Notebooks and Tutorials¶
The repository ships a maintained notebook set that mirrors the main code paths. These notebooks are tutorial-style companions to the documentation, not separate experimental branches.
How to Use Them¶
- read the corresponding site page first if you want the conceptual overview
- open the notebook locally in Jupyter or your IDE
- treat the notebook as the interactive version of the same maintained workflow
Start Jupyter from the repo root with:
jupyter lab
Notebook Map¶
| Notebook | Focus | When to use it | Link |
|---|---|---|---|
01_prepare_data.ipynb |
dataset construction and split logic | when you want to inspect the processed CSV pipeline interactively | Open on GitHub |
02_train.ipynb |
TGNN-Solv training, optional Stage 0 pretraining, and encoder/config variants | when you want to step through the curriculum, --pretrain, and tuned TGNN follow-up configs |
Open on GitHub |
03_inference.ipynb |
single-point inference, temperature scan, and AD/OOD checks | when you want to inspect one system manually | Open on GitHub |
04_evaluation.ipynb |
metrics, uncertainty, calibration, and error analysis | when you want richer post-hoc analysis than the quick CLI | Open on GitHub |
05_baselines.ipynb |
DirectGNN, descriptor baselines, RF, and external baselines | when you are comparing TGNN against non-physics alternatives | Open on GitHub |
06_ablations.ipynb |
ablation reading and architectural comparisons | when you want to isolate which component is helping | Open on GitHub |
07_temperature.ipynb |
temperature dependence, ideal solubility, and van't Hoff analysis | when you want to inspect thermal trends rather than aggregate metrics | Open on GitHub |
08_optuna_tuning.ipynb |
Optuna-based hyperparameter tuning | when you want interactive tuning for TGNN, GPS TGNN, descriptor-augmented TGNN, and DirectGNN families | Open on GitHub |
How They Map to the Site¶
| Site page | Notebook companion |
|---|---|
| Data Preparation | 01_prepare_data.ipynb |
| Training | 02_train.ipynb |
| Evaluation & Inference | 03_inference.ipynb, 04_evaluation.ipynb |
| Baselines | 05_baselines.ipynb |
| Experiments & Benchmarks | 06_ablations.ipynb, 07_temperature.ipynb, 08_optuna_tuning.ipynb |
Recommended Reading Order¶
For a first pass through the project:
01_prepare_data.ipynb02_train.ipynb03_inference.ipynb04_evaluation.ipynb
Then move on to the experiment notebooks depending on your question:
- architecture choices:
06_ablations.ipynb - temperature behavior:
07_temperature.ipynb - tuning:
08_optuna_tuning.ipynb
Important Scope Note¶
The notebooks are kept aligned with the maintained implementation, but the reproducible default remains the grouped CLI surface under:
scripts/data/scripts/training/scripts/evaluation/scripts/experiments/
Use notebooks for inspection and explanation, and the CLIs for reproducible batch runs.
Current alignment notes:
02_train.ipynbnow covers both notebook-driven Stage 0 and the maintained CLI path throughscripts/training/train.py --pretrain/scripts/training/train_with_pretrain.py08_optuna_tuning.ipynbmirrors the currentOptunaTunermodel aliases, including GPS and descriptor-augmented TGNN variants
Current scope note:
- the formal benchmark-adapter API
- checksum-based benchmark-release freezing
- the thermodynamic stress suite
are currently CLI-first rather than notebook-first surfaces.