From algorithms to experiments: Xuemei Gu contributes to Nature perspective on AI-driven scientific discovery

Can artificial intelligence do more than analyse data or optimize an existing experiment – can it actually conceive new physics experiments? A new Review in Nature, co-authored by ACP Principal Scientist Dr. Xuemei Gu, explores how AI could fundamentally change the way experiments are designed. For Gu, the topic also reflects an unusually interdisciplinary scientific journey from electrical engineering and computer science to hands-on quantum optics and, ultimately, AI-driven scientific discovery.

The design of a physics experiment has traditionally relied heavily on the knowledge, intuition and creativity of researchers. Yet the space of all possible experimental configurations can be enormous. Even a relatively small collection of optical components can be combined in an immense number of ways, leaving most possibilities beyond what researchers could systematically explore.

In a new Review published in Nature, Jonathan Klimesch and colleagues examine how artificial intelligence and computational methods could increasingly take part in this creative process. Rather than merely adjusting a few parameters of a predefined setup, algorithms can search over components, experimental topologies and parameters to identify entirely new configurations. The authors organize this emerging approach around four central challenges: defining sufficiently expressive search spaces, developing fast and reliable physics simulators, translating scientific questions into suitable objective functions, and creating AI exploration methods capable of navigating these vast spaces.

Among the co-authors is Dr. Xuemei Gu, Principal Scientist at the Abbe Center of Photonics and head of the AI Quantum Lab at Friedrich Schiller University Jena. Her own career has brought together precisely the disciplines that are converging in this new field.

A path between disciplines

Gu initially worked in electrical engineering, building practical measurement devices during her bachelor's and master's studies. An increasing interest in algorithms and computational theory then led her to pursue a PhD in computer science at Nanjing University. Her first encounter with quantum information came through a summer school – and ultimately redirected her scientific career.

During her PhD, Gu spent two and a half years at the Institute for Quantum Optics and Quantum Information in Vienna in the environment of Nobel Laureate Anton Zeilinger. There, she moved from computational descriptions of quantum systems toward actual optical experiments. In one project on sorting spatial modes of light, she used her computer-science background to develop a computational approach for finding possible experimental solutions and then implemented these ideas in the laboratory herself. She subsequently joined the University of Science and Technology of China, where she worked on experimental quantum optics and expanded her expertise in quantum information technologies. Her research included photonic quantum networks, multiparticle entanglement and quantum gates. But rather than leaving computer science behind, Gu eventually came back to it from a new perspective. She had seen first-hand how strongly modern experiments already depend on algorithms, electronic control and computational tools. This experience convinced her that advances in artificial intelligence could become powerful instruments for experimental physicists. As she later described her career path, her original intention had been to remain an experimental quantum physicist; instead, she returned to computer science to learn advanced AI techniques that could be applied to quantum experiments.

Can AI become a scientific designer?

At the Max Planck Institute for the Science of Light in Erlangen, Gu joined Mario Krenn's Artificial Scientist Lab and received an Alexander von Humboldt Research Fellowship for her project on AI-assisted discovery of novel quantum-optical imaging concepts.

In February 2025, she established the independent AI Quantum Lab at the University of Jena. The group develops AI methods for de novo quantum-experiment design, interpretable machine learning, scientific idea generation and automated discovery. This is also where Gu's research connects particularly closely to the vision developed in the new Nature Review. The article discusses a progression from comparatively narrow parameter optimization toward de novo discovery: algorithms searching spaces sufficiently broad to uncover experimental concepts that may lie outside established human design conventions. It reviews examples ranging from quantum optics and microscopy to particle physics, fusion research and gravitational-wave detection. Looking further ahead, the authors envision combining increasingly comprehensive physics simulators with large sets of scientific objectives. Such systems could allow AI to explore possible experiments across conventional disciplinary boundaries and identify unexpected ways of testing physical theories. Gu's own work already extends beyond designing individual experiments. Together with collaborators, she has investigated how machine learning, knowledge graphs and large language models can identify promising research directions and generate new scientific ideas. Two of these strands of research – autonomous quantum-physics research using LLM agents and machine-learning-based forecasting of potentially high-impact research topics – are explicitly discussed in the Nature Review. The long-term ambition is not to remove scientists from the process. On the contrary, the Review argues that AI could change their role: researchers would increasingly define scientific questions, constraints and objectives while intelligent computational systems explore possibilities that would otherwise remain inaccessible. In the authors' words, this could shift the scientist's role from an intuition-guided designer towards an “orchestrator of intelligent automated exploration.”

For Gu, whose scientific path has repeatedly crossed the boundaries between hardware, algorithms, quantum experiments and artificial intelligence, that future is already becoming a research programme in Jena.

Publication

J. Klimesch et al., “Designing physics experiments with artificial intelligence,” Nature 657, 47–58 (2026). Published online 2 September 2026. DOI: 10.1038/s41586-026-10898-6.

Link to Xumei Gu's junior research group at ACP