In Tension: Computational exploration of the design space of tensile network structures

Abstract

Cable and rope net structures are lightweight tensile systems and generally cannot resist compression or bending. Tensile network structures are often used to span long distances without intermediate supports and have found applications in art, architecture, and structural engineering due to their physical and visual lightness. However, the design of tensile net structures is generally challenging since their form cannot be arbitrarily defined. Instead a process of form-finding must be used to establish a geometry where all edges of the network carry only tensile forces. Physical models and computational methods can be used for the form-finding of tensile network structures; however the primary challenge in the design process is the adjustment of the network parameters to achieve a specific design. Recent work has shown that automatic differentiation software packages can be used to efficiently design funicular structures (that is, those that work in pure tension or pure compression) with additional designer driven objectives, but these techniques remain largely inaccessible to general designers, architects, and engineers due to the involved process of problem setup and limited interactivity of existing tools. To address this limitation, I introduce a new tool set consisting of two main components, Ariadne and Theseus. These components take advantage of automatic differentiation of objective functions for efficient tensile network simulation and provide a user interface for architects, engineers, and other designers as a plugin for a commonly used 3d modeling software. In this thesis, I outline the structure and features of this tool set, show results of networks optimized with different composable objectives, and show some fabricated examples. Next, I explore the the generation of more complex 3d network topologies through a procedural shape grammar. Finally, I explore the use of differentiable simulation in conjunction with machine learning techniques to optimize the geometry of tensile networks using semantic input and to develop an implicit representation of the space of equal edge length tensed network poses. Together, this new tool set and additional methods enable a more expansive exploration of the design space of tensile networks where design intent and practical constraints are respected.