Machine learning assisted analysis of neutron scattering: new insights into spin ice

The macroscopic manifolds of ground states in highly frustrated magnets are responsible for their rich physical behavior. Thermal and quantum fluctuations within these manifolds gives rise to liquid states with remarkable properties which are receiving intense interest. An open question is, what is the fate of these liquid states as the system cools? To gain insight into this we have undertaken a comprehensive study of Dy₂Ti₂O₇. Dy₂Ti₂O₇ is well known as a prototypical spin ice material that shows a U(1) gauge liquid behavior with magnetic monopole quasiparticles. From a combination of neutron scattering, magnetic noise, and thermodynamic measurements we have developed a model for the material using machine learning. The analysis involves use of autoencoders to identify phases and by optimizing in the network’s latent space highly accurate interaction parameters are extracted. A key part of this analysis is the ability to discriminate artifacts from physical signals in the neutron scattering data and to perform analysis on three-dimensional diffuse data sets. The extracted model is shown to reproduce glass formation in the material and provides microscopic understanding of a range of observations including the arresting of order as freezing occurs and 1/f^α magnetic noise. High performance computer modeling has also been used to map the development of short-range order and changes in dimensionality and topology in monopole pathways. Our results suggest ways in which spin glass formation could occur from simple interactions even without intrinsic disorder.