Skip to content

Latest commit

 

History

History
107 lines (80 loc) · 5.68 KB

README.md

File metadata and controls

107 lines (80 loc) · 5.68 KB
Raw

JAX-CanVeg: A Differentiable Land Surface Model

fishy

Table of Contents

Overview

Land surface process describes the water, energy, and carbon cycles exchanged among the atmosphere, canopy, and soil. Its complex interacting nature makes it challenging to model due to the associated unknown biophysical and ecophysiological parameters and less-mechanistically represented subprocesses. Differentiable modeling provides a new opportunity to explore the parameter space and capture these complex interactions by seamlessly coupling process-based and deep learning models. Here, we developed a differentiable land surface model by reimplementing an existing simulator, CanVeg, in JAX -- a Google-developed Python package for high-performance machine learning research using automatic differentiation. Anchored in differentiable modeling, we expect that JAX-CanVeg provides a new avenue for modeling land-atmospheric interactions by leveraging the benefits of both data-driven learning and process-based modeling.

Repo Structure

.
+-- src/jax_canveg
+-- examples
+-- data
+-- environment.yml
+-- doc
+-- README.md
  • src/jax_canveg: providing source codes for JAX-CanVeg.
  • examples: providing example notebooks for running JAX-CanVeg at four selected flux tower sites.
  • data: providing the observation data (including both flux tower and MODIS) and the simulated lagrangian particles.
  • environment.yml: the YAML file for creating the conda virtual environment.
  • doc: providing a list of documentation files and figures.
  • README.md: the readme file.

Installation

Please follow the procedures of this post to install JAX-CanVeg.

Getting Started

We suggest training and running JAX-CanVeg by providing a JSON-based configuration file. See this post for details.

Computational Time Benchmark

fishy

We compared the computational time of JAX-CanVeg with the legacy Matlab-based CanVeg. The benchmark was performed on one A100 GPU and one AMD EPYC 7763 CPU at four flux tower sites. See this post for benchmark details.

Example: A Hybrid Ball-Berry Model

fishy

We demonstrated JAX-CanVeg's hybrid modeling capability by applying the model to simulate the water and carbon fluxes at four flux tower sites in the western United States with varying aridity. To this end, we developed a hybrid version of the Ball-Berry equation that emulates the impact of water stress on stomatal closure (Jiang et al., 2024). The scripts for reproducing the results of the paper are available in the folder examples. We applied the differentiable JAX-CanVeg at four flux tower sites to evaluate the performance of a hybrid version of the Ball-Berry equation. The model were trained against both observed latent heat fluxes and net ecosystem exchange. Below we illustrate the application example on US-Whs (which is applicable to the other three flux tower sites, i.e., US-Me2, US-Bi1, and US-Hn1).

  • Step 1: Train the process-based and hybrid JAX-CanVeg and the pure neural networks:
python [jax-canveg-folder]/examples/US-Whs/train_models.py
python [jax-canveg-folder]/examples/US-Whs/train_dnns.py

Note

Training multiple JAX-CanVeg models will take a pretty long time. It is suggested to train it in an HPC system. We provide an example of sbatch job script here.

  • Step 2: Evaluate the simulation performance of the trained models including both JAX-CanVeg and DNNs:
cd [jax-canveg-folder]/examples/US-Whs
python postprocessing.py
  • Step 3: Calculate the parameter sensitivity of selected models:
cd [jax-canveg-folder]/examples/US-Whs
python calculate_sensitivity.py

Note

This step requires the completion of model training at all four sites. One can modify the code to calculate the sensitivity at specified sites.

Some simulations of the trained JAX-CanVeg --

fishy

License

Distributed under Simplified BSD License. LICENSE for more information.

Acknowledgements

This work was funded by the Laboratory Directed Research and Development Program at Pacific Northwest National Laboratory and the ExaSheds project supported by the United States Department of Energy, Office of Science, Office of Biological and Environmental Research, Earth and Environmental Systems Sciences Division, Data Management Program.

Citation

Jiang et al., (2024). JAX-CanVeg: A Differentiable Land Surface Model. Water Resources Research, in review.

Contacts

Peishi Jiang ([email protected])