Self-Assembly-Driven 3D Generative Computing: A Framework for Constructing Three-Dimensional Objects Through Learned Local Interactions

Most three-dimensional (3D) generative systems construct objects by predicting or refining geometric representations, including voxels, point clouds, implicit fields, meshes, and Gaussian primitives. This paper proposes Self-Assembly-Driven 3D Generative Computing (SAD-3DGC), a conceptual framework in which 3D structures emerge through repeated local interactions among computational assembly units. Instead of requiring every unit to receive a prescribed final position, the framework allows units to update their positions, orientations, internal states, and attachment relationships according to a shared or condition-dependent interaction policy. Global geometry is therefore treated as an emergent outcome of local assembly dynamics rather than exclusively as a directly predicted configuration.

The framework consists of an assembly-unit representation, a dynamic neighborhood graph, a learned local interaction function, an attachment and detachment mechanism, an optional environmental field, and a global structural evaluation objective. A target shape may condition the initial state, interaction parameters, boundary conditions, or stopping criterion without specifying every final coordinate. The proposed formulation accommodates point-based, particle-based, voxel-based, and Gaussian-based representations and can be combined with differentiable rendering or geometric reconstruction objectives.

This paper formalizes the proposed framework, presents a mathematical model and a training procedure, and outlines an experimental protocol for evaluating geometric fidelity, assembly stability, computational cost, and robustness to perturbations. The framework is a research proposal rather than an experimentally validated method. Its potential contribution lies in explicitly treating learned assembly dynamics and emergent global geometry as central components of 3D generation. Because generative cellular automata, neural particle systems, and computational models of colloidal self-assembly already investigate closely related mechanisms, empirical comparisons and a precise distinction from these approaches are necessary to establish the framework's eventual novelty.