TRENDY-Emulator

This repository accompanies the MSc thesis "TRENDY-Emulator: A Bias-Corrected Deep Learning Emulator of Terrestrial Carbon and Water Dynamics", presented in fulfillment of the Global Forestry Erasmus Mundus MSc at AgroParisTech, Montpellier. It provides the full workflow used to build, train, and benchmark the TRENDY-Emulator.

The pipeline includes:

• preprocessing TRENDY forcing and DGVM outputs
• generating spatial masks
• constructing Zarr datasets for training and inference
• model training (Base → Stable → Transfer-Learned)
• scenario prediction and NetCDF export
• ILAMB benchmarking
• plotting and analysis

1. Preprocessing

Location: scripts/preprocessing/

Processes all TRENDY forcing and ancillary datasets.

Steps:

1. Run in any order

   • climate/ClimateProcessor.sh
   • co2/co2.py
   • jsbach/jsbach_processor.py
   • luh2/LUHProcessor.sh
   • ndep/ndep_processor.sh
   • nfert/nfert_processor.sh
   • potential_radiation/potential_radiation.sh
   • population/population.py
   • avh15c1.sh — processes LAI observations (after downloading via ILAMB).

2. rolling_means.sh — computes 30-year trailing means for selected climate variables.

3. preindustrial.sh — generates pre-industrial CO₂, land-use, and climate inputs for TRENDY scenarios.

4. model_outputs/models.sh - preprocesses all the individual DGVMs

5. model_outputs/ensmean.sh - takes the ensemble mean of DGVMs.

2. Masking

Location: scripts/masking/

1. nan_mask.sh — creates a mask of pixels where all forcing + output variables are finite.
2. land_mask.sh — selects pixels where CLM, ORCHIDEE, ELM & CLASSIC agree land fraction > 0.9; combined with the nan mask.
3. tvt_mask.sh — creates the longitudinal-band Train/Validation/Test split.

3. Make Zarrs (Training & Inference)

Training Zarrs

Location: scripts/make_zarrs/training/main/

1. Run make_training_tiles.sh — generates training, validation, and testing Zarrs.
2. Run consolidate.sh and finalize.sh.

Location: scripts/make_zarrs/training/other/

3. fill_potential_rad_nans.sh — fills missing potential radiation data in the testing Zarr.
4. add_avh15c1.sh — adds LAI observations for transfer learning.

Inference Zarrs

Location: scripts/make_zarrs/inference/

5. make_inference.sh — creates inference-ready Zarrs for all scenarios.
6. add_avh15c1.sh — inserts observed LAI for scenario 3.

4. Standardisation

Location: scripts/standardisation/

1. standardisation.sh — computes global means and standard deviations for each variable.
2. merge.py — assembles the full standardisation JSON.

5. Train the Emulator

Location: pipeline/1.train/

1. Run train.sh

Three emulator versions are produced:

• Base-Emulator — trained without autoregressive carrying
• Stable-Emulator — autoregressive carrying with progressively longer horizons
• TL-Emulator — transfer-learned on AVH15C1 LAI observations

These versions were produced by manually adjusting the condigurations in train.sh.

6. Generate Predictions

Location: pipeline/2.predict/

1. Run predict.sh

Outputs:

• Zarr prediction files
• (optional) NetCDF exports (export_nc=true)
• automatic copying into ILAMB-ready structure

7. Benchmark with ILAMB

Location: pipeline/3.benchmark/

Steps:

1. Download ILAMB benchmark datasets (see https://www.ilamb.org/doc/tutorial.html).
2. Place emulator output inside MODELS/.
3. Configure build.cfg (regions, confrontations, variables).
4. Run submit.sh.

8. Plotting and Analysis

Location: pipeline/4.Analyse/

Main manuscript figures are generated via:

• create_csvs/
• plot_csvs/

Additional analysis scripts are provided in the same directory.

Usage and Citation

This work is being prepared for publication, please contact the author before use.