Generative design for additive manufacturing through data-free optimization-derived neural weight subspaces

Abstract Generative design (GD) enables exploration of diverse shape variants that exhibit trade-offs among competing design objectives related to part manufacturability and performance, offering a transformational approach towards automated part design for advanced additive manufacturing (AM). However, existing GD methods often derive design diversity either from large datasets of existing designs or through topological modifications, limiting their applicability to data-scarce engineering problems, particularly for complex parts where manufacturability and preservation of topology and geometric features are critical. In this work, we present a training-free generative design framework that leverages the optimisation trajectory of a single neural field-based shape optimisation to represent and explore neural field variants within a low-dimensional neural weight subspace. Our design space exploration is realised through a Langevin diffusion-inspired stochastic process with energy-conditioned guidance, enabling the generation of constraint-satisfying shape variants that provide distinct objective trade-offs. The framework generalises across geometrically and topologically diverse complex parts, including freeform geometries with intricate surface features and non-trivial topology. We validate the method through simulations and physical additive manufacturing experiments, demonstrating support-free 3D printing of the generated shape variants and establishing a data-free route towards manufacturability-aware generative design for complex parts.