STRUCTURAL DESIGN OF THE FLEXIBLY FORMED, MESH-

Vol. 58 (2017) No. 1 March n. 191transported. Without the need for scaffoldingdirectly underneath, there is no need for temporaryfoundations and unobstructed access is madepossible.Three structures are known to have been built witha cable-net formwork, using materials other thanfabric as shuttering. A London City Council schoolassembly hall consisting of five 22m span, 100mmthin hypars, in Southwark, Newington, London,UK, now the Pentagon Hall at the Ark GlobalAcademy, was built around 1960 following a 1:8prototype (Figure 4) [4,5]. In this case, some of thewires were left for post-tensioning. The shutteringwas mesh reinforcement and woodwool insulation,and c. C50/60 shotcrete (‘Gunite’) was applieddirectly from underneath. This underlines thebenefit of unobstructed access when using flexibleformworks.Figure 4: Pentagon Hall, London, UK, built c. 1960, andAuto Perfection car repair shop, Midland, MI, USA, builtc. 1960-1962 (above: Mike Deakin, below: CC BY 4.0Ryan Collier)Around the same time, c. 1960-1962, butindependently, the Bay Service Station in Midland,MI, USA [13], now the Auto Perfection car repairshop (Figure 4), and a clubhouse at the Purdue GolfCourse in West Lafayette, IN, USA [32],demolished in the mid-1990s, were built, both anassembly of four 14m span, 165mm thin hypars.Earlier prototypes, up to 5:8 in scale, are describedin [31]. In this case, the shuttering was XPS foam26insulation, while the cable-net was lost formworkfor traditionally placed c. C45/55 concrete.In these cases, substantial deviations from thedesign shape due to deflections are reported. VanMele and Block [27] presented a method for findingthe distribution of forces to obtain a particularshape, after it has been loaded with fresh concrete.This control allows a range of pre-defined, nonanalytical, anticlastic shapes to be designed andconstructed, with much greater accuracy [29,30].4. FORM FINDING AND OPTIMIZATIONPROCESSThe design process for the roof consists of anintegrated parametric model used for multiobjective evolutionary optimization of the shell, andsubsequent analysis of its nonlinear behaviour aswell as the flexible formwork used for itsconstruction. Figure 5 explains the computationaldesign process of HiLo, consisting of formgeneration, structural analysis, and multi-criteriashape optimization.Figure 5: Workflow of optimization and analysis (sectionsin parentheses), additional criteria in dotted linesThe process consists of boundary, topology andform generation (Sections 4.1, 4.2, 4.3). Then loadgeneration (Section 4.4) to allow for thicknessoptimization (Section 4.5). The shell geometry andmass is now fixed and can be evaluated for furtherfor cable-net forces (Section 4.6) as well as theamount of glazing along its perimeter. Theseparameters were then used to inform the shapeoptimization (Section 4.7). Some details on theimplementation are provided in Section 4.8, before

JOURNAL OF THE INTERNATIONAL ASSOCIATION FOR SHELL AND SPATIAL STRUCTURES: J. IASScontinuing with the furtheroptimization (Section 5).analysisafterThe geometry that is initially generated ismaintained throughout the entire process, actingboth as the layout of the cable net and the mesh ofthe shell itself (apart from triangulation, some nodesinserted to apply wind loads from the glass façade,and subdivision for further analysis in Section 5).4.1.Boundary generationThe shape of the roof is largely determined by thegeometry of its boundary edges, and the topology ofthe generating network. The edge consists of four orfive undulations, one for each support, curvingbetween each support position to the given height hof the roof. Each half undulation is characterised byan amplitude a h, period p, and sharpness s(Figure 6):π z ( t ) a cos 2 t ( x )2 p where t ( x ) (1)additional parameters for the edge shape (Figure 6).The resulting space is required for the exteriorinsulation, drainage, connections to the thin-filmphotovoltaics and hydronic system, providingeffective area for the supports, and ensuring that theglass facade connects to the shell at angles of 45ºto allow for proper detailing.In this case, the sharpness s can be determined froma height h, period p, width w and amplitude a:s c πn(2)n ( c 2π n π )where c arccos ( 2h a 1) and n 1w.2 pIn the final optimization, the five support positionswere fixed, leaving three parameters foroptimization: width w, amplitude a, depth d, i.e.fifteen variables for optimization.s x x.s x 1In a first optimization, four or five supportpositions, determining p, the sharpnesses s, and theroof height h were parameters for the optimization,i.e. seven or eleven variables for optimization.Figure 7: Topology generation4.2.Topology generationThe roof is then divided into five convex patches,determined by five points Bi on the shell’s boundaryand three interior points Si (Figure 7), which aresubdivided as follows (Figure 8).Figure 6: Boundary generationThe boundary curves can extend below thefoundation and can optionally be cut off. By doingthis, the roof touches down on the floor with aplanar, curved footprint. These are defined asparabolas with a certain width w and depth d; twoEach patch is then subdivided along approximatelyradial and concentric directions with respect to thesupport positions.The interior edges of the patch are divided into anequal number of segments that are as close aspossible to some desired, global edge length. Thissame number then subdivides the exterior edges ofthe patch. The resulting vertices are connected tothe corresponding vertices along the interior edges.27

Vol. 58 (2017) No. 1 March n. 191lengths. The parabolic segments get at least threevertices, to avoid degrading them into straight lines.4.3.Figure 8: Force densities interpolated from eleven valuesStarting at the outermost exterior vertices,concentric edges are created that follow the interiorboundary of the patch, crossing all radial edges inbetween. For undulations that are cut off, theexterior vertices are divided evenly over the threeexterior curve segments, based on their relativeForm generationFrom these boundary conditions, a suitable,anticlastic shape is generated using the linear forcedensity method [23]. To minimise the number ofadditional variables for optimization, the forcedensities throughout the network are determined byinterpolating nine or eleven values for four or fivesupports respectively (Figure 5). The ratio ofallowable force densities is limited to 1:20, to createreasonable shapes without too abrupt changes incurvature and resulting forces. In the case of cut-offsupports, the network potentially curves in on itself(Figure 5). This is remedied by calculating forcedensities of the network’s triangulated projectionusing the linear natural force density method [21].This tends towards a minimal surface of ourprojection, avoiding overlaps, and thus any inwardcurving. These force densities are then used in asecond form-finding procedure, which is alsopartially constrained to the original form-foundmesh.Figure 9: Load generation for thermal loads, snows loads and wind zones for main wind direction (SW) bothfor pressure ( ) and suction (-)28

JOURNAL OF THE INTERNATIONAL ASSOCIATION FOR SHELL AND SPATIAL STRUCTURES: J. IASSTable 1: Reduction factors ψ, unfavourable/ favourable load factors γ (SIA 260) and critical buckling load factor λ(IASS 1979)Self-weightDeadThermalLiveWindSnowSLS occasional 1.0 / ψ0Load1.01.0 / 0.01.0 / 0.61.0 / 0.01.0 / 0.61.0 / 0.86SLS frequent ψ1 and ψ21.01.0 / 0.00.5 / 0.000.5 / 00.43 / 01.00.7 (1.0)1.0 (0.0)0001.35 / 0.81.35 / 0.81.5 / 01.5 / 01.5 / 01.5 / 0SLS quasi-permanent ψ2ULS load factor γCLS load factor λ1.75Table 2: Load combinations with and without thermal loads (LC) used. Leading action in boldSLS quasi-permanentSLS occasionalULS4.4.LoadSelf-weightDeadThermalLC 01.00.70.2 / 01.0LC 11.0LC 21.01.0 / 01.0 / 0LC 31.01.0 / 0Snow1.01.01.01.0 / 00.861.351.350.6 / 01.5LC 60.80LC 70.80LC 81.351.350.6 / 00.86LC 91.351.351.5 / 00.860.6 / 01.01.50.6 / 0Load combinations were defined using reductionfactors ψ and load factors γ in Figure 7 followingSIA 260. The quasi-permanent load combination isused for the determination of creep and shrinkage,Live1.0LC 5For each shape, loads are automatically generatedfrom SIA 261 to be applied to the structure. Theseloads include: the self-weight of the concrete (24kN/m3); dead loads from the integrated shell (0.5kN or 0.3 kN/m2); live loads for maintenance on theroof (1 kN or 0.4 kN/m2); thermal loads due to theembedded hydronic system for a minimumtemperature of 0 ºC for optimization (Figure 6) and-20 ºC for final analysis; snow loads (μ 0.9kN/m2, Figure 6); and wind loads (Cp 1.07kN/m2). For the wind loads, half of the wind loadon the glass facade is also taken into account. Thesnow shape factor μ varies between 0 and 0.8depending on the roof angle and the wind shapefactor varies between -0.3 and 0.75 depending onthe wind direction and roof angle (we interpolatebetween facade and angled roof, i.e. zone A andm, in Figure 9).Windpressure1.0LC 4Load generationWindsuction1.51.5with dead loads and thermal loads altered (0.7 and1.0 instead of 1.0 and 0.0) to reflect the actual longterm load on the shell. The occasional loadcombinations are used for checks in theserviceability limit state (SLS) against allowabledeflections and crack width. They are also thestarting point for limit load calculations. Theultimate limit state (ULS) load combinations areused to check against allowable stresses. Limit loadcalculations were carried out to establish whetherthe load factor λ, or safety factor, according toIASS 1979 was met. This limit load state is herereferred to as the ‘critical limit state’ (CLS).4.5.Thickness optimizationBy redistributing the material in the shell, it ispossible to reduce the total volume of requiredconcrete, even though the maximum stresses staywithin the same limits. The program Karamba triesto approach a given maximum deflection of L/500 18 mm, while reducing thicknesses throughoutthe structure and keeping within a 20 MPa stresslimit. The linear elastic stiffness was reduced toonly E 5000 MPa to approximately account for29

Vol. 58 (2017) No. 1 March n. 191cracking and creep in the design. The optimizationis done for all SLS load combinations, as those inthe ULS were found to not govern the results. Thepresented result has a minimum and averagethickness of 3.0 and 7.7 cm, and a total weight of 29metric tons.4.6.Best-fit optimizationThe goal is now to find the forces in the cable-netsuch that, under given loads of the wet concrete, theresulting concrete shell takes the form of the targetshape [27]. The topology and shape of the cable net(Sections 4.2 and 4.3) is the basis for triangulatedmesh of the shell (Section 4.4 and 4.5). To enforcereasonable bounds on these forces under load (4-50kN along the perimeter), the resulting constrainedlinear least squares problem can be written as aquadratic program. Assuming the bounds have notallowed us to find an exact match with the targetshape, we compute the sum of squared deviations,which are used as target for optimization. Theconstrained linear least squares solver offers aninitial estimate of the force distribution, showinghow different solutions compare, but withinreasonable computational time. A more robustnonlinear algorithm [28] is applied to the finalgeometry to obtain the closest-fit in the detailedengineering and tendering phase, as the topology ofthe formwork may change depending on input fromthe future contractor.4.7.Shape optimizationThe roof was optimized in two rounds: initially, asingle-criterion optimization; and then a final multicriteria optimization. The optimization was carriedout for a monolithic concrete shell, and thesandwich section was taken into account in thesubsequent structural analysis (Section 5).The first optimization minimized mass, proportionalto the elastic bending energy E, subject topreliminary stress and deflection constraints (20N/mm2 and 30 mm). The energy is a function of theshape f f(x,s,h,q) with 16-22 variables (seven oreleven boundary parameters plus nine or elevenforce density parameters, for shells with four or fivesupports respectively).This stage studied different boundary conditions(positions and number of supports as well as roof30height), and their relative influence on the potentialto minimize the mass. The problem is to:minimize E ( f ( x, s, h, q ) )subject tos 20 N/mm 2 ,d 30 mm,0.11 x4 0.45,0.60 x3 0.90,1.10 x2 1.90,2.10 x1 2.43,3.45 x5 3.90,0 s1.5 10,0 h 5, and1 q1.11 10.The bounds on variables x were determined to avoidany supports close to the corners, and keep anysupports within the architecturally and functionallypreferred support zones. The bounds on variable swere subjectively set to avoid extremely steep orshallow edge curves. The bounds on variable hwere determined by a minimum ceiling clearanceand a maximum allowable roof height.The second and final multi-criteria optimization,subject to a preliminary stress and deflectionconstraints (20 N/mm2 and 1/500th of the span L),minimized four criteria:internal elastic energy(proportional to mass) as before; the buckling loadfactor (lowest, positive value); deviations of thecable net to the target shape; and, surface area ofglazing. A fifth measure of the amount of headclearance below the roof was also calculated tocompare results, measured as the sum of squaredlengths of all nodes higher than 2.15m. Thesecriteria are all a function of the shape f f(w,d,a,q)with 26 variables (fifteen boundary and elevenforce density parameters for a shell with fivesupports).This stage determined the final design as it wassubmitted to the authorities for building permission(see also Sections 4.1-2). The problem is to:

JOURNAL OF THE INTERNATIONAL ASSOCIATION FOR SHELL AND SPATIAL STRUCTURES: J. IASSminimize E , λ , z T z, A,as functions of f ( w, d , a, q ) ,subject toσ 20 N/mm 2 ,d L / 500,1.2 w1 2.0,0.9 w2.5 1.2,0.42 d1 0.82,0.45 d 2.5 0.75,7.5 a1 9.0,4.4 a2.5 9.0,1 q1.5,11 20, and1 q6.10 10.The bounds on variables w, d and a, were set tomaintain various requirements related to space forinsulation and drainage on the exterior, and toangles between the shell and the glass façade on theinterior.4.8.ImplementationThe entire design process was implemented inGrasshopper for Rhinoceros [14,22]. Several plugins for Grasshopper were included: Karamba oo for the second form-finding procedureand Octopus for multi-objective optimization.Thermal actions were based on calculations carriedout in Energy2D and ANSYS by A/S. A customVB component generated the boundaries andtopology, and custom IronPython components werewritten to communicate with external CPythonscripts; the first form-finding procedure andcalculation of prestresses in the cable-netformwork, the latter using CVXOPT’s QP solver[1] to solve the bounded least-squares problem.The shell was subsequently evaluated for variousadditional nonlinearities in Sofistik (see Section 5),as the present version of Karamba does not includelayered or volume elements to model the sandwich,non-linear material models, or third order geometricnonlinearity to evaluate post-buckling behavior.However, Sofistik is also limited as it is not capableto combine volume elements with both non-linearmaterial and geometric modelling, to load stepthermal actions, and to model the reinforcement inmore than two layers per side. The input for Sofistikis generated from Grasshopper using a customIronPython component.Karamba only offers 3-node triangular TRICelements for shell analysis. The mesh was relativelycoarse (Figure 10) to minimize computational timeduring optimization. Sofistik only offers 4-node,non-conforming, Mindlin-Reissner quadrilateralelements. The mesh was subdivided once toimprove the accuracy, particularly the resolution ofthe buckling modes (Figure 11).Figure 10: Four criteria: elastic energy (proportional to mass, shown as thickness e), buckling load factor λ for LC 0(showing first positive buckling mode with deflection w), cable-net deviations (showing constrained forces F under load),and surface area A of clear glazing31

Vol. 58 (2017) No. 1 March n. 1915. STRUCTURAL ANALYSISThe subsequent structural calculations, carried outin Sofistik, follow Swiss code SIA 262 - intendedfor conventional reinforced concrete - wherepossible, but applies ACI 549R-97 and ACI549.1R-93 for aspects related to ferrocement, andMedwadowski et al. [16], here referred to as ‘IASS1979’, for aspects related to thin-shell structuraldesign. Creep and shrinkage formulas from SIA 262are based on those in EN 1992-1-1:2004. AdoptingIASS 1979 means that we are required to perform astability analysis, by calculating the initial bucklingload, or critical load, then modifying this load - orrecalculating using a sufficiently refined model - bytaking into account: large displacements (geometricnonlinearity), material properties of concrete andreinforcement (material nonlinearity includingcreep and shrinkage) and deviations from theidealised shape (imperfections). Because theresearch unit will be replaced after 5-10 years, thereference period for design is 20 years. Loadcombinations are according to Section 4.4.5.1.Boundary conditionsAs mentioned, the shell is supported on fivelocations. Those at the rear are close to thebackbone, and assumed fixed. Those in front aresupported on a cantilevering, prestressed concretefloor slab, which are modelled as springs(stiffnesses provided by structural engineers of theNEST building, Dr. Schwartz Consulting). Onesupport is modelled as a horizontal spring as well toaccount for the local flexibility of the supportingsteel frame.5.2.The creep coefficients are φ 1.06 (inner face),2.25 (PU foam insulation) and 0.81 (outer face).The drying shrinkage strains are ε -0.19‰ (innerface), -0.10‰ (outer face). Following SIA 262,autogeneous shrinkage is not included yet, pendingdevelopment and testing of the actual concrete mix.The current values assume that the shell remains inthe formwork while curing for 28 days, and that theaverage layer thickness is 50mm. The inner face isexposed on one side and has a relative humidity of40%, while the outer face is completely enclosedand has a relative humidity of 60%. For the creep ofthe PU very little is known, and for now is takenfrom Garrido et al. [7], who investigated rigid PUfoam for sandwich panels, though of much lowerdensity.The creep coefficients and shrinkage strains wereapplied to the quasi-permanent load combination inforty incremental steps, simulating 20 years ofcreep and shrinkage. This state was then used forfurther application of the occasional SLS and theULS load combinations.Material propertiesThe reinforced concrete was modelled as a C90/105with B500A according to SIA 262, but additionalcalculations were carried out for a range betweenC35 and C90 concrete, and for AR-glass and carbonfibre TRC, to inform the detailed engineeringphase. The higher C90 concrete strength wasmainly chosen based on the resulting creep andshrinkage behavi

reinforced concrete (TRC) with glass or carbon fibre offer similar benefits, but is even more flexible. Figure 3: Examples of ferrocement and carbon- fibre TRC sections, 50mm thick, showing dense mesh reinforcement [3,24] The decision for the final material of the (steel, carbon or glassAR ) will be made in the next phase. Due to its high in-plane