Optimizing Neural Quantum States with Decision Geometry

par Mehdi Drissi (TU Darmstadt)

vendredi 10 juil. 2026, 14:00 → 15:00 Europe/Paris

100/-1-A900 - Auditorium Joliot Curie (IJCLab)

Over the past few years, the capabilities of Neural Quantum States (NQS) to solve the quantum many-body problem have undergone a tremendous progression. The variety of Hamiltonians being tackled now spans a wide range of fields of physics such as condensed matter, quantum chemistry, and nuclear physics. Much of the progress have come from the development of new network architectures to efficiently capture the richness of quantum many-body states. At the same time, there have been recently a renewed interest in revisiting the optimization algorithms employed to adjust the parameters of NQS. Currently, the main approaches are based on variants of Adam and Stochastic Reconfiguration (SR), two algorithms originally designed in the context of supervised machine learning (ML) and standard Variational Monte-Carlo (VMC), respectively.

In my talk, I will discuss our on-going efforts to develop a new optimization algorithm based on Decision Geometry, namely the Decisional Gradient Descent (DGD). The goal is to take into account the specificity of NQS optimization problems from the ground-up, such as the particular cost function and the large number of parameters which differ from standard ML and VMC problems. After introducing Decision Geometry and DGD for a general Hamiltonian, I will mainly focus on NQS modelling spin lattices with a J1-J2 interaction as a test bench. Preliminary results will be discussed comparing the current performance of DGD against the state-of-the-art SR optimizer in terms of stability and speed of convergence.

--------------
Videoconference link: https://visio.numerique.gouv.fr/nwk-bmff-xny