Chalmers Researchers Cut Optical AI Design Time 90% by Embedding Physics Laws
Updated
Updated · SciTechDaily · Jun 17
Chalmers Researchers Cut Optical AI Design Time 90% by Embedding Physics Laws
3 articles · Updated · SciTechDaily · Jun 17
Summary
Tasks that once took 30 days to train optical-design neural networks now take about three days after Chalmers researchers built electromagnetism directly into the models.
40,000 simulations can be needed in conventional training, with each data point taking 10 minutes to an hour, so the physics-informed approach sharply reduces data generation and retraining.
A trained network can then evaluate a structure’s optical properties in a millisecond, giving better estimates and avoiding obvious errors in nanophotonic material design.
The work targets artificial optical materials for thinner lenses and for quantum technologies, including photonic crystals that could help move information between quantum computers over optical frequencies.
Laser & Photonics Reviews published the study, which was backed by Swedish research funding and supercomputing resources as Chalmers develops Sweden’s first large-scale quantum computer.