Abstract
Machine learning methods are powerful in distinguishing different phases ofmatter in an automated way and provide a new perspective on the study ofphysical phenomena. We train a Restricted Boltzmann Machine (RBM) on dataconstructed with spin configurations sampled from the Ising Hamiltonian atdifferent values of temperature and external magnetic field using Monte Carlomethods. From the trained machine we obtain the flow of iterativereconstruction of spin state configurations to faithfully reproduce theobservables of the physical system. We find that the flow of the trained RBMapproaches the spin configurations of the maximal possible specific heat whichresemble the near criticality region of the Ising model. In the special case ofthe vanishing magnetic field the trained RBM converges to the critical point ofthe Renormalization Group (RG) flow of the lattice model. Our results suggestan alternative explanation of how the machine identifies the physical phasetransitions, by recognizing certain properties of the configuration like themaximization of the specific heat, instead of associating directly therecognition procedure with the RG flow and its fixed points. Then from thereconstructed data we deduce the critical exponent associated to themagnetization to find satisfactory agreement with the actual physical value. Weassume no prior knowledge about the criticality of the system and itsHamiltonian.