Processed InShaPe dataset: Large-scale multi-shape phase retrieval dataset for PBF-LB/M laser beam shaping SLM phase mask estimation.

Copyright claim: This dataset, the "Processed InShaPe dataset for FourierGSNet" is modified from the original InShaPe dataset [1] published by the publication "Deep learning based phase retrieval with complex beam shapes for beam shape correction" [2] under the terms and conditions of CC BY 4.0 LicenseModifications from the original InShaPe dataset:This dataset and the original InShaPe dataset are both paired dataset. The pairs in the original InShaPe dataset are: 12 Zernike coefficients - 3 intensity images; but the pairs in this dataset are modified from the original corresponding sample in InShaPe dataset to be: 1 Phase map - 1 intensity image.Along with the original InShaPe dataset, Yan, et al. also publicly released six optical simulation scripts [3] (also under CC BY 4.0 License) that generated the subsets of the six beam shapes. We used the Zernike coefficients given in the original InShaPe dataset of Yan, et al. to re-run the simulations and reproduced the aberrated SLM phase maps and distorted beam shape images in original size 1281×1281. We crop reproduced phase maps and beam shape images around the aperture to reduce matrices size to 427×427 to accelerate trainings. Therefore, each paired sample of the processed InShaPe dataset is an aberrated SLM phase map serving as the ground truth (supervision label) and a distorted beam shape intensity image serving as the input for phase retrieval. The same data processing is conducted on the six subsets of the six beam shapes, respectively, to obtain the same six training sets and test sets as original dataset.<br>References:[1] Shengyuan Yan, Richard Off, Anil Bora Yayak, Katrin Wudy, Anoush Aghajani-Talesh, Markus Birg, Jonas Grünewald, Mike Holenderski, and Nirvana Meratnia (2025). Large-scale multi-beamshape phase retrieval dataset based on Zernike coefficients for PBF-LB/M systems. Optica Publishing Group. Dataset. https://doi.org/10.6084/m9.figshare.27650703.v2[2] Shengyuan Yan, Richard Off, Anil Bora Yayak, Katrin Wudy, Anoush Aghajani-Talesh, Markus Birg, Jonas Grünewald, Mike Holenderski, and Nirvana Meratnia (2025). Deep learning based phase retrieval with complex beam shapes for beam shape correction. Optics Express <b>33</b>, 10806-10834. https://doi.org/10.1364/OE.547138[3] Shengyuan Yan, Richard Off, Anil Bora Yayak, Katrin Wudy, Anoush Aghajani-Talesh, Markus Birg, Jonas Grünewald, Mike Holenderski, and Nirvana Meratnia (2025). Codes for optical simulation, deep learning models training and testing, and the simulation-assisted analysis of aberration detection and beam shape correction accuracies for PBF-LB/M systems. Optica Publishing Group. Software. https://doi.org/10.6084/m9.figshare.27650946.v2<br>This simulated dataset containing six beam shapes is used for training and testing on the complex forward path application, PBF-LB/M beam shaping system, in our paper "Efficient Gerchberg-Saxton algorithm deep unrolling for phase retrieval with complex forward path". In this dataset, there are 6 sub-datasets stored in 6 zip folders named by "dense+[beamshape name]+30k_pre.zip". For each beam shape, the trainging set and the test set has 10000 and 3000 pairs of aberrated PBF-LB/M beam shape samples, respectively. Each paired sample contains one intensity profile of the aberrated beam shape in an imaging plane 5 times Rayleigh length before the back focal plane and one corresponding aberrated SLM phase mask that generated the aberrated beam shape. The aberrations in the paired samples are modelled with linear combination of Zernike polynomials amplified by randomly sampled Zernike coefficients.In order for the convenience of distributing the dataset, this repository does not contain the simulated paired samples. Instead, we give the original Zernike coefficients stored in Numpy array .npy files and the corresponding simulation script so that the users can run the simulation to reproduce the entire dataset.