Firstly, I think it wouldn't be useful when compared to existing algorithms... Cellular automata can be quite heavy to process.
I suspect the problem you will run into this that you cannot reliably say that additional variations will for sure result in significant changes in n ticks of the simulation.
What could be interesting though is Langton's Ant, as you can for sure make guarantees that the next iteration will result in some change in the output - and it should in theory always be entirely reversible if you want it to be. You could easily make it non-reversible by presenting some sub-set of the space for use in verification for example.