NaNs generally are inconsistent - because sampling involves randomness, sometimes a ray that causes a NaN will spawn, sometimes not. Generally they are caught when an animation is done, because then on some frames a NaN will happen (but, never on the same frames :) Just by repeating rendering the scene, the randomness will eventually show up). As a note, the NaN is not in the scene itself, but in the calculations done by Corona - thus, there isn't a "problem pixel" in the scene that is a NaN, but somewhere along the way a calculation in Corona results in something that can't be computed (you can take division by zero as the classic example).
Best thing to do in such cases is send us the scene, because if it happened once, it can happen again, depending on the "roll of the dice", and it's always an error in how Corona is calculating something so is something we really want to investigate. Thanks!