This sounds a bit like allowing each neuron's function to perform its own symbolic regression? But at predicting physical phenomena you might get better perform

This sounds a bit like allowing each neuron's function to perform its own symbolic regression? But at predicting physical phenomena you might get better performance per cycle from just an A-Life swarm of symbolic regression cells competing than trying to harness them as a single organism. Why do you need a NN to model what's basically a deterministic result set, and why is that a good test?