How AI Is Redesigning LIGO and Quantum Experiments for Unprecedented Sensitivity

LIGO Sensitivity Optimization

Researchers at the California Institute of Technology (Caltech) supplied the full specification of LIGO’s optical components—mirrors, beam splitters, high‑power lasers, and suspension systems—to the quantum‑optics design platform developed by Mario Krenn’s group. The AI‑driven optimizer initially generated a concept with a multi‑kilometre ring arm and thousands of elements, which was subsequently refined through iterative analysis. The final configuration adds a 3 km circular recycling cavity between the main interferometer and the photodetector. This cavity repeatedly circulates the light field, exploiting a noise‑suppression technique originally proposed by Soviet researchers to lower quantum radiation‑pressure noise. Simulations indicate that incorporating this AI‑suggested architecture from the design stage could increase LIGO’s strain sensitivity by 10 %–15 % , effectively expanding the observable volume for gravitational‑wave events.

Related link: https://arxiv.org/abs/2312.04258

Quantum‑Optics Experiment Redesign

The University of Tübingen team, led by Mario Krenn, employed the open‑source package PyTheus to let a graph‑based AI redesign the 1993 entanglement‑swapping experiment originally proposed by Anton Zeilinger. In the AI model, each optical element (beam paths, nonlinear crystals, detectors) is represented as a node in a graph, and edges encode photon‑pair creation and interference. The target state is a pair of photons entangled without a shared past, i.e., a Bell state generated from independent sources. The optimizer produced a layout requiring only four nonlinear crystals and three optical branches , compared with the original design that used eight crystals and a more complex network of beam splitters. The reduced architecture preserves the same entanglement fidelity.

Experimental verification by Xiao‑Song Ma’s group at Nanjing University (December 2024) reproduced the AI‑generated scheme and measured an entanglement fidelity exceeding 90 % , while the total number of optical components was cut by more than 40 % . This demonstrates that AI can discover compact, high‑performance quantum‑optics configurations.

Related link: https://arxiv.org/abs/2210.09981

AI‑Assisted Data Analysis in High‑Energy Physics

Kyle Cranmer (University of Wisconsin–Madison) trained a supervised machine‑learning model on simulated dark‑matter halo catalogs and on observable properties of neighboring galaxy clusters (e.g., velocity dispersion, X‑ray luminosity). The model learned an empirical mapping that predicts halo density with a root‑mean‑square error reduced by roughly 15 % relative to the standard analytical fitting functions used in cosmology.

Separately, a team at the University of California, San Diego applied a self‑supervised neural network to raw LHC collision data without providing any physics priors. After training, the network’s internal representations revealed invariance under Lorentz transformations, effectively rediscovering Lorentz symmetry from the data. The model also confirmed that particle‑production rates are statistically independent of the Earth’s rotation, illustrating AI’s capacity to uncover “zero‑hypothesis” symmetries in high‑dimensional experimental datasets.

Related links: https://arxiv.org/abs/2006.11287 , https://arxiv.org/abs/2310.00105