BioCovidLab

Team workspace for the Germany COVID-19 forecasting project in Bio-AI.

Repository Layout

BioCovidLab/
├── data/
│   ├── README.md
│   └── raw/
├── docs/
├── notebooks/
├── outputs/
│   └── figures/
├── references/
│   ├── course_materials/
│   └── papers/
├── reports/
│   ├── latex/
│   └── pdf/
├── scripts/
├── .gitignore
├── README.md
└── requirements.txt

What Goes Where

• notebooks/: project notebooks for setup, EDA, and experiment planning
• outputs/figures/: shared place to export figures from notebooks and scripts
• data/raw/: downloaded CSV and JSON source data
• references/papers/: research papers used in the project
• references/course_materials/: local lecture material copied into the repo
• reports/latex/: editable LaTeX source for the living project report
• reports/pdf/: compiled PDF snapshots for sharing
• docs/: short project-facing documentation and workflow notes
• scripts/: helper scripts for notebook generation and utilities

Recommended Workflow

1. Install dependencies from requirements.txt.
2. Run notebooks/COVID19_Forecasting_Experiment.ipynb to refresh local data and papers.
3. Use notebooks/COVID19_Dataset_EDA.ipynb for exploratory analysis.
4. Use notebooks/COVID19_Experiment_Plan.ipynb as the design and evaluation guide.
5. Update the living report in reports/latex/ and export milestone PDFs to reports/pdf/.

Setup

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
python -m ipykernel install --user --name biocovidlab --display-name "BioCovidLab"

Report Build

From reports/latex/:

powershell -ExecutionPolicy Bypass -File .\build.ps1

The compiled report will be written to reports/pdf/covid19_experiment_report.pdf.

Notes For Teammates

• Do not commit .venv/, .jupyter_runtime/, or LaTeX build artifacts.
• Keep raw data immutable in data/raw/.
• Export ad-hoc or presentation figures into outputs/figures/notebooks/ or outputs/figures/scripts/.
• Put new implementation code in scripts/ for now, or add a dedicated src/ package once model code grows.
• When adding new figures for the report, export them from notebooks and place them in reports/latex/figures/.

Presentation Notes (2026-04-29)

• The SIR ODE model is easy to interpret because it explains the epidemic through susceptible, infectious, and removed compartments.
• A key limitation is parameter estimation: the original dataset does not provide direct recovery observations, so the recovery process must be approximated indirectly.
• Because of that missing recovery information, the fitted ODE parameters can be unstable and the resulting compartment dynamics may be far from the real data distribution.
• A plain MLP can fit the observed curves more flexibly, but it has a clear risk of overfitting and may learn noise instead of epidemic structure.
• A PINN combines data fitting with the SIR ODE by adding the ODE residual as a regularization term in the loss.
• This means the PINN can return an ODE or SIR-style equation together with the neural approximation.
• However, the learned SIR equation can still remain far from the real data because the current formulation uses constant parameters over the full time horizon.
• In practice, epidemic behavior changes over time because of interventions, testing policy, reporting effects, vaccination, and behavioral shifts, so constant-parameter dynamics are often too restrictive.