Transformer Hypernetworks for Complex Structural Representation

This research introduces a novel method for generating large-scale structural systems from text-based user prompts. The proposed reconstruction approach consists of a fine-tuned named entity recognition model and an ensemble of deep neural hypernetworks, which are respectively tasked with extracting the design criteria from the user’s input and predicting each of the base structural elements for the given design task. These geometries are replicated and manipulated within the design space using a series of translation vectors that are generated by an additional transformer encoder, allowing for the construction of complex, interconnected structural systems with unprecedented representation capacity. These models achieved an average criteria extraction test accuracy of 96.89% and a weighted average intersection over union of 0.7868 across the test reconstruction datasets. The complexity of designs produced by the proposed text-to-structure models significantly exceeds the scope of previous research within the field of three-dimensional (3D) shape reconstruction, advancing the hypernetwork method and similar reconstruction techniques to meet modern engineering and architectural applications.