Finite-time cooling and accessibility of the stripe phase in the Ising antiferromagnet

What the math gives to ML

The paper identifies a transferable finite-time mechanism: local dynamics can restore local energetic constraints while failing to select a globally coherent ordered sector. The relevant control variable is not only instantaneous loss or local consistency, but also a coarsening length whose growth is slow enough that competing domains survive; the reported kinetic ordering time grows at least quadratically with system size over the simulated range. This suggests a neural-network training schedule that separately monitors local update consistency and global mode coherence, slowing learning when local optimization has occurred but global sectors remain fragmented.