Unconventional Yet Effective Experimental Designs: AI Says 'I Think This Might Work'}

In the history of physics, experimental design often demands researchers to exhaust their imagination and repeatedly test and refine: from the arrangement of optical components to the fine-tuning of particle detectors, every step embodies the wisdom and effort of scientists.

Today, AI acts as a new "collaborator," using its extraordinary "I think this might work" approach—beyond conventional thinking—to propel experimental physics into a new era of "bizarre yet effective" innovations.

LIGO Sensitivity Optimization

At Caltech, physicist Rana Adhikari and colleagues, responsible for optimizing LIGO (Laser Interferometer Gravitational-Wave Observatory), turned to AI to break through bottlenecks after the first gravitational wave detection in 2015.

They fed the available LIGO components—lenses, mirrors, lasers—into a quantum optical experiment design software developed by Mario Krenn’s team. Initially, the AI proposed ideas like a 100-kilometer ring-shaped arm with over a thousand components, seemingly wild but deeply insightful.

After months of analysis, the team discovered that the AI added a 3-kilometer ring cavity between the main interferometer and detector, effectively circulating the light and using noise suppression principles from Soviet physicists decades ago to reduce quantum noise to unprecedented levels.

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

Image: Detector located in Livingston, Louisiana.

Adhikari noted, "If AI had been used during LIGO’s construction, its sensitivity could have improved by 10-15%." In the sub-proton measurement realm, this is like detecting 10% more signals in sub-proton-level measurements—an enormous boost.

Unexpected Joys of Quantum Optics

In classical physics, every object has definite properties independent of observation. But in the quantum world, objects are described by quantum states, which are mathematical entities used to calculate the probability of finding an object at a certain position.

This includes phenomena like entanglement and shared quantum states. At the University of Tübingen, Mario Krenn’s team used software called PyTheus to attempt AI redesigns of the 1993 Zeilinger group's quantum entanglement swapping experiment.

AI abstracted elements like optical paths, crystals, and detectors as graph nodes, aiming for a "no common history" entanglement between two particles. The resulting design was much simpler—requiring only four crystals and three beam splitters—yet achieved the same entanglement.

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

In December 2024, Nanjing University’s Xiao‑Song Ma’s team successfully replicated the AI-designed setup, achieving over 90% entanglement fidelity, reducing experimental complexity by over 40%, and opening new avenues for quantum communication and networks.

Exploring Physics with AI

AI’s role extends beyond experiment design; it also aids in data analysis. Kyle Cranmer from UW–Madison used machine learning to predict dark matter clump densities, fitting a new formula that reduced prediction errors by 15% compared to traditional models.

Image: Illustration of extracting physical equations from data sets. Source: related link.

Related link: https://arxiv.org/abs/2006.11287

UCSD’s team applied AI to Large Hadron Collider data, successfully deriving Lorentz symmetry without prior physics knowledge, confirming that particle production rates are independent of Earth's rotation, demonstrating AI’s potential in high-dimensional data analysis.

Related link: https://arxiv.org/abs/2310.00105

In these cases, AI does not merely replace humans but inspires scientists to reconsider domain boundaries with "bizarre" solutions. As Caltech’s Aephraim Steinberg said: "When a team’s design, long thought impossible, is suddenly realized by AI, we realize the depth of exploration still left to discover."