Neural Particle Automata: Learning Self-Organizing Particle Dynamics

Neural Particle Automata (NPA) enable particles to learn self-organizing behaviors by replacing the fixed grid of traditional Neural Cellular Automata (NCA) with a dynamic, particle-based perception system. By utilizing Smoothed Particle Hydrodynamics (SPH) inspired kernels, NPA allows particles to perceive and interact with their neighbors in continuous space, enabling the emergence of complex patterns and morphology without being tied to a lattice.

Local Perception via SPH Kernels

NPA implements a perception system where each particle possesses a continuous position and an internal state. Instead of reading from fixed lattice neighbors, particles aggregate information from nearby particles within a specific support radius. This aggregation is performed using smooth kernels, which allows the particles to operate on irregular and dynamic configurations while preserving the locality of the system.

Key measurements gathered by each particle include:

• Density: A scalar summary of the neighborhood.
• Smoothed State: The kernel-averaged RGB state of neighboring particles.
• Density Gradient: A vector indicating the direction of where particle density increases (the "crowded" side).
• Moment Matrix: A geometric summary of how neighbor offsets and kernel-gradient directions cover the local space.

Gradient Estimation and Geometric Correction

To understand how internal states change across a space, NPA uses state-gradient estimators. A 0th-order estimator calculates differences in state values, which remains accurate for constant state fields but can be biased when particles are irregularly placed and unevenly sampled.

To solve this local geometric distortion, NPA employs a 1st-order correction using the moment matrix. By multiplying the 0th-order estimate by the inverse of the moment matrix, the system normalizes the local geometry, making the gradient estimate exact for locally linear fields, provided the neighborhood is well-conditioned. This ensures that that the perceived state-gradient is based on the field's actual change rather than the artifact of particle distribution.

Community Insights and Potential Applications

Discussion among developers and users of the project has highlighted several potential applications and theoretical connections to biological systems:

"On the outside it looks very similar to what Michael Levin found on electrical communication between living cells. There too, the organism's cells were able to structure and repair their larger-scale morphology."

Beyond biological mimicry, community members have suggested the following potential extensions to the NPA framework:

• Texture Synthesis: Using particles arranged in a grid or interpolating between particles to recreate textures, particularly in low-contrast areas.
• 3D Point Clouds: Applying the NPA approach to 3D point cloud experiments, similar to Gaussian Splatting.
• Cell Division: Implementing mechanisms for particles to divide, mimicking biological cell growth.
• Symmetry and Stability: Users observed that patterns can be disrupted enough that they cannot re-form, suggesting a further area of research into the stability of the self-organizing patterns.