SIB-CL Datasets

This repository contains the Surrogate- and Invariance-Boosted Contrastive Learning (SIB-CL) datasets for two scientific problems:

• PhC2D: 2D photonic crystal density-of-states (DOS) and bandstructure data.
• TISE: 3D time-independent Schrödinger equation eigenvalue and eigenvector solutions.

The data and loader scripts reproduce the behavior of the original PyTorch Dataset classes from the SIB-CL paper:

Loh, C., Christensen, T., Dangovski, R., Kim, S., & Soljačić, M. (2022). "Surrogate- and invariance-boosted contrastive learning for data-scarce applications in science." Nature Communications, 13, 4223. https://www.nature.com/articles/s41467-022-31915-y

Repository Structure

• phc2d.h5: HDF5 with groups:

  • unitcell/mpbepsimage/<index>: 32×32 permittivity images
  • unitcell/epsavg/<index>: scalar average permittivity
  • mpbcal/DOS/<index>, mpbcal/DOSeldiff/<index>: DOS arrays
  • mpbcal/efreq/, mpbcal/emptylattice/, mpbcal/eldiff/ for bandstructure

• tise.h5: HDF5 with groups:

  • unitcell/potential_fil (or _res<downres>): potential arrays
  • eigval_fil/<index>: eigenvalues
  • eigvec_fil/<index>: eigenvectors

Processing & Training

For processing pipelines and model training code, see the SIB-CL GitHub repository: https://github.com/clott3/SIB-CL

Citation: Loh, C., Christensen, T., Dangovski, R., Kim, S., & Soljačić, M. (2022). "Surrogate- and invariance-boosted contrastive learning for data-scarce applications in science." Nature Communications, 13, 4223. https://www.nature.com/articles/s41467-022-31915-y