Leopoldo Sarra, Ph.D.

I'm a Flatiron Research Fellow at the Flatiron Institute, and working in the Foundation Models for Science initiative. My current research interests lie at the intersection of physics and machine learning. In my research, I have been investigating how machine learning techniques can be used to support scientific discovery, exploring how a future "Artificial Scientist" could automatically learn from observations, understand relevant concepts, build new physical models and design new experiments similar to how human scientists do. I received my Ph.D. in Physics from the Friedrich-Alexander Universität in Erlangen, Germany, in association with the Max Planck Institute for the Science of Light. In the past, I earned my B.Sc. and M.Sc. degrees in Physics at Sapienza University of Rome, Italy, focusing on the statistical physics of spin glasses.


Polymathic AI

Currently, I am working on large foundation models for scientific data in the Polymathic AI team. We are exploring how to leverage scientific data from multiple sources to build AI models more generally suited for scientific applications. Please, refer to the group's website for the latest news about our efforts!

Below are some additional topics of my research:

Extraction of collective variables

One of the most useful concepts in physics is the notion of collective coordinates. However, it is not clear a-priori which low-dimensional function is best suited as a compact description of the high dimensional data. We investigated a technique to find these quantities based on their information content. We generalized the idea of mutual information, a solid concept in information theory, to work for this purpose, i.e. to quantify the amount of information between a random variable and a deterministic continuous function of it. In addition, we showed that it may be possible to also automatically extract collective variables by parametrizing features and optimizing over the parameters.

Bayesian Experimental Design and Active Learning

Bayesian experimental design allows to efficiently select measurements to characterize a physical system, by maximizing the expected information gain. Recent developments in deep neural networks and normalizing flows allow for a more efficient approximation of the posterior and thus the extension of this technique to complex high-dimensional situations. We investigated how this approach holds promise for adaptive measurement strategies to characterize present-day quantum many-body platforms. We focused on arrays of coupled cavities and qubit arrays, which both represent model systems of high relevance for modern applications, like quantum simulations and computing, and both have been realized in platforms where measurement and control can be exploited to characterize and counteract unavoidable disorder.

Discovering quantum circuit components with program synthesis

Despite rapid progress in the field, it is still challenging to discover new ways to take advantage of quantum computation: all quantum algorithms need to be designed by hand, and quantum mechanics is notoriously counterintuitive. We studied how AI, in the form of program synthesis, may help to overcome some of these difficulties, by showing how a computer can incrementally learn concepts relevant for quantum circuit synthesis with experience, and reuse them in unseen tasks. In particular, we focused on the decomposition of unitary matrices into quantum circuits, and showed how, starting from a set of elementary gates, we can automatically discover a library of new useful composite gates and use them to decompose more and more complicated unitaries.

Exploration in Reinforcement Learning

Scientists are agents that performs experiments to improve their model of the world. Their reward is mainly due to the observation of some new phenomenon and the improvement of their scientific model. In Reinforcement Learning, a similar situation happens when the reward is sparse, and it is useful to design strategies to explore the environment independently. Curiosity-driven and novelty-based techniques allow to intrinsically motivate the agent to explore the environment, in a similar way as a human scientist would do. We developed a particular novelty-based techniques, suitable for challenging exploration tasks like 3D-environments and Atari video games.

Some less recent work:

Device-independent tests of quantum channels

The device-independent approach for the characterization of a quantum system allows to draw conclusions about a system without the need of a precise knowledge of the measurement apparatus. We showed an implementation of this framework for the characterization of a quantum channel in a photonic setup.

Finite-size corrections to the mean-field spin glass

Finite-size corrections to the free energy of the Sherrington-Kirkpatrick spin glass in the low-temperature phase are a long-standing and still open problem. We investigated the role of one contribution to these corrections, the so-called longitudinal fluctuations, and determined its scaling exponent both with an analytical and numerical approach.


  • Leopoldo Sarra, Florian Marquardt "Deep Bayesian Experimental Design for Quantum Many-Body Systems" Machine Learning: Science and Technology 4 (4) 045022 arXiv DOI
  • Leopoldo Sarra, Kevin Ellis, Florian Marquardt "Discovering quantum circuits components with program synthesis" Machine Learning: Science and Technology 5 (2) 025029 arXiv DOI
  • Alaa Saade, Steven Kapturowski, Daniele Calandriello, Charles Blundell, Pablo Sprechmann, Leopoldo Sarra, Oliver Groth, Michal Valko, Bilal Piot "Unlocking the Power of Representations in Long-term Novelty-based Exploration", The Twelfth International Conference on Learning Representations (ICML 2024). Spotlight paper. arXiv OpenReview
  • Leopoldo Sarra, Andrea Aiello Florian Marquardt "Renormalized Mutual Information for Artificial Scientific Discovery" Physical Review Letters 126 (20), 200601 arXiv DOI
  • Iris Agresti, Davide Poderini, Gonzalo Carvacho, Leopoldo Sarra, Rafael Chaves, Francesco Buscemi, Michele Dall’Arno, Fabio Sciarrino "Experimental semi-device-independent tests of quantum channels" Quantum Science and Technology 4 (3), 035004 arXiv DOI
  • Giorgio Parisi, Leopoldo Sarra, Lorenzo Talamanca "Study of longitudinal fluctuations of the Sherrington–Kirkpatrick model" Journal of Statistical Mechanics: Theory and Experiment 2019 (3), 033302 arXiv DOI


  • Decomposing Quantum Unitaries into Circuits with Program Synthesis, APS March Meeting 2023, March 6th, 2023, Las Vegas, United States
  • Renormalized Mutual Information for Artificial Scientific Discovery, Machine Learning at Galileo Galilei Institute Workshop & Conference, September 9th, 2022, Florence, Italy
  • Renormalized Mutual Information for Artificial Scientific Discovery, Origins Data Science Labor Forum, June 17th, 2022, Munich,Germany/Online
  • Renormalized Mutual Information for Artificial Scientific Discovery, Learning to Discover workshop, April 28th, 2022, Paris/Online
  • Deep Bayesian Experimental Design for Quantum Many-Body Systems, APS March Meeting 2022, March 16th, 2022, Chicago, United States
  • Unsupervised feature extraction in simple Physical Models through Mutual Information Maximization, GDS Virtual Session - Deep Learning in Physics Short Talks, May 8th, 2020, Online
  • Towards Artificial Scientific Discovery, IMPRS Annual Meeting and Autumn Academy 2019, October 8th, 2019, Hersbruck, Germany
  • Study of Longitudinal Fluctuations of the Sherrington-Kirkpatrick spin glass, APS March Meeting 2019, March 7th, 2019, Boston, United States


I usually commit to open source the research code. Please visit my GitHub page for the updated list.

Name Description
Polymathic AI In this repository, we release the code of the Foundation Models for Science projects and related works
unitary-synthesis Discovering quantum circuit components with program synthesis
active-learning Bayesian experimental design for quantum many-body systems
rmi Renormalized mutual information for artificial scientific discovery