This is one of the only attempt to represent a CA rule as a CNN I have come across. The author uses a deep CNN to learn a rule and studies various information-theoretic quantities in the activation patterns to evaluate the complexity of the rules.
I am personally very interested by the paper since it is an interesting direction for creating neural-network based rules that can be sampled and efficiently stored and applied. This work is also cited in (Mordvintsev et al. 2020).