Transfer Learning For Many-Body Systems

Abstract

We study the ground-state properties of fully-polarized, trapped, one-dimensional fermions interacting through a Gaussian potential using a variational wavefunction represented by an antisymmetric artificial neural network. We optimize the network parameters by minimizing the energy of a four-particle system with machine learning techniques. We introduce several methods to enhance the efficiency, accuracy, and scalability of the artificial neural network approach. We use Early Stopping to terminate the training when the energy converges. We apply Transfer Learning to pre-train the network with different interactions or particle numbers to reduce the memory cost and improve the performance. We employ a re-weighting scheme to speed up the Metropolis- Hastings sampling for estimating the energy expectation value. We show that, in particular, this re-weighting scheme can speed up the training time by up to a factor of 10.