Research team is getting closer to creating an AI physicist
With SciExplorer, scientists at the Max Planck Institute for the Science of Light (MPL) are introducing an AI science agent that can be deployed across a broad range of physical experiments. Based on a large language model, this artificial scientist agent is a generalist: it is capable of automating the process of scientific research without the need for task-specific fine-tuning. The results were recently published in Physical Review X.
At the core of the natural sciences lies the drive to generate new knowledge for a deeper understanding of systems around us. Scientific discovery is based on a cycle of observations, analyses, and generating new hypotheses. The research team led by Prof. Florian Marquardt, Director at MPL and Head of the Theory Division, has now presented a new, agent-based approach that can autonomously carry out the cycle of this scientific process for physical systems.
SciExplorer is an artificial scientist based on a large language model (LLM) that automates this process. The capabilities of these LLMs are evolving rapidly. Access to external tools also boosts their performance in impressive ways. Recently, the first attempts to exploit this so-called agentic approach in the domain of science have shown promising results. However, these efforts have so far been tailored to specific tasks, workflows, and tools.
Exploring Unknown Physical Models: From Classical Mechanics to Wave Physics to Quantum Many-Particle Physics
SciExplorer, by contrast, takes a generalist approach and is not limited to specific tasks or systems. In addition, this agent-based AI scientist works with a minimal set of generic instructions and code-based general-purpose tools. It generates an autonomously running workflow that, much like its human counterpart, develops a solution using heuristics through a multi-step process. Starting with minimal information about the system under investigation, it selects suitable experiments within the simulation unaided, analyzes, and visualizes their results. It also formulates hypotheses, which it then tests. What is remarkable about SciExplorer is its ability to uncover the underlying models of a wide variety of physical systems – ranging from classical mechanics to wave physics to quantum many-body systems.
For example, it achieves strong performance on tasks such as recovering equations of motion from observed dynamics and inferring Hamiltonians from expectation values.
“Because a large variety of modern physics experiments are controlled through code-based interfaces, in the future SciExplorer could be directly applied to laboratory settings. This could be helpful for complex fluids, cold atomic gases, strongly correlated electronic and spin systems, or quantum simulators,” says Maximilian Nägele, a doctoral student and first author. “We are now living in an era in which Artificial Intelligence can solve expert-level scientific questions, without having been trained on specific tasks. The evolution is rapid, and in the next few years we will experience important breakthroughs in science based on these techniques,” adds Florian Marquardt.
Original Publication in Physical Review X
M. Nägele, F. Marquardt, “Agentic Exploration of Physics Models”,
Phys. Rev. X 16 (2026)
DOI: doi.org/10.1103/xnqc-q6nt
Scientific contact
Max Planck Institute for the Science of Light, Erlangen
Prof. Florian Marquardt
Division ›Theory‹
www.mpl.mpg.de / florian.marquardt@mpl.mpg.de
