Integrating machine learning and structural ptimization for efficient design of grid-shell structures

Abstract

Structurally efficient forms for grid-shell structures are generally created using computationally expensive iterative processes. Form-finding methods aim to develop optimal forms in which members experience minimal internal forces. Current computational processes suffer from scalability issues due to their high processing cost. This research work describes an Al-accelerated workflow that uses Graph Neural Networks (GNNs) and Potential Energy Method (PEM) to accelerate form-finding of dome-shaped grid-shells. The GNN model is tramed to calculate optimal node positions and cross-sectional areas of members without conducting iterative simulations by encoding grid-shells in graphs. The findings reveal that the GNNs, in comparison to the conventional PEM, enable predictions of node configuration that are nearly 99% accurate while reducing computational costs significantly. This model can suggest initial member sizes using a stress-ratio technique combined with EN 1995-1-1 design guidelines to propose to the designer initial section sizes. Future work will explore the applicability of GNNs to other computational form-finding methods and generalize them to any grid-shell layouts by incorporating physics equations.