08/24/21
Yang Deng

********************************************************************************************************************************
*                                               									       *
* This is the dataset repository for "Benchmarking Deep Learning Models in Artificial Electromagnetic Material Problems".      *
*                                                                                                                              *
* This repository contains two new datasets included in the work: All-dielectric metasurface and Nanophotonics particles.      *
* All-dielectric metasurface data is under directory "ADM".                                                                    *
* Nanophotonics particles data is under directory "Nano".                                                                      *
*															       *
********************************************************************************************************************************

Details for all-dielectric metasurface (ADM) data:

The ADM data contains geometry and absorptivity spectrum pairs {g,s}. The input geometry has 14 dimensions. The output spectrum has 2001 dimensions. The total amount of ADM dataset is near 60,000 simulations {g,s} pairs, and we took out 10% of the data as an independent testset.
The rest of 90% of the data builds up the training dataset to train neural networks mapping ADM's geometry to spectrum. We categorize the geometry data into data_g.csv and spectrum data into data_s.csv.
Similarly, for the testset, we categorize the geometry data into test_g.csv and spectrum data into test_s.csv. All the simulations are generated by computational software CST Microwave Studio.


Details for nanophotonics particles data:

The nanophotonics data contains geometry and scattering responses spectrum pairs {g,s}. The input geometry has 8 dimensions. The output spectrum has 201 dimensions. The total amount of nanophotonics particles dataset is 50,000 {g,s} pairs, and we took out 10% of the data as an independent test set.
The rest of 90% of the data builds up the training dataset to train neural networks mapping geometry to scattering responses. We categorize the geometry data into data_g.csv and spectrum data into data_s.csv.
Similarly, for the testset, we categorize the geometry data into test_g.csv and spectrum data into test_s.csv. The data is generated following "Peurifoy, John, et al. "Nanophotonic particle simulation and inverse design using artificial neural networks." Science advances 4.6 (2018): eaar4206."