Freezing-Front-Driven 3D Generation: A Process-Based Framework for Constructing Three-Dimensional Ice Structures Through Phase-Transition Dynamics

Three-dimensional (3D) generative methods increasingly produce geometric structures through learned distributions, implicit representations, point-based primitives, and neural rendering. However, many natural ice structures emerge through a fundamentally different process: the progressive transformation of liquid water into solid ice under spatially and temporally varying thermal conditions. This paper proposes Freezing-Front-Driven 3D Generation (FF3DG), a conceptual framework for constructing 3D geometry through the evolution of freezing fronts, stochastic nucleation, anisotropic crystal growth, thermal transport, and phase-dependent morphological competition.

The central idea is to treat freezing as a generative process rather than merely a physical effect applied to a previously generated object. A coupled state representation tracks temperature, phase fraction, nucleation sites, crystal orientation, local growth conditions, and the evolving solid–liquid interface. The resulting structure is reconstructed into an explicit or implicit 3D representation, including meshes, voxel volumes, signed distance fields, or Gaussian primitives. A proposed learning-based extension enables conditional generation from target geometries, environmental parameters, and desired morphological properties while retaining physically meaningful constraints.

The framework is intended to support several classes of structures, including icicles, frozen layers, dendritic ice networks, and ice-filled porous ground. Its potential contributions are a unified generation architecture, process-conditioned geometry synthesis, multiscale representation, and controllable freezing histories. Because related physical simulations and ice-growth models already exist, the proposed contribution is architectural and methodological rather than a claim that freezing-based 3D simulation is new. Experimental validation, computational benchmarking, and comparative evaluation remain future work.