Computational co-design of fibrous architecture

Fibrous architecture constitutes an alternative approach to conventional building systems and established construction methods. It shows the potential to converge architectural concerns such as spatial expression and structural elegance, with urgently required resource effectiveness and material efficiency, in a genuinely computational approach. Fundamental characteristics of fibre composite are shared with fibre structures in the natural world, enabling the transfer of design principles and providing a vast repertoire of inspiration. Robotic fabrication based on coreless filament winding, a technique to deposit resin impregnated fibre filaments with only minimal formwork, as well as integrative computational design methods are imperative to the development of complex fibrous building systems. Two projects, the BUGA Fibre Pavilion as an example for long-span structures, and Maison Fibre as an example of multi-storey architecture, showcase the application of those techniques in an architectural context and highlight areas of further research opportunities. The highly interrelated aesthetic, structural and fabrication characteristics of fibre nets are difficult to understand and go beyond a designer’s comprehension and intuition. An AI powered, self-learning agent system aims to extend and thoroughly explore the design space of fibre structures to unlock the full design potential coreless filament winding offers. In order to ensure feedback between all relevant design and performance criteria and enable interdisciplinary convergence, these novel design methods are embedded in a larger co-design framework. It formalizes the interaction of involved interdisciplinary domains and allows for interactive collaboration based on a central data model, serving as a base for design optimisation and exploration. To further advance research on fibre composites in architecture, bio-based materials are considered, continuing the journey of discovery of fibrous architecture to fundamentally rethinking design and construction towards a novel, computational material culture in architecture.


Fibre composites in technology and nature
Fibrous composites joined the spectrum of architectural materials only relatively recently. While membrane and textile architecture, in the form of tents and marquees, has a millennia-long history and significant influence on architectural theory, as for example in Semper's theory of Bekleidung, fibre reinforced composites have only been explored since the mid-1950s. They are made-up of fibrous elements with primarily tensile capacity, which are embedded in a matrix material that surrounds them and holds them in their relative positions. The fibre and matrix elements remain distinct in the composite material, but their combination typically results in properties and performance characteristics that significantly surpass the ones of the individual constituent parts (Fig. 1).
Enabled by the discovery of possibilities for mass producing glass strands and the development of suitable resins in the years between the first and second world war, fibre reinforced plastics (FRP) were initially mainly

Open Access
Architectural Intelligence *Correspondence: fabian.kannenberg@icd.uni-stuttgart.de conceived as cheaper, lighter and more independently available replacements for materials such as wood and steel for the naval and aircraft industries. Only after the war, especially in the late 1950s and 1960s, architects began to experiment with these new materials, with the Monsanto House and the Futuro House being amongst the most prominent examples of their respective decade (Genzel & Voigt, 2005). However, these so-called houses of the future were short-lived, mainly because of building-physics deficiencies and their inherent conceptualisation of architecture as site-unspecific, serially-produced consumer products.
Individualisation enabled by CNC technologies has given rise to a revival of architectural use of FRP since the 1990s, mainly as the economical material of choice for free-form surfaces and claddings. At the same time, the use of fibrous composites, especially glass fibre reinforced plastics (GFRP) and carbon fibre reinforced plastics (CFRP) has gained significant momentum in various industries and led to large-scale applications as the main structural material, as for example in commercial airliners, mass-produced cars and wind turbine blades. However, most of these applications share a common production principle, regardless whether in the aerospace, automotive or in the building sector: They all depend on one of a variety of moulding devices, may it be positive or negative moulds in fibre lay-up and fibre placement processes, internal male mandrels in filament winding or dies in fibre pultrusion. In architecture, this requirement for elaborate, expensive and often wasteful moulds still limit the use of fibre composite materials to extravagant projects with individualized building elements, or components produced in larger quantities with identical geometry. Moreover, most available fabrication technologies available in the building sector rely on the use of semi-finished products, such as fibre mats or textiles, which results in production waste and offcuts, as well as predetermined fibre layouts and orientations.
Nature shows a profoundly different way how fibre composites can be utilized. In fact, most load-bearing structures in living nature are fibrous composites, but, despite their omnipresence, all fibres are based on a small number of fundamental materials (Jeronimidis, 2004). The basic polymer types of biological fibres are cellulose in plants, collagen in animals, chitin in insects and crustaceans, and silks in spider webs. The vast diversity of biological structures all originate in this limited palette of fundamental materials, but their fibres' arrangement, directionality and density, as well as the chemical makeup of the matrix, is highly differentiated (Neville, 2011). Natures takes maximum advantage of the inherent anisotropy of fibrous systems by locally varying the layout, orientation and concentration of fibres in response to the required characteristics of the larger material structure, and thus fully utilizes the material effectiveness and local variability of fibrous systems (Knippers & Speck, 2012).
As biological fibre composites share their fundamental characteristics with man-made, technological fibre composites, such as GFRP or CFRP, natural fibre structures offer a vast repertoire of design principles for potential transfer to architecture (Hensel et al., 2010), especially in the light of recent advances in digital design, simulation and fabrication (Menges & Knippers, 2015 This entails the study of three highly interrelated research areas: (1) the development of robotic fabrication processes that allow for the production of individual fibrous building elements with varied fibre arrangements, orientations and densities, that do use all material without any production waste or offcuts, but that do not require any moulds or mandrels; (2) the development of related computational methods that enable the design of such highly differentiated fibrous systems on the level of each individual fibre roving, which are bundles (3) the resulting, genuinely digital fibre composite building systems, which are extremely lightweight and materially-efficient.
A key enabling technology for fibrous architecture is the robotic fabrication process of coreless filament winding (CFW) developed by ICD and ITKE to produce light-weight, large scale fibre composite for architectural structures. It is based on conventional filament winding, a process used in the composite industry to produce compression vessels and tanks where resin impregnated fibre strands are deposited on a mandrel in different pattern configurations. CFW however, does not rely on any mandrel thus unlocks a plenitude of component morphologies independent of any mould. It rather constitutes a continuous process of formation where the shape of the building part is the direct result of the reciprocal interaction of free spanning fibres between two boundary frames (Figs. 2 and 3). During the process, the predefined shape of the building component emerges only from the interaction of the filaments.
Both modular and monocoque composite structures were built in the past using CFW (Menges & Knippers, 2015). While monocoque structures may offer a larger design solution space due to more geometric freedom, they are limited by transportation and handling volume which becomes specifically problematic for large scale architectural applications. One of the key potentials of CFW is that it allows for bespoke form and individual fibre layup for each component without any economic disadvantage. In addition, there is no production waste or material off-cuts. This significantly extends the design space of components produced on identical boundary frames and allows for differentiated fibrous morphologies despite being a modular structure.
With only a target geometry but no distinct mould, the prediction of the interaction of free spanning fibres, bundle cross sections and the resulting shape of the composite through bespoke simulation methods becomes imperative to all design steps and requires a highly integrative computational approach. The CFW process enables fabrication of building structures based on the individual placement of several hundred thousand meters of spatially arranged and locally differentiated fibres (Fig. 4). The possibility to individually design and place them in turn poses a significant challenge to current architectural design methods, not only because of Beyond typical linear workflows and established design and production phasing, the strong interdependence between the CFW fabrication process and resulting design and performance characteristics necessitates a feedback-based development of design and engineering methods, fabrication and construction processes, and the resulting material and building systems, an approach referred to as co-design . The development of an increasingly integrative approach to computational design and robotic coreless filament winding has been explored and tested along a series of full-scale projects in recent years. Two examples indicating the potential of the resulting fibrous architecture are the BUGA Fibre Pavilion, a long-span, high-performance structure (Figs. 5 and 6), and the Maison Fibre, a speculative model for fibrous multi-storey habitats (Figs. 7 and 8).
The BUGA Fibre Pavilion  was commissioned by the Bundesgartenschau Heilbronn as a temporary exhibition pavilion, with a floor area of around 400 square meters. Its distinctive structure is entirely made from more than 150.000 m of glass-and carbon fibres and achieves a free span of more than 23 m (Fig. 9). The 60 loadbearing, primary structural elements were produced by robotic CFW. Based on a minimal, adjustable edge frame that holds the anchor points for the winding process, first a lattice of translucent glass fibres was built up, onto which black carbon fibres are placed where they are structurally needed. Much like in natural fibre systems, CFW allows to exploit the fibre's natural anisotropy leading to local differentiation of material and more efficient material configurations, which can be tailored to The extensive testing and approval procedures required for this building project showed that each one of these highly load-adapted components can take up to 250 kilonewtons of compression force (Fig. 10), which equals around 25 tons, while remaining exceptionally lightweight. In fact, with 7.6 kg per square meter, the pavilion fibre structure is approximately five times lighter than a more conventional steel structure.
The pavilion translates this innovation on a technical level into a unique architectural expression. The interplay between the black carbon filament bundles and the translucent glass fibre lattice create a stark contrast in texture that is highlighted by the pavilion's fully transparent, mechanically-prestressed ETFE skin. The distinctive architectural articulation intensifies the spatial experience for the visitors and exposes the underlying design principles of the fibrous tectonics (Zechmeister et al., 2020) in an explicable yet expressive way.
Maison Fibre (Dambrosio et al., 2021), as exhibited at the Venice Architecture Biennale 2021 (Fig. 11), is the first multi-storey architecture that consists entirely of a fibrous structure made by robotic CFW fabrication. Employing an integrative computational design, simulation and fabrication approach, it was possible to design the slab elements requiring only 2 % of the component volume as material volume. The code-compliant, loadbearing fibre structure of the upper floor weighs just 9,9 kg/m 2 , with the wall elements having even lower specific weight. The extraordinary lightweightness is not only exposed structurally but also architecturally, resulting in a distinctive expression stemming from the radically different nature of fibrous systems compared to conventional solid slabs and walls (Fig. 12). This extremely low material consumption coupled with the very compact, robotic production suggest future applications based on onsite fabrication, not only during the initial construction process, but also during expansion or conversions. Moreover, the project's integrative computational design method and robotic CFW process is transferable to a wide variety of alternative fibrous materials, ranging from mineral fibre systems that can withstand extreme temperature stresses, as for example basalt, to fully renewable natural fibre systems, such as hemp or flax fibres.
In order to further advance coreless-wound fibre composite architectural systems, two main challenges need to be tackled: The compound aesthetic, structural and fabrication characteristics of fibre nets and fibre bodies are difficult to fully comprehend and their complex interrelations tend to be too great a remove from the modalities of a designer's sense and intuition. Therefore, new computational methods for the exploration of the full design space of coreless-wound fibre composite architectural systems are required, which is currently investigated by employing artificial intelligence (AI) methods for selflearning fibre agents. In order to ensure feedback with all relevant design and performance criteria and allow not only for exploration but also interdisciplinary convergence, these design methods then need to be embedded in a larger co-design framework.

Artificial intelligence for self-learning fibre agents
In early design phases of large-scale coreless-wound fibre composite architectural systems, considerable effort is required to develop project-specific solutions due to the high complexity of interdependent parameters of design, structure and production. A particular challenge constitutes the fibre layup. Thousands of fibres need to be finely calibrated and their location, orientation and density steered towards the desired aesthetics, required performance and producibility. Since fibre components are produced sequentially, by winding a fibre between anchor points to gradually layer and form-find a fibre network, their material behaviour during production needs to be integrated from the start (Reichert et al., 2014). As coreless winding significantly reduces the amount of formwork, with no mandrel or mould for the fibres to rest on, the prediction of interaction of free spanning fibres and the resulting shape of the composite is paramount. The design process therefore requires computational methods to simulate the result of a winding scheme and evaluate its quality and physical integrity. The development of the winding sequence, the so-called fibre syntax, has to consider fabrication constraints and requires constant feedback from simulation to evaluate the aesthetical, structural and functional requirements of the resulting fibre net. In order to extend the design space beyond structures based on experience and intuition, we aim to develop general design methods that are applicable beyond project-specific solutions, unlocking the full potential of coreless filament winding.
Agent-based modelling (ABM) has proven to be an effective method to handle the complex interrelations and reciprocities of multifaceted performance criteria. At ICD, an agent-based design framework was developed (Groenewolt et al., 2018), which combines bottom-up, rule-based, performance-driven ABM methods with the possibility to directly steer design exploration and allow for interaction between computational processes and human design intent, intuition and competence, enabling advanced forms of human-machine interaction in design. The integration of artificial intelligence methods has become a central part of the development of this agentbased design framework, as they allow to go beyond predetermined, hard-coded and fixed behavioural rules towards self-learning, intelligent agent behaviours.
In the context of fibrous structures, the task of finding an optimal winding path is assigned to a virtual fibre agent. This autonomous manipulator agent (Stieler et al., 2022) is deployed in a simulation environment and generates the continuous fibre paths by moving from winding anchor to winding anchor while depositing material and thus altering the resulting geometry (Fig. 13). Since the amounts of possible solutions are vast, predetermined agent behaviours can be used to find initial valid fibre layups. Visual analytics can aid in the assessment and expose relationships between input parameters and the evaluation criteria (Abdelaal et al., 2022).
However, the definition of agent behaviours relies on expert knowledge to predict which sequence of actions will lead to the desired outcome. As expert knowledge is scarce and, even if the related expertise is present, design solutions can be counterintuitive, this approach severely limits the full exploration of the design space, leaving potentially interesting but unknown designs undiscovered.
This challenge has been addressed by employing Reinforcement Learning (RL), an AI method that deals with learning associations between observations, actions and rewards (or punishments) (Sutton & Barto, 2018). To go beyond heuristic agent behaviours, the virtual fibre agent is trained in a simulated environment (Juliani et al., 2020). By repeatedly building up and evaluating fibre nets, the agent is exploring and observing its environment while being rewarded for actions that lead to desired outcomes. The agent's ultimate objective is to maximise the reward it receives; thus, it tries to explore and exploit effective behaviours that yield the highest reward. These rewards are shaped by high level goals based on essential design characteristics, such as the amount of fibre-fibre interaction, fibre-segment lengths and fibre orientation. The behaviours are learned with Proximal Policy Optimization (Schulman et al., 2017), a RL algorithm which can be used with either discrete or continuous action spaces. Enabled by parallelized computation, many agents can be trained simultaneously, all contributing to the learning process (Fig. 14).
Assuming a sufficiently varied training curriculum (Bengio et al., 2009), the virtual agent could work well on a large variety of fibre components. By varying the winding point configurations during the training, the agent aims to maximise the expected return across a distribution of environments and could therefore potentially generalise the learned behaviours to unseen winding point configurations. In traditional optimisation approaches, a change of the winding point configuration would require to start the optimisation from scratch. In contrast, the proposed method is also open to more interactive settings in which the designer can actively interact with the virtual agent (Arzate Cruz & Igarashi, 2020). From a larger perspective, this also questions how a designer's tacit knowledge, that is based on experience and intuition, can be learned by (and taught to) virtual agents. Such advanced forms of human-machine interaction in design are required to enable the full exploration and exploitation of the design space offered by genuinely digital architectural systems such as coreless-wound fibre composites, in order to tap their potential to consume a lot less material and resources by leveraging a higher level of -computationally enabled -design intelligence. However, they also need to be embedded in a larger computational co-design framework to work in a fully integrative and interdisciplinary manner.

Computational co-design framework for fibre composite in architecture
The development of computational design strategies for coreless-wound fibrous composite systems by ICD and ITKE over the past years put considerable effort on the integration of digital simulation of engineering problems and fabrication processes, advancing computational design from being merely focused on architectural concerns towards truly integrative computational design, negotiating multiple and oftentimes contradictory requirements and related streams of information (Reichert et al., 2014). While this way of immediate collaboration proved to be very productive for well established, small interdisciplinary teams, it is difficult to be scaled up to larger groups of collaborators and building scales significantly exceeding the size of a small-scale building demonstrators. To exhaustively leverage the potential of both computational design and the wealth of knowledge provided by a large team of experts from different fields, interactive and reciprocal approaches to the use of digital technologies are required (Gil Pérez et al., 2022b). Existing solutions and software packages in architecture, engineering and construction (AEC) however, are largely based on datasets of conventional, established building systems and construction processes and lack the flexibility and adaptability to be used for novel and experimental structures (Poinet et al., 2020) Rather than relying on incongruous datasets and prevalent iterative, linear process chains, co-design is based on constant, cyclical feedback. It aims to assess multiple areas of research simultaneously and in constant exchange to enable mutual, fundamental innovation of the construction sector . Novel building systems, such as coreless-wound fibre composite systems, and related manufacturing and engineering processes add additional challenges to the development of suitable collaboration strategies, stemming from geometric uncertainties involved in CFW . A project independent, open and interactive computational framework for co-design, not restricted to any specific project or component typology and easily expandable for the use across different scales was developed. It clusters multidisciplinary specialist collaborators into multiple domains of expertise. The involved areas of expertise include a simulation domain comprising fibre interaction simulation and structural simulation, a fabrication domain including hardware setup configuration and the generation of robot toolpaths, the evaluation of structural, geometric and fabrication data in an evaluation domain as well as comprehensive data management performed by a data integration domain. The domains interact based on a hierarchy of nested solution spaces (Fig. 15). A morphological solution space defines the validity of a building part according to criteria established by the building system, a fibre syntax solution space contains all valid fibre layups based on mechanical feasibility, structural integrity and aesthetic considerations and a fabrication solution space contains the producible overlap between morphological solution space and fibre syntax solution space considering a given fabrication setup (Gil Pérez et al., 2022b).
The developed framework formalizes the domain's interactions based on the requirements of CFW, independent of project scale or typology. The main interface for all collaborators alike is an open source object model, customized and extended to fulfil the requirements of Fig. 15 Nested solution spaces of CFW, governing the different domain's interaction. (Gil Pérez et al., 2022b) data storage and exchange of coreless wound composite building systems. The object model hierarchically stores information on simulation, fabrication and quality evaluation and allows for comparison and correlation of process relevant aspects between different domains (Fig. 16). Being implemented on top of an open source project, the object model can easily be extended and allows for interoperability between different software platforms common in the AEC industry (Poinet et al., 2021). Thus, it remains accessible and customizable also for non-experts which considerably lowers the technological threshold and allows for community-based, shared development in the future.
On a more global level, a computational co-design framework is a prerequisite for design and implementation of large-scale fibrous architecture. Given a fibrous building system's rise in complexity in proportion to its increase of scale, accurate simulation tools gain significance as comprehensive physical testing gets increasingly labour and time intensive. A deeper understanding of the fibre systems interdependencies across different domains thus becomes imperative and serves as a base for agentbased design optimisation based on adaptive, nonheuristic behaviours, enabling more informed design decisions already in early project stages. Further interoperability can be provided by interfacing with a semantic description of manufacturing hardware which would enable automated feedback on fabrication constraints and parameters given a specific building part. Changes of the parts geometry or of fabrication setup components can be considered and their impact on the design solution space evaluated which would contribute to a more efficient design and planning process.
Beyond design optimisation, the computational codesign framework establishes the boundary conditions for design exploration based on the interaction of the designer with intelligent, virtual agents leading the way towards unprecedented design solutions outside of the designer's own imagination or intuition.

Conclusion and outlook
Fibrous architecture constitutes a novel way of addressing the imperative to position architecture and its sociocultural dimension in the light of the ecological challenges the built environment is facing, by converging concerns that relate to spatial expression and structural elegance, as well as resource effectiveness and material efficiency, in a genuinely computational approach.
In order to further advance the related research, it needs to expand from bio-inspired structures to also include bio-based materials. The integrative computational design and fabrication technologies, including the self-learning fibre agents and computational codesign framework, presented in this essay are generally open and expandable towards a broad range of material systems, including bio-based matrix materials as well as bio-based fibres. Particularly promising candidate natural materials are hemp and flax fibres, materials that are fully naturally renewable, biodegradable, and regionally available in Central Europe (Gil Pérez et al., 2022a) (Fig. 17).
In many ways, the journey of discovery and development of fibrous systems in architecture, of which this essay provides a short glimpse, serves as model of how digital technologies provide a vehicle to rethink design and construction, one that is capable to take us places far beyond the well-trodden path of established architectural morphologies and known structural typologies en-route towards a novel, computational material culture in architecture. Fig. 17 The livMatS pavilion, a load-bearing fibre structure entirely comprised of natural flax fibres, offering resource-efficient alternatives to conventional construction methods.© IntCDC University of Stuttgart