They used molecular dynamics (MD) simulation to train the model - couldn't the correlation length be calculated from the MD simulations directly without the graph network to gain the same insight?