In the mid-1960s, in a wind tunnel at the TU Berlin, Ingo Rechenberg and Hans-Paul Schwefel shaped a jointed plate without solving for it. They changed the joint angles at random, measured the drag, kept the better plate and repeated. Each generation was one physical experiment. There was no equation for the optimum and no drawing of the answer. They called the method evolution strategies.

That loop is the foundation of evolutionary architecture. It also shows the one decision inside it that no algorithm makes for you: what “better” means. Someone chose drag. The plate did the rest.

What it is: Evolutionary architecture is design by breeding. You do not draw the building. First you write a genome, the parameters that describe which forms are possible: bay widths, blade tilts, node positions, member sections. Then you write a fitness function, the score that defines “fit”. Last, you set the rules of variation, which decide how candidates recombine and mutate. Selection then works through hundreds or thousands of candidates, generation after generation. PAZ’s own concept panel on the subject says the form that comes out at the end “is less an invention than a survivor.” The unit of design stops being one object and becomes a population.

Read those three inputs as an auditor would. The genome controls what can exist. The variation rules control how far the search wanders. The fitness function controls what wins. Whoever writes the score designs the building. Everyone else reviews the survivors.

Why it works: Evolutionary search needs no gradients. You do not differentiate the problem. All you need is a score for each candidate. That matters because building design has rugged search spaces. Truss topology is full of on/off choices, such as whether a bar exists or not. Steel sizing jumps between standard sections. Façade panelisation and shading layouts have many local optima. A gradient solver can stop on a local ridge and report it as the summit. A population samples many ridges at once, and recombination lets good partial answers from different ridges meet.

The structural logic is older than the software. In 1904 A.G.M. Michell showed that the lightest frames align their members with the directions of principal strain, and that lightly stressed material is mostly dead weight. Evolutionary structural optimisation (ESO), published by Y.M. Xie and G.P. Steven in 1993, moves toward Michell’s load paths numerically. You run a finite-element analysis, remove the least-stressed elements and run it again. The bidirectional variant, BESO, lets material grow back where stress concentrates.

Goals often compete: stiffness against mass, winter gain against summer overheating, cost against embodied carbon. In those cases NSGA-II, published by Kalyanmoy Deb and colleagues in 2002, returns a Pareto front. That is the set of designs where no objective can improve without another getting worse. It is the method’s most honest output. It maps the trade-off. It does not make it.

The people who built the tools state the limits themselves. David Rutten wrote Galapagos for Grasshopper in 2010. In his 2013 Architectural Design essay “Galapagos: On the Logic and Limitations of Generic Solvers”, he set out that generic solvers are slow, cannot promise a global optimum and earn their place through exploration. The cost is simple arithmetic. A population of 40 over 60 generations means 2,400 evaluations, and each one may be a full FE or daylight run. Plan your budget in evaluations, not generations.

←TODAY: In 2026 every Grasshopper file is one component away from a population search, and whoever opens the canvas writes the fitness function.
→3012: In Zurich-3012 the score is public record: versioned, signed and open to challenge by the people the building houses.
Fulcrum: Evolution moves authorship up into the rules, and governance means deciding who may write them.

Origins: The concept got its name in 1995, when John Frazer published An Evolutionary Architecture with AA Publications in London. The cybernetician Gordon Pask wrote the foreword. The book collected years of work at the Architectural Association with Julia Frazer and their students. They treated buildings as organisms, each with a genetic code, rules for growth and an environment that rewards some variants and lets others die out.

The line goes back further. D’Arcy Wentworth Thompson’s On Growth and Form (1917) turned fish, crabs and skulls into one another on coordinate grids. For him, form was the trace of forces, not of taste. John Holland’s Adaptation in Natural and Artificial Systems (1975) made Darwin executable as the genetic algorithm. Cedric Price’s Generator (1976–79) was an unbuilt retreat for the Gilman Paper Corporation at White Oak, Florida: a kit of cubes, screens and walkways that a crane could move. John and Julia Frazer wrote the programs that let it propose new layouts, even when its users had stopped changing it. Accounts that credit Generator to Price alone leave out the two people who wrote its intelligence. Richard Dawkins’ Biomorphs in The Blind Watchmaker (1986) put the same loop on a home computer.

At the AA, Michael Hensel, Achim Menges and Michael Weinstock collected morphogenetic strategies in the Architectural Design issue “Emergence” (2004). At the University of East London, Paul Coates taught a generation of architects to program their own tools (Programming.Architecture, 2010). At ETH Zürich, Kristina Shea’s Engineering Design and Computing Lab carries computational design synthesis into engineering. Her structural shape annealing runs a close relative of the same cycle: generate, analyse, accept or reject, repeat. In 2017 the borrowing went the other way. Neal Ford, Rebecca Parsons and Patrick Kua published Building Evolutionary Architectures, and Pramod Sadalage joined them for the second edition. They used the fitness function to keep software systems healthy while they change.

In practice: Built projects show where the score can sit. At the Qatar National Convention Centre (2011), Arata Isozaki and Mutsuro Sasaki used extended ESO to grow a branching steel canopy along a front of roughly 250 metres. There the fitness function is structural, and its result is the face of the building. For the ICD/ITKE Research Pavilion 2013–14 in Stuttgart, Achim Menges and Jan Knippers borrowed a score that biology had already run: the double-layered forewings of flying beetles. Robots wound the glass- and carbon-fibre modules without a mould.

For governance, the precedent is Project Discover, the Autodesk office in Toronto’s MaRS Innovation District (2017). It was designed by The Living with Autodesk Research: David Benjamin, Danil Nagy, Damon Lau, John Locke, Jim Stoddart, Lorenzo Villaggi, Ray Wang and Dale Zhao. A multi-objective genetic algorithm ranked around 10,000 floor layouts against six measured goals, from adjacency and work-style preference to daylight and views. Employee surveys fed the scores, so the people who would work in the building helped write the function that chose it. The thing to copy is not the layout. It is the way the future users agreed to the score.

On a Swiss desk, this happens in Grasshopper on Rhino. Galapagos handles one objective. Octopus, by Robert Vierlinger at Bollinger+Grohmann, and Wallacei, by Mohammed Makki, Milad Showkatbakhsh and Yutao Song, handle several objectives and draw the front. Wallacei is built on NSGA-II. A small Zürich studio tuning a south-façade louver array, or a Bern timber engineer sizing a glulam grid, reaches for these tools once there are more variants than anyone can check by hand in an afternoon. Cite every penalty to its source: serviceability limits from SIA 260, embodied greenhouse-gas values from the KBOB Ökobilanzdaten list, targets from SIA 2040. When the survivors leave the canvas, carry them into Archicad through the PAZ Grasshopper↔Archicad Library, so the winning parameters stay attached to the element that gets built.

From where I write, in the late 2070s, the lasting damage from optimisation rarely came from the solver. It came from the default weight nobody reopened. Offices that left their carbon weighting inside a vendor’s hosted service all describe the same thing. A number was set when grey emissions were easy to ignore, and it kept steering designs for years because nobody had the standing to change it. The fix already exists in 2026. Keep the fitness function as plain, versioned text that your office owns, and you can export it, challenge it and rerun it on any solver. Write that exit clause before you write the score.

Atelier: An office adopting AI-assisted search this year meets the same question at every scale, from the louver array to the floor plate to the model that proposes it: who wrote the score, and can anyone else read it? On Monday, open the last Galapagos or Wallacei file the office ran. Write its fitness function out as one plain-language paragraph: every objective, every weight, every penalty with its SIA or KBOB source, and the random seed. File it next to the model.

Hack: Keep every non-dominated design visible, instead of letting a weighted sum quietly pick one. A weighted sum collapses three objectives into one number and hides the decision inside the weights. A Pareto filter keeps every design that no other design beats on all counts. Paste this into Rhino 8’s CPython 3 component, or any Python 3, and feed it the objective values your Wallacei or Octopus run exported:

# toy values: (embodied kg CO2e/m2, cost CHF/m2, daylight deficit 0-1), lower is better
designs = [(410, 3200, 0.18), (395, 3350, 0.22), (430, 3100, 0.15), (440, 3400, 0.25)]
beats = lambda a, b: all(x <= y for x, y in zip(a, b)) and a != b
front = [d for d in designs if not any(beats(o, d) for o in designs)]
print(front)  # three survivors; the fourth is beaten on every count

Three designs survive, and none of them is “the answer”. Each takes a different position on carbon, cost and light. Choosing between them is a decision that a named person signs.

Evolution moved authorship one level up, into the rules. Follow it there. Before the next run starts, find out who in your office writes the score, and put their name on it.