MKoumans/pidu-data

PIDU simulated datasets

Simulated infrared spectra (Transfer Matrix Method, random Drude-Lorentz layer and oscillator parameters) with their ground-truth parameters, used to train and test the models of Physics-infused Deep Unfolding for Automated Material Parameter Extraction of Optical Data (Koumans, Stevens, van Sloun, van Mechelen).

Code: https://github.com/MKoumans/PIDU · Checkpoints: MKoumans/pidu-optical-parameter-extraction

Contents

Path	Content
data/datasets/usecase1/*.npz	Use case 1: train / val / test sets
data/datasets/usecase2/*.npz	Use case 2: train / val / test sets, with (wsn) and without (nwsn) the second system
data/labels/*.xlsx	Label mappings used during dataset generation
outputs/case<N>/outputs/loss_{CNN,DU,PIDU}_<i>_001/	Test exports of the trained runs (8 seeds per model; use case 2 includes test_exp, the predictions on the measured spectra)
outputs/case<N>/loss/loss_{CNN,DU,PIDU}_<i>_001/	Per-epoch train / val / test loss of those runs
outputs/case<N>/logs/train_<TYPE>_<i>.log	Training logs of those runs
models/usecase<N>/<TYPE>-t01-<i>.pt	Checkpoints of those runs (the Fig. 2 timings average over seeds)
outputs/case<N>/outputs/loss_lma_{baseline,noisy,random}/	TRF (LMA) baseline: use case 1 on the 1024 simulated test spectra, use case 2 on the 75 measured spectra

Each .npz holds the clean spectra (data_y_*), noised spectra (data_y_noised_*), initial and target parameters (data_x_*), and the wavenumber axis (wn_*).

Not included: the experimental SrTiO₃ measurements themselves (the predictions and fits on them are).

Usage

pidu-hub download --usecase 1 2                    # everything
pidu-hub download --what data logs --usecase 1      # datasets and training logs only

Citation

To be added.