Wear-Driven Generative Modeling: A Conditional Framework for Three-Dimensional Shape Synthesis Through Progressive Material Removal

Three-dimensional generative modeling commonly constructs object geometry by sampling learned distributions, denoising latent representations, predicting point sets, or optimizing continuous scene representations. However, many natural objects acquire their characteristic shapes through progressive material removal rather than direct construction. River pebbles, rounded mineral grains, and polished surfaces exemplify processes in which repeated contact, friction, and impact progressively modify local geometry. This paper proposes Wear-Driven Generative Modeling (WDGM), a framework for synthesizing three-dimensional shapes through a controlled sequence of surface-removal operations. Given an initial geometry, WDGM maintains a geometric state, a local material-state representation, and a history of environmental interactions. A conditional wear operator estimates spatially varying material removal from contact exposure, local geometry, material properties, and environmental conditions. The resulting geometry is updated while preserving non-negativity of removal, approximate mass consistency, and surface validity. A conditional controller selects the wear environment and process intensity to guide the evolving object toward target shape characteristics, such as roundness, volume, surface roughness, and aspect ratio. A differentiable surrogate or learned transition model may be used to support target-conditioned trajectory optimization. Unlike methods that rely exclusively on direct shape prediction or generic geometric smoothing, the proposed framework explicitly represents progressive material loss as a generative mechanism. This work establishes a formal modeling framework, defines candidate objectives and evaluation protocols, and identifies the experiments required to determine its effectiveness. It does not claim empirical superiority or validated novelty. The proposed approach is intended for procedural shape synthesis, natural-particle modeling, digital asset creation, and physically informed three-dimensional generation.