--- title: "Self-Replicating Neural Cellular Automata Make Digital Diversity Measurable" description: "A September preprint separates phenotypic from genotypic diversity across self-replicating neural cellular automata, including 24 long 200 × 200 runs." date: 2026-09-18 topic: artificial-life tags: ["neural-cellular-automata","self-replication","digital-ecosystems","evolutionary-computation","emergence"] production: ai-assisted source: https://notes.suffro.com/posts/self-replicating-neural-cellular-automata-diversity --- # Self-Replicating Neural Cellular Automata Make Digital Diversity Measurable A new preprint on self-replicating neural cellular automata asks a useful artificial-life question: if a digital ecosystem can reproduce, mutate, compete, and persist, how should its diversity actually be measured? In [*Self-Replicating Neural Cellular Automata: Quantifying Emergent Phenotypic and Genotypic Diversity in an OpenEnded Substrate*](https://arxiv.org/abs/2609.19902), submitted on September 17, Sanyam Jain, Felix Simon Reimers, and Stefano Nichele study an in-silico cellular automaton in which every cell carries a small neural network. The work is less about claiming that the system is “alive” than about making two properties of an evolving digital population observable: what cells look and behave like, and how different their inherited neural parameters are. That distinction matters because an artificial-life system can appear visually diverse while becoming genetically uniform, or maintain many genotypes that collapse into similar phenotypes. Treating those as separate quantities gives the authors a way to study persistence and evolutionary structure rather than relying on attractive animations alone. ## Reproduction is built into the substrate The simulated world is a two-channel cellular-automata grid. Each pixel corresponds to a cell and is associated with a small neural network that senses its local Moore neighborhood. Cells do not persist indefinitely. Reproduction is the mechanism that keeps a lineage present: a living neighbor is cloned and the copied neural-network weights can be perturbed by mutation. The cell state is therefore the phenotype, while the network weights act as the genotype. This is an important operational choice. Reproduction itself is not something that evolves from arbitrary instructions. The substrate already defines how a cell is copied and how mutations enter the lineage. What evolution can change is the inherited neural rule that determines local behavior, which in turn affects whether descendants survive and spread through the grid. The project’s [public implementation](https://github.com/s4nyam/Self-Replicating-NCA), originally developed as Jain’s master’s-thesis codebase, makes the setup concrete. It exposes parameters for inheritance probability, parameter perturbation, cell lifetime, activation functions, initial population density, and grid size. The repository also contains the large-run configurations and released experimental archives used to study longer trajectories. ## The system produces ecological structure Starting from a small seeded population, the simulations develop spatially organized populations that the authors describe as coexisting, competing, and sometimes dominant species. The paper reports a five-fold sweep covering 1,680 small runs, followed by 24 long configurations on 200 × 200 grids for 1,000 generations. Twenty of those 24 long configurations remained persistent and self-maintaining. That is evidence of robust persistence across many tested settings, but it is not evidence that persistence is guaranteed. Four long configurations did not meet that outcome, and the result is specific to the tested neural cellular-automata rules and parameter ranges. The authors use “species” as an operational category inside the simulation, not as a claim of biological speciation. Its value is analytical: it provides a way to describe lineages and population structure as the grid evolves. ## Phenotype and genotype do not move together The main contribution is the measurement framework. For phenotypic diversity, the project uses four coarse-grained tools based on quantities such as cell-type frequency, global entropy, cell variance, and local organization. These summarize how heterogeneous the visible cellular states and spatial patterns become over time. For genotypic diversity, the authors compare methods that map neural-network weights to lineage-like identifiers. One samples a subset of weights, while another hashes the full parameter vector. The full-genome hash reveals lineage structure that the sparse probe can miss. Across the experiments, the authors report a trade-off: parameter settings that increase phenotypic diversity tend to reduce genotypic diversity, and settings that preserve more genotypic variation tend to produce less phenotypic diversity. That result is more interesting than a simple “more diversity is better” story. It suggests that diversity in an artificial ecosystem is not one scalar property. A population can explore many internal genotypes without producing correspondingly distinct macroscopic states, or converge genetically while still expressing varied spatial behavior. The study does not establish that this trade-off is universal. It is an empirical pattern in this substrate and parameter sweep. Other mutation operators, update rules, network architectures, or ecological constraints could change the relationship. ## What counts as evolution here Several artificial-life concepts are explicit enough to be separated cleanly. **Reproduction** is the cloning of a living neighboring cell into a new location. **Heredity** is carried by the copied neural-network weights. **Mutation** is introduced by perturbing those weights during inheritance. **Evolution** is operationalized as changes in inherited network parameters and lineage composition across generations as cells reproduce, disappear, and compete spatially. **Emergence** refers to population-level structures and diversity patterns that arise from local cell interactions rather than from a global controller specifying the final ecosystem. Those definitions make the system scientifically useful without requiring a stronger claim that the simulated cells are literally alive. ## Why the result matters Artificial-life research often faces a measurement problem. It is relatively easy to show a simulation in which intricate patterns appear. It is harder to say whether the system is maintaining heritable variation, producing genuinely different phenotypes, preserving lineages, or merely generating visual complexity. This work attacks that problem directly. By measuring genotype and phenotype separately, it gives researchers a more disciplined way to compare long-running digital ecosystems and to detect cases where one kind of diversity is masking the collapse of another. The public code and experimental archives also make the substrate more useful than a one-off visual demonstration. The repository predates the new preprint and contains the simulation, analysis tools, and configurations for the long experiments, so the September paper is best understood as a formalization and larger empirical treatment of an existing research system. The strongest conclusion is therefore methodological rather than metaphysical. The experiments show that a self-replicating neural cellular automaton can sustain structured populations across many tested conditions, and that its evolutionary diversity looks different depending on whether it is measured in cell states or inherited neural parameters. That gives artificial-life work a more precise vocabulary for asking what, exactly, is changing when a digital ecosystem appears to evolve. ## Sources - [Jain, Reimers, and Nichele, “Self-Replicating Neural Cellular Automata: Quantifying Emergent Phenotypic and Genotypic Diversity in an OpenEnded Substrate”](https://arxiv.org/abs/2609.19902) - [Self-Replicating-NCA implementation and experiment repository](https://github.com/s4nyam/Self-Replicating-NCA)