hpcom: GPU-accelerated optical fibre link simulator

Python · TensorFlow · GPU · open source (GPLv3) · GitHub · PyPI · Docs

Simulating light propagating through fibre is the bottleneck of most research in optical communications, and it is even worse when you need millions of labelled examples to train a machine learning model. hpcom is the simulator I wrote for that job at Aston, and it now underpins several of my publications.

• Fast. Split-step Fourier and Manakov solvers on TensorFlow/GPU, with simulation speed-ups of up to 2000×.
• Composable API. Transmitter, Fiber, EDFA, Receiver, Link and Simulation are validated, frozen dataclasses. Custom link components drop in without plumbing changes.
• Swappable backends. A propagator interface sits between the physics and the numerics (NumPy and TensorFlow today; JAX planned).
• Trustworthy. Numerical tests pin soliton shape preservation, energy conservation, AWGN BER against theory and digital back-propagation round trips, each with tolerances cited from the literature.
• Reproducible. One seed on a Simulation splits into independent random streams, so identical seeds give identical results, including amplifier noise.
• Dataset generation for ML workflows at scale, and a live demonstration at ECOC 2023 (paper).