Using deep techniques learning to investigate phase transitions in physics

Deep learning techniques, particularly neural networks, are powerful tools for pattern recognition, often uncovering hidden patterns. It is well known they are effective for recognizing phase transitions, even for systems without an explicit order parameter. Conventionally, training is performed in the vicinity of the critical region. In this talk, I will consider 2D Ising, 3D Ising and Heisenberg models. Neural networks are trained on lattice configurations for these systems at extreme temperatures far from the critical point, specifically at T=0 and T=∞. Once trained, these ML models are tested on the models' configurations over a broad range of temperatures, generated through MCMC simulations. By treating the average predictions from the output layer as an effective order parameter, we can accurately locate the critical temperature. Furthermore, we apply histogram reweighting techniques to the neural network predictions to help extract critical exponents associated with these phase transitions.