StructDiff:

 Discrete graph diffusion for constraint guided structure generation
Abstract. The automated generation of optimal timber wall frames using specific length constraints remains challenging due to computationally intensive nesting algorithms. This, combined with the scarcity of domain-specific training data in construction applications related to other fields such as medical imaging and image processing, slow the development and implementation of machine learning solutions. This paper presents a novel graph diffusion model pipeline that utilizes a UNet-inspired architecture for the generation of timber wall partitions with constrained element length inputs. The approach addresses the need for workflows focused on specialized, domain-specific AI diffusion models by integrating structured initialization techniques with discrete diffusion processes for 2D wall generations. The system functions by focusing on splitting the prediction task into multiple stages and initializing the diffusion process from a constrained initial state. The focus on customizable embedded vectors as task specific conditioners displayed rapid model personalization across structural domains and strength in the diffusion process. Using only synthetic data for training, we deliberately target real world applications with intended goals using the focused method for task specific model generation.  The experimental results demonstrate that this model framework can achieve high success rates for generating valid structural frames with a training time of a few hours on consumer hardware, generating high-quality graph data in minimal diffusion steps. Through the implementation of these structure specific targets and specific input parameters such as partial lists of available elements into model system, this method is shown to be able to produce valid timber frame fabrication data for implementation into industry production.

Keywords: Diffusion model, Machine Learning, Structural Graph Generation, Generative AI, DDPM, Reciprocal Framework, Material Nesting