Teaching AI cells to self-organize using learned developmental blueprints

Self-organising systems like cellular automata can grow complex patterns from local rules, but biological development relies heavily on pre-existing spatial cues (maternal morphogen gradients) to bootstrap the process. This work pairs a Neural Cellular Automaton with a learned coordinate-based pattern generator (SIREN) and trains both simultaneously, letting the model discover what information to embed in initial conditions versus what to compute on the fly. Information-theoretic analysis shows the joint approach outperforms purely self-organising baselines in robustness, encoding capacity, and symmetry-breaking. Notably, effective learned pre-patterns don't merely approximate target outputs — they reshape developmental dynamics to make convergence easier, mirroring a non-trivial memory-compute trade-off seen in biology.