Reproduce “Physics-Based Deep Learning for Fiber-Optic Communication Systems” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Christian Häger, Henry D. Pfister

Researchers at Chalmers University of Technology and Duke University introduced Learned Digital Backpropagation (LDBP), a physics-based deep learning model that re-parameterizes the split-step method for fiber-optic communication systems. LDBP achieved substantial reductions in computational complexity for nonlinear equalization, outperforming traditional DBP methods and revealing that the *cascaded* filter response, not individual step accuracy, is key to performance.

Key insights

  • The numerical split-step method for solving the Nonlinear Schr bd r bd dinger Equation (NLSE) has a functional form analogous to a deep multi-layer neural network, enabling a physics-based deep learning approach.
  • Analysis of learned filter responses revealed that LDBP optimizes the *cascaded* frequency response of its filters to be nearly perfectly flat, rather than requiring individual step filters to be perfect inverse chromatic dispersion approximations.
  • Combining data-driven optimization with established domain knowledge can guide model design, hyperparameter selection, and initialization, leading to new scientific insights for existing problems.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.