The dataset contains 11,000 generated ARPES spectral images (x_train and x_val, ) along with their corresponding ground truth segmentation masks ( y_train andy_val).
The data is provided in HDF5 (.h5) format.
Data Access & Code: The simulation code used to generate this dataset is publicly available at link.
The dataset is partitioned into a training set and a validation set:


Training set: 10,000 samples (x_train, y_train). intended for model training.


Validation set: 1,000 samples (x_val, y_val) intended for evaluating model performance.


The input spectra have a shape of (128,128,1), while the corresponding segmentation masks have a shape of (128,128,2). This specific dimensionality was chosen to ensure compatibility with U-Net architecture, which requires input dimensions to be powers of two (e.g., 128×128). 
This dataset was used to train and evaluate the models proposed in: Pustovit, Yu. V., & Ohloblia, M. O. (2026). Semantic segmentation of ARPES spectra for electronic dispersion visualization. Low Temperature Physics, 52(2), 161–168. https://doi.org/10.1063/10.0042266