Artificial Scientist Lab

Experiments are our windows to the universe. Yet, the space of all possible experiments is enormously large. Did humans already find all useful experiments, or are there yet undiscovered but exceptional experimental ideas that can lead to new ways to explore the world?

We can split this question into four pillars: extremely large and complex search spaces, fast and reliable simulators, meaningful objective functions, and clever AI-exploration algorithms.

AI-designed Quantum Experiments: We started using AI for the design of physics experiments in 2014, published in 2016, where our first program Melvin discovered numerous new experiments for high-dimensional multipartite quantum entanglement, several of which were later built in laboratories: Nature Photonics 2018, Nature Photonics 2016, and Phys. Rev. Lett. 2017. These solutions contained surprising results, and we were able to conceptually understand several of them, for example an entirely new way to create photonic quantum entanglement, denoted as Entanglement by Path Identity, as well as a new bridge between quantum optics and graph theory that led to the discovery of new interference effects.

Since then we have developed many further methods. PyTheus (spearheaded by Carlos Ruiz González and Sören Arlt) is an algorithm for designing vastly diverse quantum experiments, for quantum state generation, the design of single- and multi-photon transformations, and new communication protocols. One surprising new discovery, a new way to entangle independent photons, has been experimentally implemented by the experimental group of Xiaosong Ma in Nanjing, China, a former PhD colleague of Mario from Anton Zeilinger's lab: Phys. Rev. Lett. 2024.

AI-designed microscopes: Driven by human ingenuity and creativity, the discovery of super-resolution techniques, which circumvent the classical diffraction limit of light, represents a leap in optical microscopy. However, the vast space of possible experimental configurations suggests that some powerful concepts and techniques might not have been discovered yet, and might never be discovered through a direct human design approach. AI-based exploration techniques could therefore provide enormous benefits by exploring this space in a fast, unbiased way. We developed XLuminA (spearheaded by Carla Rodriguez), an open-source computational framework built with JAX, a high-performance computing library for Python. XLuminA speeds up simulation by four orders of magnitude, allowing us to explore the space of possible microscope concepts much faster than before.

AI-designed Gravitational Wave Detectors: Gravitational waves are created by some of the most extreme events in the universe, such as the collision of black holes or the explosion of stars. These ripples of space-time then propagate through space towards Earth. When they reach us, they are extremely faint signals. It took 100 years since their prediction by Einstein to detect them. That became possible in 2016 through the international LIGO collaboration. The experiment built by LIGO is an interferometric system based on Michelson's interferometer, designed by ingenious human scientists. The question is: did humans find the best way to measure gravitational waves, or are there practical new experiments that are significantly more sensitive? Together with Rana Adhikari and Yehonathan Drori from the Caltech LIGO Lab, we are exploring this question using AI. We discovered more than 50 blueprints that, at least theoretically, outperform the best previous setups. We spent months exploring the best-performing solutions, and while we were able to extract simple conceptual cores from some of them, we were unable to understand the big picture behind most of the solutions. This indicates an important challenge for the future of AI-driven scientific discovery.

Similar to the microscopy case, we took the original simulator for gravitational-wave optics, Finesse, developed by Andreas Freise's group, and reproduced its crucial core in JAX, which provided an enormous speed-up. The simulator, Differometor, was developed by Jonathan Klimesch.

If an AI discovers solutions that outperform all human solutions, it must contain new tricks and ideas that we could learn from. As scientists, we want to understand these new tricks and concepts discovered by the machine.

The first time we faced this problem was in a case where our algorithm Melvin discovered new high-dimensional multipartite entangled systems that seemed to be more strongly entangled than previously thought possible. After analyzing the solutions for weeks, we found that it had discovered an entirely new way to generate quantum entanglement in photonics. We described this technique as Entanglement by Path Identity. In another instance, we investigated entanglement generation with PyTheus. We discovered that PyTheus invented a new trick to generate high-dimensional entanglement: it discovered a structure resembling a probabilistic multi-photon emitter just by building two-photon emitters. When we understood this technique, we were immediately able to generalize it to other situations. This new principle, spearheaded by Sören Arlt, was published as Emulating multiparticle emitters with pair-sources: digital discovery of a quantum optics building block.

A crucial step is the idea of Meta-Design (see the figure above, pioneered by Sören Arlt). Here, instead of designing solutions for individual questions and trying to understand them afterwards, we trained a language model that produces Python code. The resulting Python code solves whole classes of questions at once. Instead of having to analyze each solution independently, we can inspect the generated Python code, which contains the main principle of a generalizable solution.

Inspired by the work of computational sociologists and network theorists, for example James Evans, Jacob Foster, Albert-Laszlo Barabasi, and many others, who showed how to compress information from millions of scientific papers into knowledge graphs and make quantitative statements about the research community as a whole, we aimed to do the same for quantum physics. Thus, in 2020, we developed SemNet, where we built a knowledge graph for quantum physics based on 650,000 papers written over 100 years. This evolving knowledge graph allowed us to predict what researchers will work on in the future, based on what they worked on in the past. The predictive performance was surprisingly high. Three years later, we ran an AI competition called Science4Cast on a closely related topic, the prediction of new topic combinations from a knowledge graph, with similarly strong predictive performance.

An important question is whether one can also predict which research directions that have never been investigated before will be of high impact in the future. For that, spearheaded by Xuemei Gu, we developed Impact4Cast, a knowledge-graph system augmented with citation information, which uses supervised neural networks to predict which future works will have high impact.

Ultimately, we are interested in whether these predictions could actually be useful. For that, we launched the project SciMuse: we performed a large-scale interdisciplinary evaluation of research ideas generated using connections between knowledge graphs and LLMs and evaluated them with more than 100 research group leaders in the Max Planck Society. Doing so produced a dataset of more than 4,500 human-expert-evaluated research ideas, which we can use to predict the interest level of an idea before it is even fully defined. This will be a crucial capability of future artificial scientists.

We have developed AI-Mandel, an agent system that can autonomously generate scientific ideas and then execute these ideas using intelligent tools such as PyTheus. To demonstrate that it works, we used two outputs of AI-Mandel and wrote scientific papers about their results: Automated Discovery of Non-local Photonic Gates and Automated discovery of high-dimensional multipartite entanglement with photons that never interacted.

The quest to create artificial scientists comes with important questions about science itself. For example, what should we do if we can no longer understand AI-generated solutions? What does it even mean to "understand" in a scientific context? Addressing this question, we wrote a perspective paper called On scientific understanding with artificial intelligence, based on dozens of comments from domain experts in biology, chemistry, physics, and AI research. Specifically, the paper classifies algorithms into three categories, Computational Microscope, Artificial Muse, and Agent of Understanding, and shows how AI can extend scientific understanding.

Most recently, in Philosophy of Autonomous Science: Ten Questions for the Coming Age of Artificial Scientists, we argue that the move from powerful AI tools to autonomous scientists is not only a technical challenge, but also a philosophical and conceptual one. The Philosophy of Autonomous Science (PAS) program asks how core epistemic aims such as understanding, curiosity, surprise, interest, creativity, and novelty can be translated into computable, nonanthropocentric objectives for artificial scientists. Drawing on our research experience, we outline ten questions for PAS and invite philosophers, scientists, and AI researchers to shape the principles for safe and successful autonomous scientific discovery.