Deep Machine Learning Investigation of Phase Transitions
We explore the possibilities of using neural networks to study
phase transitions. The main question is the level of accuracy which can
be achieved for the estimates of the critical point and critical exponents
of statistical physics models. We generate data for two spin models in
two dimensions for which analytical solutions exist, the Ising model and
Baxter-Wu model, which belong to the different universality classes. We
applied six neural networks with three different architectures to the data
and estimated the critical temperature and the correlation length exponent.
We find that the accuracy of estimation does depend on the neural
network architecture. The critical exponents of Baxter-Wu model are
estimated by the deep machine learning technique for the first time.