Dresden 2026 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

DY: Fachverband Dynamik und Statistische Physik

DY 58: Focus Session: Physics of AI – Part II (joint session SOE/DY)

DY 58.1: Invited Talk

Friday, March 13, 2026, 09:30–10:00, GÖR/0226

What can we learn from neural quantum states? — Brandon Barton¹⁰, Juan Carrasquilla¹⁰, Anna Dawid⁹, Antoine Georges^3,6,7,8, Megan Schuyler Moss^1,2, Alev Orfi^3,4, Christopher Roth³, Dries Sels^3,4, Anirvan Sengupta^3,5, and •Agnes Valenti³ — ¹Perimeter Institute for Theoretical Physics, Waterloo — ²University of Waterloo, Waterloo — ³Flatiron Institute, New York — ⁴New York University, New York — ⁵Rutgers University, New Jersey — ⁶Collège de France, Paris — ⁷École Polytechnique, Paris — ⁸Université de Genève, Genève — ⁹Universiteit Leiden, The Netherlands — ¹⁰ETH Zürich, Switzerland

Neural quantum states (NQS) provide flexible parameterizations of quantum many-body wave-functions that serve as powerful tools for the ground-state search. At the same time, NQS offer something that standard machine-learning tasks and datasets fundamentally lack: a known underlying Hamiltonian and quantum-physics tools that allow direct examination of the encoded wavefunction. This additional structure makes NQS an interesting platform for probing the behavior of classical neural networks themselves. I will first show how pruning and scaling-law phenomena change when the learning task is the quantum wavefunction itself, and link effects depend on the underlying Hamiltonian. I will then discuss generalization and double descent through the lens of quantum observables, by analyzing how NQS fail at the interpolation threshold. Finally, I will discuss how these results relate back to practical consequences for training and architecture search in the context of the ground state search for quantum many-body systems.

Keywords: Neural quantum states; Machine learning; Quantum phase transitions; Overparameterization