The prototypical scientific process in many technical and natural sciences domains comprises ideation, definition of hypotheses, experimental design, execution, evaluation, and iteration. This scientific process has often been understood to constitute the pinnacle of intellectual achievement, and also a blueprint for the development of efficient and effective business processes. In this talk, I want to sketch an overarching picture of the use of AI in science with particular results from the Stuttgart Center for Simulation Science – and beyond. 

Prof. Dr. Steffen Staab heads the Institute for Artificial Intelligence at the University of Stuttgart, Germany, and holds a chair for Web and Computer Science at the University of Southampton, UK. Coming from a research background in computational linguistics, Staab is interested in foundations and usage of semantic technologies investigating the likes of knowledge graphs, machine learning, natural languge processing and databases. Prof. Staab is a fellow of the Association for Computing Machinery (ACM), the European AI Association (EurAI), the Asia-Pacific AI Association (AAIA), and ELLIS – the European Laboratory for Learning and Intelligent System Society.

Artificial intelligence (AI) is a potentially disruptive tool for physics and science in general. One crucial question is how this technology can contribute at a conceptual level to help acquire new scientific understanding or inspire new surprising ideas. I will talk about how AI can be used as an artificial muse in physics, which suggests surprising and unconventional ideas and techniques that the human scientist can interpret, understand and generalize to its fullest potential [1]. I will focus on AI for the design of new physics experiments, in the realm of quantum-optics [2, 3] and quantum-enhanced gravitational wave detectors [4] as well as super-resolution microscopy [5].

[1] Krenn, Pollice, Guo, Aldeghi, Cervera-Lierta, Friederich, Gomes, Häse, Jinich, Nigam, Yao, Aspuru-Guzik, On scientific understanding with artificial intelligence. Nature Reviews Physics 4, 761 (2022).

[2] Krenn, Kottmann, Tischler, Aspuru-Guzik, Conceptual understanding through efficient automated design of quantum optical experiments. Physical Review X 11(3), 031044 (2021).

[3] Ruiz-Gonzalez, Arlt, et al., Digital Discovery of 100 diverse Quantum Experiments with PyTheus, Quantum 7, 1204 (2023).

[4] Krenn, Drori, Adhikari, Digital Discovery of interferometric Gravitational Wave Detectors, in press: Phys. Rev. X (2025), (https://arxiv.org/abs/2312.04258)

[5] Rodríguez, Arlt, Möckl, Krenn, Automated discovery of experimental designs in super-resolution microscopy with XLuminA, Nature Comm. 15, 10658 (2024)

Dr. Mario Krenn is the research group leader of the Artificial Scientist Lab at the Max Planck Institute for the Science of Light in Erlangen, Germany, and will soon start as a full professor for “Machine Learning for Science” at the University of Tuebingen in the Department of Computer Science.
Dr. Krenn has introduced AI systems that design quantum experiments and hardware, several of which have been realized in laboratories, and developed algorithms to inspire unconventional ideas in quantum technologies. His ERC Starting Grant project, ArtDisQ, aims to transform physics simulators to accelerate the discovery of advanced quantum hardware.

Quantum machine learning holds significant promise for enhancing various aspects of machine learning, including sample complexity, computational complexity, and generalization. The field has made substantial strides in recent years. However, a key objective—developing quantum algorithms that clearly outperform classical methods for practically relevant unstructured data—remains elusive. In this talk, we will explore this challenge from multiple perspectives. We will examine cases where separations can be identified, such as in abstract instances of generator [1] and density modeling [2], in training classical networks using quantum algorithms [3], for short quantum circuits [4], and for quantum analogs of diffusion probabilistic models [5]. At the same time, we will address challenges arising from dequantization in both noise-free [6] and non-unital noisy settings [7]. These insights will also encourage thinking beyond traditional approaches. We will reconsider the concept of generalization [8] and explore examples of explainable quantum machine learning [9] and single-shot quantum machine learning [10]. Ultimately, we will use these insights to reflect on the potential and limitations of applying quantum computers to machine learning problems involving unstructured noisy data and put this research line into the wider context of the intersection of research on artificial intelligence and quantum computing [11].

[1] Quantum 5, 417 (2021).

[2] Phys. Rev. A 107, 042416 (2023).

[3] Nature Comm. 15, 434 (2024).

[4] arXiv:2411.15548 (2024).

[5] arXiv:2502.14252 (2025).

[6] Quantum 9, 1640 (2025).

[7] arXiv:2403.13927 (2024).

[8] Nature Comm. 15, 2277 (2024).

[9] arXiv:2412.14753 (2024).

[10] Quantum 9, 1665 (2025).

[11] arXiv:2505.23860 (2025).

Jens Eisert is a theoretical physicist at the Freie Universität Berlin, working on notions of quantum computing and the study of complex quantum systems. He is known for his rigorous work in quantum information science, which spans the traditional fields of physics, mathematics, and computer science, and is pragmatically motivated by application, protocol, or phenomenon. In the past, he has made several important contributions to the field of quantum computing, focusing on what quantum computers can do and their ultimate limitations, making him one of the most seen and cited researchers in his field. Before joining the Freie Universität Berlin, he was a full professor at the University of Potsdam and an assistant professor at Imperial College London. For his work, he has received an ERC AdG, an ERC CoG, a EURYI Award (the predecessor of the ERC StG), and a Google NISQ Award.

In the first part of this talk, Christine Muschik will outline avenues on which modern machine learning techniques can benefit the development of quantum technologies. In the second part, she’ll cover a concrete example of such an application and explain how  AI-assisted quantum sensors can play an important role in protecting our drinking water.

Prof. Christine Muschik holds a University Research Chair at the University of Waterloo. She holds a faculty position at the Institute for Quantum Computing and an associate faculty position at the Perimeter Institute for Theoretical Physics.
Christine Muschik received a number of awards for her work on quantum computers, including an Ontario Early Researcher Award, a CIFAR Azrieli Global Scholar Fellowship for “Research Leaders of Tomorrow”, and a Sloan Fellowship for outstanding early career researchers.
Her work on taming quantum systems to perform hard calculations is internationally recognised and has been featured by Scientific American and Forbes. In 2016, her results on programming quantum computers has been named as one of the “Top 10 breakthroughs in Physics” of this year.
Today she pushes the frontier of scientific computing by developing new quantum-enhanced methods.

Artificial neural networks (NNs) have revolutionized language translation and image generation and may similarly transform quantum technologies. In particular, NNs show strong potential to outperform traditional numerical methods in learning Hamiltonians that govern quantum systems based on experimental data. This promise is especially notable in architectures like graph neural networks (GNNs), which are inherently invariant to system size. This allows them to infer Hamiltonians for quantum systems that are larger or structurally different from those seen during training, potentially reaching regimes inaccessible to conventional approaches such as tensor networks.
In this work, we explore this potential in the context of Rydberg-atom arrays—a leading platform for quantum simulation. The experimental control of these arrays is limited by the imprecision in the positions of optical tweezers during array assembly, introducing uncertainty in the realized Hamiltonian. To model this, we use the Density Matrix Renormalization Group (DMRG) to generate ground-state snapshots of the transverse field Ising model for many realizations of Hamiltonian parameters. Correlation functions reconstructed from these snapshots serve as input data for training the GNN.
We demonstrate that our model exhibits a remarkable ability to extrapolate beyond its training domain, both in system size and geometry. It accurately infers the underlying Hamiltonian parameters, marking a significant step toward bridging the gap between simulation and experiment in large-scale quantum systems.

Anna Dawid leads a research group as an assistant professor at the Leiden Institute of Advanced Computer Science (LIACS) and the Leiden Institute of Physics (LION) at Leiden University in the Netherlands. Her scientific interests include interpretable machine learning for science, ultracold platforms for quantum simulation, and the theory of machine learning. She is passionate about transforming automated computational methods (especially neural networks) into a new, unique lens for gaining fresh insights into difficult, well-established scientific problems.
Before joining Leiden, she was a research fellow at the Center for Computational Quantum Physics at the Flatiron Institute in New York. In 2022, she earned her PhD in physics and photonics under the supervision of Prof. Michał Tomza (Faculty of Physics, University of Warsaw) and Prof. Maciej Lewenstein (ICFO – The Institute of Photonic Sciences, Spain). Earlier, she completed a master’s degree in quantum chemistry and a bachelor’s degree in biotechnology at the University of Warsaw.
She is the first author of the book Machine Learning in Quantum Sciences, to be published by Cambridge University Press in May 2025. She is also a recipient of the START fellowship from the Foundation for Polish Science (2022) and a participant in the 74th Lindau Nobel Laureate Meeting (2024).

Since 2008 and the advent of t-SNE, the domain of dimensionality reduction has undergone a paradigm shift.
Previous approaches to project data, like principal component analysis, or to embed data based on distance preservation, like multidimensional scaling, can hardly compete with t-SNE or its closely related variants such as UMAP.
The paradigm of t-SNE is to preserve neighbourhoods of points from the data space to the visualisation space, instead of distances, and t-SNE stands for t-distributed stochastic neighbour embedding.
In fact, t-SNE is packed with counter-intuitive empirical tweaks and their theoretical understanding is still work in progress.
Immunity against distance concentration and a strong inductive bias towards cluster formation have made t-SNE an extremely useful tool for visualisation in various applications domains, like cell biology.
However, t-SNE suffers from a few rare drawbacks, like its poor preservation of global structure, fragmentation of clusters, or complicated hyper-parameterisation.
In addition to analysing t-SNE, we will present some recent developments that might improve t-SNE or, alternatively, rehabilitate other simpler paradigms of dimensionality reduction, like multi-dimensional scaling.

John Aldo Lee was born in 1976 in Brussels, Belgium. He received the M.Sc. degree in Applied Sciences (Computer Engineering) in 1999 and the Ph.D. degree in Applied Sciences (Machine Learning) in 2003, both from the Université catholique de Louvain (UCL, Louvain-la-Neuve, Belgium). He is currently a Research Director with the Belgian fund of scientific research (Fonds de la Recherche Scientifique, F.R.S.-FNRS). During several postdocs, he developed specific image enhancement techniques for positron emission tomography in the Centre for Molecular Imaging and Experimental Radiotherapy of the Saint-Luc University Hospital (Belgium). He is also an active member of the UCL Machine Learning Group, with research interests focusing on data visualization, dimensionality reduction, neighbor embedding, intrinsic dimensionality estimation, clustering, and image processing. John A. Lee is currently the head of the center MIRO (Molecular Imaging, Radiotherapy, and Oncology) in UCLouvain Brussels. His team activities are aimed at image-guided treatment planning in radiation oncology and particle therapy. This includes adaptive treatments, motion management, image registration, AI-based automatic organ segmentation and dose prediction, as well as Monte Carlo simulations for proton therapy, robustness analysis, robust planning, and planning for non-conventional delivery techniques like arc proton therapy.

Jason Pye (Nordita (Nordic Institute for Theoretical Physics))

Out-of-Time-Ordered Correlators (OTOCs) have become important theoretical tools to probe the spreading of quantum information (scrambling) in many-body systems. They have been central to recent discussions in quantum chaos, thermalization, many-body localization and scarring, as well as black hole physics. However, computing the OTOC for a general quantum system is prohibitively expensive for classical computers, whereas simulation using current noisy quantum devices becomes difficult beyond the short time regime. Here we present a machine learning approach for approximating OTOCs. We explore four parameterized sets of Hamiltonians describing local one-dimensional quantum systems of interest in condensed matter physics. We frame the problem as a regression task, by sampling parameter values and generating small batches of labeled data with classical tensor network methods for systems with up to 40 qubits. Using this data, we train a variety of standard kernel machines and demonstrate that they can accurately learn the OTOC as a function of the Hamiltonian parameters. This method leads to an overall reduction in computational costs if one is interested in understanding the behaviour of the OTOC over large regions of a Hamiltonian parameter space. It allows one to perform a small number of evaluations of the OTOC and then use kernel methods to approximate the OTOC for the remaining parameter values of interest, which can be evaluated more efficiently than continued uses of tensor network methods.

Based on https://doi.org/10.1103/PhysRevB.111.144301

Manuel Lautaro Hickmann (Institute for AI Safety and Security – German Aerospace Center (DLR))

In our work we investigated the prospect of using hybrid quantum tensor network based algorithms for aeroelastic problems, leveraging synergies between machine learning and quantum computing. Quantum tensor networks for Machine Learning can be realized by Variational Quantum Circuits using a tensor network inspired internal gate structure. We apply this method to a simplified aeroelastic configuration, aiming to estimate flutter stability and regress generating parameters from time series data.

Using a dimensionality reduction via classical 1D convolutional neural networks, we compress the data into an 8-dimensional feature vector, which is then used as input for our quantum neural network. Notably, both the classical and quantum parts are trained together in an end-to-end approach, allowing for joint optimization of the entire hybrid model. We achieved outstanding results for binary time-series classification (F1-score > 0.9) and promising results for regressing parameters from the time series.

Our findings demonstrate the potential of hybrid quantum tensor network-based algorithms for aeroelastic problems, although hyperparameter selection remains a challenge. This work contributes to the development of hybrid quantum machine learning (QML) for complex problems. For more details see: https://indico.qtml2024.org/event/1/contributions/177/

Alessio Paesano (JoS Quantum GmbH)

We propose considering Quantum Key Distribution (QKD) protocols as a use case for Quantum Machine Learning (QML) algorithms. We define and investigate the QML task of optimizing eavesdropping attacks on the quantum circuit implementation of the BB84 protocol. QKD protocols are well understood and solid security proofs exist enabling an easy evaluation of the QML model performance. The power of easy-to-implement QML techniques is shown by finding the explicit circuit for optimal individual attacks in a noise-free setting. For the noisy setting we find, to the best of our knowledge, a new cloning algorithm, which can outperform known cloning methods. Finally, we present a QML construction of a collective attack by using classical information from QKD post-processing within the QML algorithm.

Anant Agnihotri (Fraunhofer IAF)

We investigate the optimization of kernel generation for quantum support vector algorithms focused on data classification tasks. To enhance classification efficiency, we implement classical post-processing techniques. Our approach begins with preprocessing high-dimensional data using Principal Component Analysis (PCA), which effectively reduces dimensionality while retaining essential features that contribute to the classification task.
Following the preprocessing step, we generate a training kernel using the ZZ feature map. In the subsequent post-processing phase, we utilize the overlap with all quantum states—not just the all-zero state, which is the standard practice in quantum kernel methods. This innovative approach allows us to compute kernel entries as a weighted sum of these overlaps, enabling us to determine the kernel values with fewer shots, thereby improving efficiency.
We implement our method on the MNIST dataset, aiming to differentiate between handwritten digits ‚0‘ and ‚1‘, as well as to distinguish between ‚even‘ and ‚odd‘ digits. In our analysis, we compare the kernel score, defined as the fraction of unseen data points accurately identified by the standard quantum kernel, against that of the kernel enhanced by our post-processing method. Our findings reveal that the post-processed kernel significantly outperforms the standard kernel, particularly when dealing with a higher number of qubits.

Dr. Max Scheerer (FZI)

Quantum computing holds immense promise, but developing quantum programs remains error-prone due to inherent hardware noise and algorithmic complexity. This work explores how generative AI (GenAI) can assist in improving the reliability of quantum programs. We examine two directions in which GenAI can contribute to improved reliability: automated bug detection and correction, and support for fault-tolerant execution. We evaluate the efficacy of GenAI using two benchmarks representative of these contexts.