On the generalizability of artificial neural networks in spin models

The applicability of artificial neural networks (ANNs) is typically limited to the models they are trained with and little known about their generalizability, which a pressing issue in practical application ANNs unseen problems. Here, by using task identifying phase transitions spin models, we establish systematic generalizability such that simple two-dimensional ferromagnetic Ising model can be applied $q$-state Potts different dimensions for $q \geq 2$. same scheme highly nontrivial antiferromagnetic model. We demonstrate similar results obtained reducing exponentially large state space spanned training data one comprises only three representative configurations artificially constructed through symmetry considerations. expect our findings simplify accelerate development machine learning-assisted tasks spin-model related disciplines physics materials science.