OPTIMIZING HARDWARE DESIGNUSING GENERATIVE AI

Sammanfattning

This thesis presents a Transformer-based Generative Artificial Intelligence (GenAI) frameworkfor automating network device enclosure design and addressing both computational and time inefficienciesof standard modelling workflows. The method uses boundary representation (B-rep),which maintain topological and geometrical details in computer-aided design (CAD) models, allowfor hierarchical generation of vertices, edges and faces. A Transformer encoder-decoder architectureprocesses vertices sequentially, while EdgeModel and FaceModel form edges and faces respectivelyusing pointer networks. In EdgeModel, edges are formed by the dot-product similarity of vertex embeddingfeatures. In FaceModel, a closed surface is created from edges using the features generatedby the multilayer perceptrons (MLPs). Due to hardware constraints, EdgeModel and FaceModelimplementations were lightweight, relying on Transformer vertex generation. The results showedthat the Transformer was capable of learning and generating vartices coordinate, while faced challengesin edge / face generation, indicating the need of more data and computational resources.This work has contributed to the area of AI driven CAD generation, with focus on the balancebetween industrial design constraints and generative flexibility.