CaDRe

Reference implementation of CaDRe (Causal Discovery and Representation learning), ICML 2026.

CaDRe jointly recovers the latent dynamic process and the observed causal graph from a multivariate time series, with nonparametric identifiability guarantees.

• Paper: OpenReview · arXiv
• Authors: Minghao Fu, Biwei Huang, Zijian Li, Yujia Zheng, Ignavier Ng, Guangyi Chen, Yingyao Hu†, Kun Zhang†

_{† Equal advising. Affiliations: MBZUAI, CMU, UC San Diego, Johns Hopkins University.}

Install

git clone https://github.com/MinghaoFu/CaDRe.git
cd CaDRe
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Layout

Path	What lives here
SSM/	Main CaDRe trainers and evaluators (CESM2, WeatherBench).
LiLY/	VAE encoders / decoders, flow priors, latent dynamics modules.
Caulimate/	Lightweight utilities for data simulation, graph operations, metrics.
forecasting/	Forecasting evaluation notebooks.
analyze/	Climate visualisation and downstream analysis.
dataset/	Data pre-processing scripts.
scripts/	Training and inference launchers.
scripts/repro/	One-line reproducer for each paper experiment.
tests/	Smoke tests.

Reproduce the paper

Set the environment variables (each script reads them):

export PROJECT_ROOT="$(pwd)"
export DATA_DIR=/path/to/data        # CESM2 / WeatherBench / ERSST
export CKPT_DIR=/path/to/checkpoints
export LOG_DIR=/path/to/logs

Run any of the following:

./scripts/repro/run_synthetic.sh         # §5.1
./scripts/repro/run_neural_baselines.sh  # App. D, Fig. 8
./scripts/repro/run_causalrivers.sh      # App. D, Fig. 9(b)
./scripts/repro/run_higher_order.sh      # App. D, Table 16
./scripts/repro/run_cesm2.sh             # §5.2, Tables 5, 6
./scripts/repro/run_weatherbench.sh      # §5.2, Tables 5, 6
./scripts/repro/run_ersst.sh             # §5.2, Table 6
./scripts/repro/run_forecasting.sh       # §5.2, Tables 5, 18-20
./scripts/repro/run_visualize.sh         # §5.2, Figs. 5, 10, 11
./scripts/repro/run_all.sh               # everything, sequentially

Data

We use four public datasets. We do not redistribute the raw data; download from the official sources and place under $DATA_DIR/:

• CESM2 Pacific SST — NCAR CESM2 control run.
• ERSST v5 — NOAA NCEI.
• WeatherBench — Rasp et al. 2020.
• Weather (Jena) — BGC Jena.

Citation

@inproceedings{fu2026learning,
  title     = {Learning General Causal Structures with Hidden Dynamic Process for Climate Analysis},
  author    = {Fu, Minghao and Huang, Biwei and Li, Zijian and Zheng, Yujia and
               Ng, Ignavier and Chen, Guangyi and Hu, Yingyao and Zhang, Kun},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026}
}

License

MIT. See LICENSE.

Contact

• Minghao Fu — minghao.fu@mbzuai.ac.ae
• Biwei Huang — bih007@ucsd.edu