatial cross-validation replicated).
2. Initialize observational “student” model with pretrained weights.
3. Fine-tune using observational train/validation splittings, masking missing data.
4. Evaluate on test splits (temporal and spatial).

An adversarial domain-alignment alternative introduces a discriminator $D_{in}=12$5 on hidden features $D_{in}=12$6 to predict “simulation vs. observation,” optimized by minimizing prediction loss $D_{in}=12$7 and maximizing discriminator loss $D_{in}=12$8 via a gradient-reversal layer.

Quantitative improvements from transfer learning are substantial; for example, LSTM performance on CO₂ spatial (observations) increases from R² = 0.339 (scratch) to 0.648 (pretrain-finetune, DayCent) and 0.666 (adversarial). For GPP, R² advances from 0.503 (scratch) to 0.750 (pretrain-finetune) and 0.781 (adversarial). Similarly, N₂O spatial prediction by Pyraformer can benefit modestly (R² from 0.754 to 0.824 with adversarial alignment). Full RMSE/MAE tables are cataloged in supplementary materials III–IV and IX–XVI (Cheng et al., 2 Feb 2026).

6. Reproducibility, Protocol Transparency, and Data Accessibility

The AgroFlux AI-ready benchmark is structured to maximize transparency and reproducibility. Comprehensive documentation includes protocol sheets, data schemas, random seed usage, and fold lists, permitting direct reproduction of baseline and transfer learning workflows. All datasets are released under the CC-BY license via a HuggingFace dataset card (details pending acceptance), using CF-compliant metadata conventions for interoperability.

A plausible implication is that widespread adoption of such AI-ready datasets, with rigorously harmonized protocols and evaluation frameworks, could significantly accelerate methodological innovation and evidence-based model comparison in GHG flux prediction and broader environmental modeling contexts (Cheng et al., 2 Feb 2026).