Generative adversarial networks (GAN) have emerged as powerful tools for generating high-fidelity data across multiple domains by learning complex spatial and geometric relationships directly from structured datasets. GAN generates output grounded in quantitative data-driven distributions, enabling precise and physically consistent generation. In additive manufacturing (AM), parts are designed in CAD, then processed by a slicer to produce layer toolpaths as part of the conventional G-code generation. In this study, a generative model was investigated to directly produce executable layer-wise tool points and thus, toolpaths solely based on given geometry conditions (e.g., shape identity and scale). A GAN-driven framework, comprising a conditional autoencoder (CAE) and a conditional Wasserstein generative adversarial network with gradient and Gaussian penalty (C-WGAN-GGP), is developed for AM this generation, controlled by explicit conditions such as shape type and size. First, the CAE is trained to reconstruct layer contours based on point sets ( x, y ) to a compressed real latent vector that preserves geometry features and scale. During GAN training, the generator network learned to produce fake latent vectors that closely matched real latent vectors, making them indistinguishable to the critic. Afterward, giving a specified condition to the trained GAN model produces latent representations that can be subsequently decoded into point sets ( x, y ). Finally, the generated point sets are post-processed through sequencing algorithms to convert unordered points into printable toolpaths. This framework bridges design intent and fabrication by generating condition-aware toolpaths on demand, accelerating the design-to-manufacture cycle and scalable AM workflows.
Ali et al. (Sun,) studied this question.