Neural Particle Automata bring self-organizing AI to dynamic particle systems

Hyunsoo Kim and colleagues from EPFL and KAIST have introduced Neural Particle Automata (NPA), a new neural architecture that generalizes Neural Cellular Automata (NCA) from fixed grids to dynamic particle systems. Unlike classical NCA where cells are pinned to pixels or voxels, NPA models each cell as a particle with continuous position and internal state, updated by a shared learnable rule. This Lagrangian approach enables clear individuation, heterogeneous dynamics, and computation concentrated only where activity occurs. To handle the challenge of dynamic neighborhoods and avoid quadratic scaling, the authors replace grid-based perception with differentiable Smoothed Particle Hydrodynamics (SPH) operators, backed by memory-efficient, CUDA-accelerated kernels that allow scalable end-to-end training.

The NPA framework demonstrates strong performance across three diverse tasks: morphogenesis (simulating biological pattern formation), point-cloud classification, and particle-based texture synthesis. In all cases, NPA retains key NCA behaviors such as robustness and self-regeneration—the ability to recover from perturbations—while enabling new behaviors specific to particle systems. The work positions NPA as a compact neural model for learning self-organizing particle dynamics, with potential applications in physics simulation, robotics, and computational biology. The paper is available on arXiv with code and data expected to follow.