Variational Digital Twin

This repository contains benchmark workflows for variational digital twin modeling across three application domains:

• NASA Battery: battery aging and voltage prediction
• HTTF: thermal forecasting
• PSML: renewable power forecasting

Paper

Published article DOI: https://doi.org/10.1016/j.egyai.2026.100756

Environment setup

Create a reproducible conda environment from the provided specification file:

conda env create -f environment.yml
conda activate variational-digital-twin

Optional notebook automation dependency:

pip install papermill

Reproducing paper artifacts

1) Run model training/evaluation scripts

Run the experiment scripts relevant to your study (NASA_Battery, HTTF, and/or PSML) to produce model outputs and figures.

2) Collect outputs into paper_results/

From the repository root, run:

python scripts/generate_paper_results.py --clean

The collection script performs the following steps:

1. Executes the configured plotting scripts:
   • NASA_Battery/plot_static_vs_rolling.py
   • HTTF/static_training/plot_model_comparisons.py
2. Collects selected figures and result artifacts from NASA_Battery/, HTTF/, and PSML/.
3. Copies collected files into paper_results/ using source-relative paths.
4. Writes paper_results/MANIFEST.md with an index of included artifacts.

If figures are already generated and you only want to re-collect files:

python scripts/generate_paper_results.py --skip-plots --clean

Reproducibility note

Some training pipelines are stochastic. Exact numerical values may vary between runs, while overall performance trends should remain similar.