Predicting Bandstructures of 2D Square Photonic Crystals

Introduction

This repo hosts the code used in [https://doi.org/10.1515/nanoph-2020-0197] where a simple neural network with CNN and fully connected layers was used to predict the bandstructures of 2D square lattice Photonic Crystals.

Dataset

The datasets can be downloaded from this link. A description of the datasets are provided here (also check the /metadata/ tag inside the .h5 files):

This .hdf5 file contains MPB calculations of square photonic crystals of either TE or TM polarization, as indicated in /metadata/ attribute runtype. The computational resolution is indicated in /metadata/ attribute resolution. Besides the /metadata/ group, this .hdf5 file contains a /shapes/ group. This group in turn contains 20,000 numerically labelled groups (1, 2, etc.) which give input and results for distinct unitcell calculations. Each such numeric subgroup (e.g., /shapes/1/) contain several subfields:

• kvecs: a flatted 23×23 grid of 2-dimensional k-vectors, in units of reciprocal lattice vectors (i.e. $2\pi k/a$)
• eigfreqs: eigenfrequencies at the corresponding kvecs, in units of $c/a$ (i.e. $\omega a/c$). Includes first 6 bands
• bandgap_1: a 2-element array with elements:
  • 1st element: bandgap size between 1st and 2nd band, in units of $c/a$. If negative, value is NaN
  • 2nd element: gap-to-midgap ratio (i.e. relative gap size), in dimensionless units (i.e. a fraction)
• unitcell: a group with subfields that describe the unitcell:
  • epsilon: the dielectric function on the unit cell, as input to MPB, in a 256×256 grid
  • epsilon_comput: the downsampled version of epsilon used internally in MPB. Has size equal to /metadata/resolution/
  • epsilon_average: the average of /unitcell/epsilon/, for convenience
  • boundary: $[N,2]$ array giving the boundary between inside/outside materials in the unitcell. It is assumed that the unit- cell lies in the $[-0.5, 0.5]^2$ box (i.e. lengths are given in units of lattice constants $a$)
  • epsin: value of dielectric function inside inclusion
  • epsout: value of dielectric function outside inclusion
  • areain: relative area of inclusion (to entire unitcell area)

Usage

Upon cloning, the code can be run directly using run.py where the path to the directory containing the dataset should be specified using the --path_to_h5 flag. The default hyperparameters defined in the parser will give the optimal results reported in the paper.