The early-stage design of reinforced concrete shear wall buildings entails the evaluation of numerous candidate layouts, a task rendered slow and computationally costly by reliance on conventional finite element analysis. Existing deep-learning method...
The early-stage design of reinforced concrete shear wall buildings entails the evaluation of numerous candidate layouts, a task rendered slow and computationally costly by reliance on conventional finite element analysis. Existing deep-learning methods address this problem only in part: they treat layout generation and structural assessment as disjoint tasks and presuppose architectural drawings that are unavailable at the earliest design stage. This thesis introduces DiffShearwall, an integrated framework that unifies these tasks by coupling discrete-diffusion layout generation with graph-neural-network drift prediction in a single inference-time workflow that requires no finite element analysis.
Given only the plan dimensions, the number of floors, and a target shear wall ratio, and without recourse to architectural drawings, the diffusion model synthesizes candidate layouts, and the graph neural network predicts the wind-induced inter-story drift of each candidate in both principal directions; the candidates are then ranked by predicted drift, and the selected layout is exported as a labeled drawing and an ETABS-importable spreadsheet so that the chosen design can be opened directly in commercial software for detailed verification. The full pipeline is deployed as a Model Context Protocol server, rendering it operable through conversational interaction with any compatible AI assistant.
The framework is developed and validated on a dataset of 1,907 building configurations spanning 17 rectangular plan types under ASCE 7-22 wind loading. The drift prediction surrogate attains a coefficient of determination of R² = 0.997 on the validation set, and the generative component produces layouts that conform to the target shear wall ratio. Taken together, the results establish a viable path from a natural-language design request to an analysis-ready structural model, extending deep-learning-assisted structural design from isolated generation or isolated prediction toward an integrated, drawing-free, and finite-element-free workflow for early-stage shear wall design.