What this kind of structure is called
The same envelope goes by six different names depending on who is describing it, and each name is accurate about something different. It is worth setting them out, because a client, an architect, a structural engineer and a steel fabricator will all use a different one for the same object.
- Free-form structure — our catalogue name. It says the thing that matters commercially: the surface is not describable by a single radius or a flat plane, so nothing on it repeats.
- Gridshell — the structural engineer's word. It says the surface itself carries the load, through a grid of members acting together as a shell, rather than through a series of parallel frames. It is the most precise of the six.
- Doubly curved surface, or doubly curved roof — the geometric description. The surface curves in two directions at once, which is what makes it stiff and what makes every panel on it unique.
- Curved steel structure, or curved roof steel structure — what it is called on site and in a steel fabricator's workshop, where the concern is bending members and setting out nodes rather than naming a geometry.
- Curved glass façade, or curved curtain wall, where the surface turns vertical and is glazed. On DOMAINE the envelope does both: it is a roof over part of the building and a glazed wall over the rest, and the two are one continuous geometry.
- Parametric façade — what it is called at design stage, because a surface like this is generated by a definition rather than drawn line by line, usually in Rhino and Grasshopper. That matters more than it sounds: the definition, not the drawing, is the thing that has to survive from concept to fabrication.
- Roof and wall, one surface Over the building it is a doubly curved roof; where it turns down to the ground it is a curved glass façade. Nothing marks where one becomes the other.
- A grid working as a shell The members and nodes carry load together, as one surface — a gridshell — rather than as a row of frames.
- Generated, not drawn The surface comes from a definition — control points and rules — and every node coordinate comes out of it. That is what makes it a parametric façade.
Section, proportions schematic.
DOMAINE is all of these at once. The engineering behind them is set out on Free-form Structures; what follows is what happened on this particular building.
What we inherited
DOMAINE was designed and begun by a contractor from another country, using their own proprietary façade system. Part way through construction that company left the project and the contract was terminated. The reasons were never made clear to us and are not ours to speculate about.
What was left behind was a building with a partly erected free-form steel structure, a façade system that could not be bought from anyone, and no continuity between what had been designed and what had been built.
We were approached mid-project. The first thing our engineers did was go to site and look at it, because nothing useful can be decided about a structure of this kind from drawings alone.
A free-form structure is not a product that can be picked up where somebody else put it down. It is a single continuous geometry, and a geometry is either consistent or it is not. Taking one over means inheriting three things at once, and none of them can be bought:
- The geometry — which may or may not be what is actually standing on site
- The system — profiles, connection nodes and gaskets that exist only as parts of a scheme nobody sells any more
- The errors — whatever has already been built out of position, now carrying load
The whole project followed from establishing which of those three we were dealing with, and how badly.
The survey: finding out what was actually built
Before anything was designed, quoted or ordered, the entire erected structure was surveyed and an as-built model was produced from that survey.
This is not a formality. On a conventional building you can take the drawings as approximately true and let the fixings absorb the difference. On a curved steel structure you cannot, because there is no repeating dimension anywhere in it to check against — no standard bay, no typical column, no level that runs through. Every node on a gridshell sits at its own coordinate. If you do not measure it, you do not know it.
The as-built model was then overlaid on the design model and the two were compared node by node, which produces a deviation map — a picture of the built structure coloured by how far each part of it sits from where it was supposed to be.
The structure had moved, substantially, and not in one direction.
In some areas it stood +15 cm high against the model. In others it sat −18 cm low. Across one continuous surface, that is a 33 cm spread between the highest and the lowest deviation.
- Highest deviation
- +15 cm
- Lowest deviation
- −18 cm
- Spread across one surface
- 33 cm
- Envelope area
- 2,500 m²
- The design model The surface as designed. On a gridshell every node has its own coordinate, and there is no repeating dimension to check against.
- The steel as built Surveyed node by node and overlaid on the model. In places it stood up to 15 cm high; in others, 18 cm low.
- A 33 cm spread Between the highest and the lowest deviation, across one continuous surface.
Schematic. The vertical deviation is exaggerated, and the drawing is not the survey data.
Why 33 centimetres is a different problem on a doubly curved surface
On an ordinary building, a slab edge 15 cm out of position is absorbed by a bracket. Brackets exist for exactly that reason: the structure is built to a coarse tolerance and the façade is built to a fine one, and the bracket is the component that converts one into the other.
A doubly curved gridshell does not work that way, for four reasons that compound — and they are the same four reasons that make a curved steel structure expensive to get wrong at any stage, not only this one.
Every panel is unique, and every panel is cut to four specific points. A panel on a doubly curved surface is not a rectangle from a stack. It is a particular quadrilateral, cut to meet four particular nodes at four particular angles. Move one node and all four panels meeting at it change shape at once.
The node angle changes, not just its position. The connection component at each intersection is made for the specific angles at which its members arrive. Displace the node and those angles change. A component made for the design geometry no longer closes on the built geometry, and it cannot be persuaded to — the adjustment built into a node is millimetres, because that is all a node on a sealed surface can afford.
The gasket stops sealing before the frame stops fitting. A gasket seals by being compressed by a designed amount inside its groove. Change the geometry of the joint and the compression changes along its length — more on one side, less on the other. The side with less is where the water comes in. This failure appears long before anything looks visibly wrong.
The error does not stay where it is. A node out of position pulls on its neighbours, which are connected to their neighbours. Deviation on a continuous surface propagates outward, and the further from the correct geometry the surface runs, the harder each subsequent panel is to close.
-
Four panels change at once Every panel is cut to four particular nodes. Move one node and all four panels meeting at it change shape.
-
The node angle changes A node component is made for the angles its members arrive at. Displace the node and it no longer closes — the adjustment built into it is millimetres.
-
The gasket stops sealing Change the joint's geometry and the compression changes along it — more on one side, less on the other. The side with less is where the water gets in, long before anything looks wrong.
-
The error travels A node out of position pulls on its neighbours, and they on theirs. The further the surface runs from the model, the harder each next panel is to close.
Displacements exaggerated.
Which is why 33 cm on a rectilinear building is an inconvenience, and 33 cm on a free-form surface is the reason a project stops.
The decision: correct the steel, not the geometry
Faced with a built structure that does not match its model, there are three real options. We went through all three with the client before choosing.
-
1 · Rebuild the steel Demolish the out-of-position areas and re-erect them. Certain to work — and the slowest, most expensive answer on a project already behind.
-
2 · Re-cut the surface to the steel Possible, because every panel is unique anyway. But it freezes the error into the building, where it reads as a dent in the reflections — and beyond a certain deviation a node cannot close at all.
-
3 · Correct the steel, then clad The route taken: move the standing structure back to the model, then build the designed surface onto it. The harder option, and the only one that ends with the building the architect designed.
Section, deviation exaggerated. The third is the route taken.
1 · Rebuild the steel. Demolish the out-of-position areas and re-erect them correctly. Certain to work, and the most expensive and slowest answer available on a project already behind.
2 · Re-model the surface onto the as-built and re-cut every panel to it. Technically possible, and more possible on a free-form structure than on any other kind, precisely because every panel is already unique — a panel cut to a wrong geometry is no harder to make than one cut to the right geometry.
We rejected it, for two reasons.
The first is that it freezes the error into the finished building. On a doubly curved surface the eye does not read dimensions; it reads curvature. This is the difference between a curved glass façade and a flat one: a flat elevation is judged by its joints, and a curved one is judged by how its highlights run across it. A 33 cm irregularity absorbed into the geometry reads as a dent in the reflection from across the street. It cannot be hidden by glass selection, by finish, or by anything else. It would have been permanent and it would have been visible.
The second is that the node components have angular limits. Beyond a certain deviation the connection simply cannot be made to close, whatever the panel does. Some areas were beyond it.
3 · Correct the structure back to the model, then build the designed surface onto it.
This is what we did, and it is the harder of the two realistic options — it means moving parts of a structure that is already standing and already carrying load. It is also the only one that ends with the building the architect designed.
Rebuilding a system nobody sells
In parallel with the structural work, the façade system itself had to be recreated, because it could not be bought.
The original contractor's system was proprietary — a curved curtain wall and roof glazing system built around profiles that exist nowhere else. Substituting a catalogue system was not available to us: a substantial part of the building was already clad in the original profiles, and the new work had to interface with the old work, joint for joint. Whatever we made had to match what was already on the building.
So the system was reverse-engineered — in three families, each with its own problem.
The aluminium profiles
Each profile section was measured from the existing material, drawn, and then made — which on an extruded profile means cutting a die.
This is the part of the story most people underestimate. An extrusion die is not a drawing; it is a tool, and the first billet through a new die never comes out exactly right. Aluminium does not flow evenly through a complex section: it runs faster where the section is thick and lags where it is thin, and the profile comes off the press with a twist or a dimension out. The die is then corrected against the trial, and run again. That cycle — trial, measure, correct, trial — is what separates a profile that fits the building from one that is nearly the right shape.
-
The first trial comes out uneven Aluminium runs faster through the thick parts of a section and lags in the thin ones, so the first profile off a new die comes out twisted or with a dimension out.
-
Trial, measure, correct The die is corrected against each trial and run again, until the profile fits the building rather than nearly matching the drawing.
CORTEX produces its own aluminium profiles. On this project that was not a marketing line. It was the reason the project could be completed at all.
The connection nodes
The components at each intersection of the grid, made for the specific angles at which members meet there. These were reverse-engineered from the existing work, tooled, and produced as prototypes.
The gaskets
The family everybody forgets, and the one that decides whether the building leaks.
A gasket is an extrusion too, and it needs its own die. Its section geometry and its hardness together determine how far it compresses when the joint closes, and therefore whether the joint seals. A gasket that is nearly the right shape sits in the groove looking correct and does not seal.
-
The gasket the profile was designed around Its section and its hardness set how far it compresses when the joint closes. Compressed by the designed amount, it seals.
-
A "similar" gasket Nearly the right shape, it sits in the groove looking correct and never reaches the compression the joint needs — the most common cause of water ingress in a takeover.
Section through a glazing joint, schematic.
Fitting a "similar" gasket into a profile designed around a specific one is the single most common cause of water ingress in a takeover project. So the gaskets were reverse-engineered and tooled with the same care as the aluminium.
First article before purchase
Prototypes of the profiles, the nodes and the gaskets were produced and submitted to the client for quality approval before any material was purchased.
-
Measure
Each section taken from the material already on the building, and drawn.
-
Tool
Dies cut for the profiles and the gaskets, and the nodes tooled — trialled and corrected.
-
Prototype
First articles of all three families, made on the production tooling.
-
Approve
Submitted to the client for quality approval, side by side with the parts already built.
-
Produce
Only then was material bought, production started, and the components shipped to site.
That sequence is deliberate and it is the correct one for reverse-engineered work. The risk on a project like this is not in the drawing. It is in whether the physical part that comes off the tool actually matches the physical part already standing on the building — and the only way to answer that is to make one and put the two side by side.
Solar analysis: on a curved surface, every panel faces a different way
While the system was being tooled, the envelope's performance was modelled.
On a flat façade there is one orientation, or four if the building has four sides, and a single glass specification covers each of them. On a curved glass façade every panel has its own orientation, because a doubly curved surface turns continuously rather than in four steps. Each panel faces a slightly different direction, at a slightly different tilt, and therefore receives a different amount of sun over the year — and the difference between the most and least exposed panels on a roof like this is not marginal.
-
A flat façade: one orientation Every panel faces the same way and takes the same sun, so one glass specification covers the elevation.
-
A curved surface: every panel its own Each panel has its own orientation and tilt, so the solar load varies across the surface. Non- transmitting panels where it is highest, Low-E glass elsewhere — in few, contiguous zones.
Section, sun from the right. Zone boundaries schematic.
So the structure was modelled against the sun: the sun's position for each day of the year, the intensity of the radiation arriving, the angle at which it strikes each part of the surface, and the energy that radiation carries into the building through it. The output is a map of annual solar load across the whole envelope.
That map was then used to specify the glass panel by panel, in zones:
- Low-E glazing where daylight was wanted but the heat that comes with it was not
- Non-transmitting panels in the zones carrying the highest solar load, where the gain was worth more to exclude than the view or the light was worth to keep
The point of doing it this way is that you buy performance where it is needed instead of specifying for the worst case across all 2,500 m². A single high-performance specification applied uniformly is over-specified across most of a curved surface and, occasionally, still not enough in the one place that matters.
Yerevan sits at about forty degrees north, in a continental climate — a hot, high-sun summer and a cold winter that brings snow. The same roof therefore has a summer solar problem and a winter snow problem, which is why the next analysis is not a separate exercise.
Wind and snow: a shape the codes do not cover
Wind and snow loading on this envelope was established by computational fluid dynamics rather than by code coefficients alone.
That is not a preference. The load codes — EN 1991-1-4 and ASCE 7, the standards referenced across our system specifications — give pressure coefficients for shapes they define: flat roofs, pitched roofs, domes, vaults, rectangular buildings. A doubly curved roof is none of those, and a gridshell that turns from roof into wall without a break is further from them still. There is no table to read it off.
That leaves two routes. Take the most onerous code case that could plausibly apply and design the whole structure to it — heavy, expensive, and still capable of missing the local peak. Or compute the actual pressure field over the actual geometry.
Wind. CFD gives the pressure and suction distribution over the whole surface, and on a curved roof the governing case is almost always suction, not pressure. Suction peaks are local and they are entirely geometry-dependent: they sit where the flow separates, and where the flow separates depends on the exact shape. You cannot guess where they are, and a structure sized on an average is under-designed exactly at those points.
Snow. The ground snow load is a starting number, not the design load. On a curved roof snow does not lie evenly — it slides off the steeper zones and accumulates in the shallow ones, in valleys, and at changes of slope, so the local accumulation in a trough can be several times the uniform value.
And the case that governs a shell is usually the unbalanced one: snow on one side and not the other. A shell carries symmetric load efficiently, in membrane action, which is why it can be light. It carries asymmetric load in bending, which is what it is worst at. A free-form roof checked only for uniform snow has not been checked for its worst day.
-
Wind: suction where the flow separates On a curved roof the governing case is usually suction. It peaks where the flow leaves the surface — and where that happens depends on the exact shape.
-
Snow on both sides: membrane action A shell carries symmetric load within its surface, efficiently — which is why it can be light.
-
Snow on one side: bending Asymmetric load is carried in bending, which a shell is worst at. A roof checked only for uniform snow has not been checked for its worst day.
Section, loads and deflection exaggerated.
Correcting the structure on site
With the analysis complete and the components in production, the site team's first task was not to install anything. It was to put the structure back where it belonged.
That work involved:
- Jacking and re-levelling — bringing the out-of-position areas back to the design levels, against the deviation map produced by the survey
- Relocating load-bearing columns into their correct positions
- Working through it in a sequence that kept the standing structure supported and stable throughout
-
1 · Support Temporary support under the areas that are out of position, before anything moves. The structure is already carrying load.
-
2 · Jack and re-level Brought back to the design levels against the deviation map, and surveyed as it goes.
-
3 · Relocate and release Load-bearing columns moved to their correct positions, then the temporary support released in a planned order — and the result checked against the model before any façade component is fixed.
Section, deviation exaggerated.
Moving load-bearing elements of an erected structure is a different operation from erecting one. The structure is already carrying load, and load that is released at one point has to go somewhere. The sequence, the temporary support and the order in which corrections are released are the engineering; the jacking itself is the easy part.
The correction was verified back against the model before any façade component was fixed to the structure. There is no point manufacturing to a geometry and then installing onto a different one — which, in a sense, is the sentence that describes the whole original problem.
Then the weave
Only then did the surface go on.
On a gridshell the order in which the surface closes is part of the design, and on a curved steel structure it is not recoverable afterwards. Each panel locates against the ones already placed, so tolerance does not simply disappear — it moves, and it accumulates in whatever direction the sequence runs. Closing the surface in the wrong order pushes every small deviation toward one edge, where it arrives as a joint that cannot be closed.
-
Closed from one edge Each panel locates against the last, so small deviations add up in the direction of travel and arrive together at the far edge — as a joint that cannot be closed.
-
Locked, then closed in a set order With the geometry held and the sequence planned, the remaining tolerance is spread across many joints instead of driven into one corner.
A row of panels in plan, joint widths exaggerated.
The sequence is set so that the geometry stays locked as the weave progresses and any remaining tolerance is distributed across the surface rather than driven into a corner.
What this project demonstrates
Four things, and they are worth stating plainly because very few façade companies can claim all four on the same building:
- Taking over a live project from a departed contractor — assessing it honestly, reporting what is actually there, and completing it
- Surveying and modelling an erected structure, comparing it against its design model, and producing a deviation map that a client can make decisions from
- Producing aluminium profiles, connection nodes and gaskets for a system that cannot be bought — reverse-engineered, tooled, trialled, corrected and approved as physical first articles before any material was ordered
- Correcting a standing structure rather than absorbing its errors into the finished building
DOMAINE is a curved steel and glass envelope on a shopping centre — a building type where the envelope is the brand and the programme rarely forgives a second attempt. It is also CORTEX's first delivered project in Armenia, on Komitas Avenue in Yerevan.