ElectroTS

Long-horizon forecasting of electrolysis stability from early-stage operating data through temporal learning

🔍 Overview

ElectroTS is a data-driven electrolysis prognostic framework that integrates data-adaptive rectification with temporal sequence modeling to enable robust prediction of lifetime, degradation rates, and uncertainty in electrolytic systems.

Using only early-stage operating data, ElectroTS forecasts device stability over operating durations exceeding tenfold the input window and beyond 1,000 hours, providing actionable guidance for rapid screening of materials and operating conditions. Few-shot transfer learning enables efficient cross-platform generalization across both low-temperature aqueous (AEMWE) and high-temperature solid-state (SOEC) electrolysis systems.

---

🚀 Installation

Python ≥ 3.10 is required. We recommend uv for environment management.

pip install -e .

---

📌 Datasets

We construct two experimental stability datasets explicitly designed for long-horizon forecasting:

Dataset	System	Abbrev.
Anion-exchange membrane water electrolysis	Low-temperature aqueous	AEMWE
CO₂ solid oxide electrolyzer cells	High-temperature solid-state	SOEC

Data availability: datasets will be made publicly available upon journal acceptance. Please check back or watch this repository for updates.

---

🔥 Model Training

Method 1: training scripts

bash scripts/train_aemwe.sh   # AEMWE leave-one-out training
bash scripts/train_soec.sh    # SOEC fine-tuning (few-shot transfer)

Method 2: CLI

After installation, the electro-ts command is available:

electro-ts \
  --model_id my_experiment \
  --root_path ./data/aemwe_all_p0.1 \
  --model ElectroTS \
  --time_cutoff 150 \
  --seq_len 1000 \
  --label_len 125 \
  --pred_len 250 \
  --e_layers 4 \
  --d_model 400 --d_core 400 --d_ff 400 \
  --learning_rate 0.0001 \
  --lradj cosine \
  --train_epochs 20 \
  --patience 20 \
  --batch_size 32 \
  --save_model

Equivalent alternatives:

python train.py   [same args]   # convenience script
python -m electro_ts [same args]  # module invocation

---

🗂️ Repository Structure

ElectroTS/
├── train.py                   # Convenience entry point
├── pyproject.toml             # Package metadata and dependencies
├── scripts/
│   ├── train_aemwe.sh
│   └── train_soec.sh
└── electro_ts/                # Installable package
    ├── cli.py                 # Argument parser + leave-one-out loop
    ├── exp/
    │   ├── exp_basic.py       # Base experiment class
    │   └── aemwe.py           # Train / validate / test / predict
    ├── models/
    │   ├── ElectroTS.py       # Model
    │   └── layers/
    │       ├── embed.py
    │       └── transformer_enc_dec.py
    ├── datasets/
    │   ├── data_factory.py
    │   └── data_loader.py     # ElectroDataset
    └── utils/
        ├── tools.py           # EarlyStopping, AverageMeter, LR scheduler
        └── eval.py            # decay_rate, lifetime, compute_errors

---

📝 Citation

If you find this work useful, please cite:

@article{electro_ts_2026,
author = {Qiucheng Xu  and Junwu Chen  and Alexander Muroyama  and Xiangli Yi  and Ariana Serban  and Philippe Schwaller  and Xile Hu },
title = {Long-horizon forecasting of electrolysis stability from early-stage operating data through temporal learning},
journal = {ChemRxiv},
volume = {2026},
number = {0430},
pages = {},
year = {2026},
doi = {10.26434/chemrxiv.15002581/v1},
URL = {https://chemrxiv.org/doi/abs/10.26434/chemrxiv.15002581/v1},
note={Preprint}
}

---

🌈 Acknowledgements

This work was supported by the EPFL Solutions4Sustainability CCUS project and by the NCCR Catalysis (grant number 225147), a National Centre of Competence in Research funded by the Swiss National Science Foundation.

---

📫 Contact

If you have any question, welcome to contact me at:

Junwu Chen: junwu.chen@epfl.ch