Flow Lenia: Mass-Conservative Open-Ended Evolution

• Flow Lenia is a mass-conservative, continuous-state, multi-channel cellular automaton that extends the Lenia framework to enable open-ended evolution and emergent complex patterns.
• It employs affinity-driven flow dynamics with localized, heritable rule parameters and a semi-Lagrangian reintegration scheme to ensure strict mass conservation.
• Experimental implementations show that Flow Lenia robustly generates self-replicating, migratory, and ecosystem-like behaviors across multi-species simulations.

Flow Lenia is a mass-conservative, continuous-state, multi-channel cellular automaton that extends the Lenia family to enable intrinsic open-ended evolution, multi-species interactions, and the robust emergence of spatially localized, adaptive, and complex patterns. By coupling affinity-driven flow dynamics with localized, heritable rule parameters, Flow Lenia constitutes a computational platform for studying artificial life phenomena such as autopoiesis, evolving morphogenesis, and ecosystem-like self-organization in high-dimensional, continuous media (Plantec et al., 2022, Plantec et al., 10 Jun 2025, Michel et al., 21 May 2025, Adachi et al., 2024).

1. Mathematical Structure and Update Rules

Flow Lenia operates on a finite or periodic two-dimensional lattice $L \subset \mathbb{Z}^2$, where each site $x$ holds a non-negative concentration vector over $C$ channels: $A^t(x) \in \mathbb{R}^C_{\ge 0}$ at time $t$. The model generalizes the classic Lenia framework by enforcing exact mass conservation and embedding update-rule parameters in the matter field itself.

Affinity Computation

At each step, an affinity map $U^t$ is computed per channel using a family of radial convolution kernels and non-linear growth functions: $$U^t_j(x) = \sum_{i=1}^K h_i(x)\, G_i\bigl( (K_i * A^t_{c^i_0})(x) \bigr) \cdot \mathbf{1}_{[c^i_1 = j]}$$ where $K_i$ is a (typically normalized) radial kernel, $G_i$ is a parametrized growth function (usually Gaussian-shaped), and $h_i(x)$ is the local weight (possibly variable due to parameter localization).

Flow and Mass-Advection

The evolution of matter is governed by the equation: $$F^t_j(x) = (1 - \alpha^t(x)) \nabla U^t_j(x) - \alpha^t(x) \nabla A^t_\Sigma(x), \qquad \alpha^t(x) = \bigl[ (A^t_\Sigma(x)/\theta_A)^n \bigr]_0^1$$ where $A^t_\Sigma(x) = \sum_{j=1}^C A^t_j(x)$, $\theta_A$ and $n$ set the scale and sharpness for diffusion onset.

Matter is then transported via a semi-Lagrangian "reintegration tracking" scheme, discretizing the continuity equation

$A_j^{t+\Delta t}(q) = \sum_p A_j^t(p)\, I_j(p, q)$

where $I_j(p, q)$ is the fraction of mass from source cell $p$ that arrives at destination cell $q$, ensuring global conservation $\sum_q A_j^{t+\Delta t}(q) = \sum_p A_j^t(p)$.

Parameter Localization and Mixing

Each cell carries a parameter vector $P^t(x)$ specifying local kernel weights (and potentially kernel radii, growth means, etc.; i.e., $P^t(x) \in \Theta \subset \mathbb{R}^{d}$), which governs local affinity calculation and thus matter flow. Parameters advect and mix along with the mass. When multiple sources contribute to the same cell, their parameters are aggregated using either weighted averaging or softmax sampling proportional to fluxes and masses: $$\mathbb{P}[P^{t+\Delta t}(q)=P^t(p)] \propto \exp\left( A^t_\Sigma(p) I(p, q) \right)$$ Optionally, random mutations (Gaussian noise) are injected at specified intervals or locations to enable evolutionary exploration of the parameter space.

2. Theoretical Foundations and Well-posedness

The original Lenia update can be formalized as a (possibly discontinuous) integro-differential evolution equation with hard state bounds. However, the vector-field formulation lacks the continuity required for standard Cauchy–Lipschitz theory due to the non-Lipschitz nature of the update at state boundaries. The rigorous existence and uniqueness of Flow Lenia dynamics are established instead via the theory of arc fields on metric spaces (Calcaterra et al., 2022).

In this setting, the CA rule is interpreted as a time-dependent arc field $X_t$, satisfying technical conditions (E1, E2, linear speed bound), guaranteeing the convergence of forward-Euler schemes to a unique continuous flow: $$X_t(f)(x) = \bigl[ f(x) + t\,G\bigl(K * f\bigr)(x) \bigr]_{0,1}$$ This approach generalizes well to unbounded densities as present in Flow Lenia, where state clipping is replaced by mass-conserving transport and advection.

3. Key Innovations: Mass Conservation and Parameter Localization

A central departure from original Lenia is Flow Lenia’s strict mass conservation:

• All matter transport is local and redistributive: At each step, the global sum of mass is algebraically preserved by construction via the reintegration tracking scheme, avoiding the typical explosive or dissipative regimes seen in vanilla Lenia.
• Parameter localization: The spatial rule field $P^t(x)$ is co-advected with matter, enabling the emergence of multi-species worlds in which each spatial region may be governed by different effective rules. Interactions at boundaries occur through mixing rules, simulating speciation, hybridization, and competition analogously to biological systems (Plantec et al., 2022, Plantec et al., 10 Jun 2025).

This mechanism allows for:

• Continuous, in-situ evolution of update laws.
• Multi-species ecosystems, symbiotic assemblies, and emergent intra-world genetics.
• Unbounded open-ended evolution, as local rules and morphologies co-evolve.

4. Experimental Exploration and Emergent Dynamics

Experimentally, Flow Lenia robustly generates a wide variety of spatially localized patterns (SLPs). In random search regimes, SLPs dominate the outcome space (≈80%), often exhibiting migratory, oscillatory, self-dividing, or self-replicating behavior (Plantec et al., 10 Jun 2025, Plantec et al., 2022).

Directed optimization—using evolutionary strategies (OpenES, GA)—produces robust behaviors such as directional locomotion, periodic turning, obstacle navigation, and chemotaxis. These emerge efficiently both in single-species and multi-species contexts. Flow Lenia supports stable "creatures" where original Lenia fails due to uncontrolled mass growth or dissipation.

Large-scale multi-species simulations demonstrate the following phenomena:

• Continuous turnover of parameter species: Birth, death, takeover, and hybridization are all observed in the parameter field.
• Spatial evolutionary trees: PCA projections of the evolving parameter space show branching patterns reminiscent of phylogenetic trees.
• Resource-based selection and ecosystemic interactions: When coupled with food fields, decay, or dissipation, dynamics include explicit foraging, reproduction, and selection.
• Emergent phase transitions: Tuning model "temperature" or decay rates yields regimes with rich critical behaviors, supporting ongoing novelty and complexification.

Custom metrics—such as evolutionary activity, non-neutral activity, and diversity—quantitatively capture the ongoing novelty and open-endedness of the system (Plantec et al., 10 Jun 2025, Plantec et al., 2022, Michel et al., 21 May 2025).

5. Discovery and Quantification of Complexity

Flow Lenia’s capacity for morphogenetic and behavioral complexity has been mapped using both hand-crafted and information-theoretic metrics.

Compression-based complexity, measured by multi-scale image compressibility (e.g., PNG encoding ratios), is used both to characterize the achievable range of Flow Lenia's patterns and as an explicit evolutionary target (Adachi et al., 2024, Michel et al., 21 May 2025). Experiments using genetic algorithms readily steer the system to produce structures with desired complexity—from highly compressible blobs to intricate, noise-like, multi-scale textures. The accessible complexity range is empirically bounded (≈[0.21, 0.62] per (Adachi et al., 2024)), determined by model hyperparameters.

This approach is complemented by multi-scale entropy measures and the analysis of spatio-temporal compressibility (e.g., video file sizes), supporting the detection and exploration of emergent complexity and the evaluation of “interestingness” in evolved ecosystems (Michel et al., 21 May 2025).

6. Automated and Human-In-the-Loop Ecosystem Exploration

Systematic exploration of Flow Lenia’s configuration space is accelerated by curiosity-driven automation, notably via Intrinsically Motivated Goal Exploration Processes (IMGEP) (Michel et al., 21 May 2025). In this method:

• System-level descriptors (evolutionary activity, compressibility, entropy) form a goal space.
• New parameterizations are sampled or perturbed to explore the space maximally.
• Behavior, coverage, and diversity under IMGEP exploration significantly outperform pure random search both quantitatively and qualitatively, rapidly discovering novel ecosystem dynamics, speciation events, and complex interaction networks.

Interactive tools enable real-time navigation through the high-dimensional space of discovered behaviors, supporting scientific analysis and iterative hypothesis testing.

7. Implications and Ongoing Directions

Flow Lenia establishes a scalable, physically interpretable CA-based platform for studying self-organization, collective behavior, and open-ended evolution under strong conservation and heritability constraints. Its unification of rule and pattern via parameter localization blurs the conventional boundary between genotype and phenotype, making it an apt testbed for the study of artificial life as “life-as-it-could-be” (Plantec et al., 10 Jun 2025).

Anticipated directions include:

• Finer definitions of species and individuality via clustering in phenotype or behavioral space.
• Incorporation of further environmental heterogeneity (obstacles, gradients, resource fields).
• Richer genetic operators and multi-level selection dynamics.
• Extension to other classes of parameterizable, spatially extended dynamical systems (Michel et al., 21 May 2025).

Flow Lenia thus provides both a concrete computational artefact and a conceptual framework advancing the theoretical and experimental study of autopoiesis, emergent intelligence, and evolutionary dynamics in artificial media.