| dc.description.abstract | Neural networks, a class of artificial intelligence techniques that extract patterns from data and are suitable for tasks such as function approximation and image classification, are finding applications in computational condensed matter physics, where simulation data can be reliably generated in large quantities. For example, they can classify phases in classical and quantum lattice models using Monte Carlo simulation data. In this application, one of their useful characteristics is their generalization capacity: their ability to identify a similar phase transition when the model Hamiltonian is altered. A controversial point in the theory of spin glasses, systems characterized by frustration and a complex energy landscape, is the existence of a spin-glass state in the presence of random fields. However, a Monte Carlo study by Young and Katzgraber has ruled this out for the 3D Edwards-Anderson model with Gaussian interactions. This suggests a excellent test for the generalization capacity of a neural network trained to identify the spin-glass state. In this work, a 3D convolutional neural network was implemented and taught to recognize the spin-glass state of the 3D Edwards-Anderson model using data from Monte Carlo simulations. The inference results collected are found to be consistent with the absence of a spin-glass state in a field. The strengths and weaknesses of the phase classifier are evaluated and further directions are discussed. | en |
