Speaker: DinhDuy Vu (Harvard) Title:Â How to build your neural networks? Abstract:Â Neural networks (NNs) form the backbone of modern artificial intelligence. In physics, they have been applied to a wide range of tasks, from optimization and classification to high-accuracy simulations. A central question arises: is it sufficient to apply a general-purpose NN directly to physical problems, or must the architecture be tailored to the physics at hand? In this talk, I will argue for the latter. By embedding physical intuition into the NN design, one can achieve significantly enhanced performance. As a case study, I will discuss the use of approximately gauge-symmetric NNs to characterize the dynamical preparation of quantum spin liquids. Our approach enables simulations on lattices of up to 384 atoms, revealing that the prepared state resembles a non-equilibrium finite-size quantum spin liquid. In particular, we find that the topological entanglement entropy is quantized---but only at specific length scales---thus setting the optimal system size and sweeping rate. We further examine the stability of the prepared states using the same NN-based framework. If time allows, I will also share recent developments on incorporating "path information" into NNs, an approach that boosts performance in systems with nontrivial phase structures such as non-Abelian gauge theories and frustrated antiferromagnets.