Free-form Grid Structure Form Finding based on Machine Learning and Multi-objective Optimisation

Abstract

Free-form structural designs are reshaping modern architecture with their expressive aesthetics and novel spatial configurations. However, their digital transformation still encounters obstacles, particularly in aligning fluid geometries with rigorous structural requirements, and in assembling high-quality datasets suitable for advanced machine learning (ML) methods. Traditional form-finding approaches often neglect critical factors—such as material behavior, load paths, and construction logistics—resulting in discrepancies between conceptual 3D models and built outcomes. To overcome these challenges, this paper presents a novel pipeline that integrates ML with multi-objective evolutionary optimisation, using glued laminated timber (GLT) as a case study. Central to our method is a Transformer-based neural network that harnesses NURBS representations of 3D geometry, creating a structured dataset for curvature prediction. These ML-generated forms are then refined through an evolutionary optimisation process targeting minimal structural mass, stress, and strain energy. Experimental results show notable improvements in design performance, with reductions in mass (3.6%), stress (up to 15%), and strain energy (68% under mesh load) . This synergy of ML-driven geometry and robust optimisation significantly advances digital construction practices by fostering a data-driven, automated workflow for complex free-form design. Practical implications extend across the building and infrastructure sectors, enabling better alignment between conceptual design and final construction, lowering material consumption, and mitigating deviations during fabrication. By coupling aesthetic exploration with structural rigour, the framework underlines a crucial step toward more sustainable, efficient, and constructible free-form structures.