Cropland is trapped outside a safe and agronomically viable nitrogen use

Supplementary data for the manuscript "Cropland is trapped outside a safe and agronomically viable nitrogen use".CSV files Datasets.zipHistorical (1961-2022) and SSP-RCP scenario dataset (2025-2100 at 5-yr intervals) used for the ML pipeline. Both dataset contains the share of functional climate in each country, population, GDP per capita. Historical dataset contains crop composition (share of different crop groups in total cropland), total N inputs, N yields, critical N inputs and viable N yield (N_att).Scenarios_operating_space_no_aug.csv Data generated for the 17 million diagnostic scenarios for different RCP-SSP combinations. This includes crop composition, N inputs, N yields, viable yields as well as the corresponding operating spaces.Model Files (.pth)static_cvae_model.pthTrained PyTorch model containing the Static Conditional Variational Autoencoder (CVAE) for crop composition generation. This model generates single-year crop compositions conditioned on climate (Arid, Temperate, Cold, Tropical zones) and socioeconomic variables (GDP per capita, population, year). The model was trained on historical data (1961-2019) with 8 crop categories and includes:Model architecture: Encoder-decoder with 32-dim latent space, 128-dim hidden layersCountry embeddings for country-specific patternsAnti-collapse mechanisms (dropout, condition bottleneck, capacity annealing)Scalers for input normalizationConfiguration metadata (version 1.3.0)temporal_cvae_model.pthTrained PyTorch model containing the Temporal CVAE for generating realistic crop composition trajectories over time (2025-2050). This model uses LSTM layers to capture temporal dependencies and agricultural transition dynamics. It generates 25-year sequences of crop compositions while maintaining temporal consistency. Includes:LSTM-based encoder-decoder architecture (64-dim latent space, 128-dim hidden layers)Multi-head temporal attention mechanismCountry embeddings and condition bottleneckingTemporal smoothness constraintsTrained on 5-year historical sequences to predict future trajectoriesMachine Learning Model Files (.pkl)rf_multioutput_model_no_aug.pkl Random Forest multi-output regression model trained without data augmentation. This is a baseline model for predicting multiple agricultural outputs simultaneously. The large file size indicates it contains:Multiple decision trees (ensemble model)Feature importance rankingsTrained on original historical data without synthetic augmentationUsed for comparison against augmented model variantstemporal_trajectories.pkl Pickle file containing generated temporal crop composition trajectories for all countries and SSP-RCP scenarios (2025-2050). This file stores:Multi-dimensional arrays of crop compositions over timeDimensions: [countries × scenarios × years × crops]Generated using the Temporal CVAE modelIncludes 25-year projections (2025-2050) for SSP1-5 and RCP2.6-8.5 combinationsStochastic samples capturing uncertainty in future trajectoriesssp_rcp_trajectories.pkl Pickle file containing SSP-RCP scenario-specific crop trajectories. Similar structure to temporal_trajectories.pkl but organized by Shared Socioeconomic Pathways (SSP1-5) and Representative Concentration Pathways (RCP2.6, 4.5, 6.0, 8.5). Contains:Country-level crop composition projectionsClimate and socioeconomic conditioning variablesMultiple stochastic realizations per scenarioMetadata linking scenarios to climate/socioeconomic assumptions