Comp

Five arenas of computation

Design, Communication, Capturing Reality, Fabrication, and Operation — the five areas where computation reshapes architectural practice

Wicked problem (Rittel & Webber)

A design problem with no clean test for right/wrong and no stopping rule; every framing already commits the designer to a particular approach

Notes on the Synthesis of Form (Alexander, 1964)

Proposes decomposing a design problem into sub-requirements, then synthesizing them into form via a tree of diagrams

A Pattern Language (Alexander et al., 1977)

Organizes architectural knowledge as a network of patterns, each a problem-response pair with confidence rankings and cross-links

WikiHouse

Open-source modular timber housing system designed for small CNC machines, with the entire detail catalogue open and forkable

Frei Otto's soap-bubble experiments

Physical models where surface tension computes minimal-surface shell forms, without a digital solver

Fun Palace (Price & Littlewood, 1961)

Concept of a building as an event-driven machine: rearrangeable modules, gantries, and screens

Architecture Machine Group (Negroponte, MIT)

Late-1960s research aiming to make the computer a dialogue partner in the design process

Alexander's three design processes

Direct intervention (vernacular building), intuitive composition (sketching), and formal designing (building an explicit model first)

Four families of design models

Parametric, rule-based/procedural, optimization-based, and machine-learning-based models

Parametric model

A design expressed as a fixed structure with variable inputs; changing a parameter updates the geometry

Shape grammar (Stiny & Gips, 1972)

A set of replacement rules applied repeatedly to generate complex compositions from simple seeds

Palladian Grammar (Stiny & Mitchell, 1978)

Showed that explicit composition rules could generate Palladio's villas plus a larger family of "impossible" villas following the same grammar

Sol LeWitt / instruction vs. execution

Idea that in rule-based/procedural design, the instruction is the artwork and the execution is downstream, raising questions of authorship

Optimisation (in design)

Define a fitness function and let an algorithm search for inputs that maximize it, once the design space is too large to browse by hand

Bird's Nest Stadium / Kaisersrot

Genetic algorithm used to breed structural lattice configurations, balancing constructability against expression

Perceptron (Rosenblatt, 1958)

Early neural-network hardware trained on photographs to distinguish categories, an ancestor of today's deep learning

Classification and embedding

A network projects inputs (images, floor plans) into a latent space where similar examples land near each other, enabling interpolation

Conditional generation (pix2pix)

A network learns to map one type of image to another, e.g. a semantic facade map to a photorealistic facade

"Designing as sampling"

Framing generative AI as recomposing existing material from a corpus, comparable to Roman spolia or architectural kit-bashing

PBR vs. Generative AI

Two opposite rendering pipelines: PBR uses explicit programmed rules to simulate light; GenAI learns a mapping entirely from data

Rendering vs. Photorealistic vs. PBR

Rendering: any 3D-to-image process; Photorealistic: indistinguishable from a photo; PBR: simulates physical light behavior to predict what is observed

Forward ray tracing (photon tracing)

Rays are shot from the light source and bounce until they hit the camera; physically correct but very inefficient

Backward ray tracing (eye-ray tracing)

Rays are shot from the eye through each pixel into the scene; the basis of all modern ray tracers, since every ray contributes to the image

Cornell Box

A physical benchmark scene used to compare a rendering against a real photograph

Global illumination

Traces many recursive light bounces, including indirect light; the most physically accurate rendering approach

BRDF

Function describing how a surface reflects light: the ratio of outgoing to incoming light energy for given directions

Phong reflection model

A classic shading model combining ambient, diffuse, and specular components

Diffuse shading formula

Irradiance falls off with 1/(4pir^2); brightness is scaled by the cosine (foreshortening) term between the surface normal and light direction

Snell's Law

n1sin(theta1) = n2sin(theta2); describes how light bends across a material boundary depending on the index of refraction

Total internal reflection

When light travels from a high- to a low-IOR medium beyond the critical angle, all light is reflected and none escapes

Fresnel equations

Describe how much light is reflected vs. transmitted at an interface, depending on angle, wavelength, polarization, and IOR

Microfacet theory (D, F, G)

Models rough surfaces as a statistical distribution of tiny facets: D = distribution, F = Fresnel term, G = shadowing/masking

Data-driven BRDFs

Physical measurement (e.g. via a spherical gantry) captures real material reflectance as a look-up table instead of an analytic formula

Autoencoder

An encoder compresses input data to a low-dimensional latent code; a decoder reconstructs the original from it

Variational Autoencoder (VAE)

Encodes inputs as probability distributions rather than points, producing a continuous, smooth, generative latent space

Diffusion model

Learns to reverse a process of progressively adding noise to an image, generating new images by denoising pure noise

Latent diffusion (Stable Diffusion)

Runs the diffusion process in a compressed latent space; a text prompt is injected via CLIP encoding to steer generation

ControlNet

Adds a structural conditioning input (edge map, depth map, pose) to a pretrained diffusion model for precise compositional control

Digital Michelangelo Project (1996)

Early structured-light scanning project (Curless & Levoy) that digitized Michelangelo's sculptures and architecture

Points/Surfaces vs. Fields paradigms

Classical methods (LiDAR, SfM) produce explicit geometry; modern field methods (NeRF, 3DGS) represent scenes as continuous density/color fields

LiDAR

Active sensor measuring distance via time-of-flight of laser pulses; produces a colored point cloud with millimeter accuracy

Stereo vision / disparity

Passive depth estimation from the horizontal shift (disparity) of matching features between two camera views

Depth from disparity formula

Z = f * B / d — depth equals focal length times baseline divided by disparity

Epipolar geometry

Constrains the stereo correspondence search from a 2D area to a 1D line (the epipolar line), given calibrated cameras

Structured light / active stereo

Projects a known pattern onto a surface; its deformation reveals depth, even on textureless surfaces

Structure from Motion (SfM)

Reconstructs 3D structure and camera poses simultaneously from many unordered photographs

"Building Rome in a Day"

Landmark SfM project that reconstructed parts of Rome from about 75,000 tourist photographs in roughly a day

COLMAP

Widely used open-source Structure-from-Motion pipeline, developed at ETH Zurich

NeRF (Neural Radiance Fields, 2020)

Implicit neural function mapping 3D position and viewing direction to color and density, rendered via volume rendering

Volume rendering (NeRF)

Pixel color is computed by integrating color and opacity contributions along a ray through sampled points

Limitations of NeRF

Static scenes only, no editing, one network needed per scene, and slow to both train and render

3D Gaussian Splatting (3DGS, 2023)

Represents a scene as an explicit, differentiable set of 3D Gaussian "blobs," rendered in real time via GPU rasterization

3DGS Gaussian parameters

Mean (position), covariance (shape/orientation), opacity, and color (via spherical harmonics)

SuGaR

Adds regularization constraints to 3DGS to align Gaussians with true surfaces, improving mesh extraction

Four regimes of making

Craft (the hand), Industry (the mould), Digital Fabrication (the code), and Biofabrication (the organism)

Semper's four techniques (+ a fifth)

Textile, Ceramics, Tectonics, and Stereotomy; digital fabrication is proposed as a possible fifth technique

Jacquard loom

Often cited as the first digital machine: a fabrication device taking instructions from punched cards rather than a human hand

Binder-jet printing (Digital Grotesque)

Deposits binder droplets into a bed of loose sand, fusing grains; offers near-total geometric freedom at sub-millimeter resolution

Tor Alva

The tallest 3D-printed building in the world (Mulegns, Switzerland, 2025): hollow printed concrete columns with integrated reinforcement and bolted, disassemblable joints

4D printing

Printed parts that change shape after fabrication, using material composition built into the part's micro-structure

Mario Carpo — second digital turn

Argues digital fabrication ends standardization's cost advantage: similarity, not identity, becomes the basic unit of production

Adolf Loos — "Ornament und Verbrechen" (1910)

Polemic declaring ornament a crime, tied to an industrial moment when custom detail meant extra cost

Mycelium-based building materials

Fungal networks grown through agricultural waste to produce low-carbon panels for interior fit-out

Bio-cementation

Bacteria deposited in a granular substrate grow a cement-like binder over weeks by metabolizing CO2

Picoplanktonics (Ling, 2025)

Printed forms embedding living cyanobacteria that keep photosynthesizing and sequestering CO2 throughout the work's lifetime

Space Popular / portals

Treats the threshold between augmented and physical reality as an architectural element architects will increasingly design

Immersive design studio (ETH)

Lets designers inhabit unbuilt spaces at full scale using real-time game-engine rendering

Ambient capture and privacy

As 3D capture of buildings becomes continuous and ambient, privacy stops being an afterthought and becomes part of the design

Buckminster Fuller's geodesic dome

Argued for a design science in which computation serves global problems through geometric efficiency

General Panel System (Wachsmann & Gropius)

Argued affordability requires industrialization: prefabricated modules that pack flat and ship globally

BUGA Wood/Fibre Pavilions (Menges & Knippers)

Industrial robots weave or place fibre composites into ultra-lightweight, individually portable structures

NCCR Digital Fabrication

Over a decade, demonstrated robotic equivalents for nearly every traditional construction process

Reyner Banham — Well-Tempered Environment

Argued that servicing systems (like air conditioning) are more architecturally consequential than any move in plan or facade

Villa Girasole (1935)

A building that rotated bodily on a circular track to follow the sun

Amazon fulfilment centres

Shelves reconfigure continuously around workers — architecture designed, operated, and maintained by a building that is also a computer

Models and what they are for

Models are useful because they are not the thing itself: they suppress detail and fix some variables to make a space of possibilities legible

Michael Polanyi — "we can know more than we can tell"

Captures why machine learning is useful for things easy to recognize but nearly impossible to write down as explicit rules

Autodesk Spacemaker / Forma

Applies rule-based logic to urban massing studies, showing the consequence of each rule change in real time

Toyo Ito's Serpentine Pavilion (2002)

Geometry generated by an iterative subdivision-and-rotation rule, executed manually without a computer

Sou Fujimoto facade optimization

Apparently random balconies are typically the output of optimizing for a minimum quota of sunlight and view per unit

"Unreasonable effectiveness of data" (Norvig/Halevy)

Argues that with enough data, explicit models may become unnecessary — a provocation for architectural theory

Fitness function (optimisation)

Encodes what "good" means numerically; choosing it is a value judgement disguised as a number

Farnsworth House rendering (opening example)

Alessandro Prodan's photorealistic image, chosen to show how imagery communicates atmosphere that blueprints cannot

Key roles of visualisation

Design refinement, creative decision-making, storytelling (spatial narrative), and as a sales tool for client presentations

Depth of field

A real camera-optics effect where objects outside the focal plane appear blurred

Subsurface scattering

Light enters a material, bounces inside, and exits at a different point; essential for realistic marble and skin

Participating media / volumetric effects

Light scattering within fog, smoke, or clouds; required for atmospheric rendering

Plato's "divine fire" (~350 BC)

Hypothesised that fire emitted by the eye encounters light from objects — the earliest conceptual form of eye-ray tracing

Dürer's ray tracing device (1525)

Used a string and assistant to mechanically establish linear perspective in drawings

Arthur Appel (1968)

Introduced the first computer ray casting algorithm, using one ray per pixel plus shadow rays

Don Greenberg (1974)

Created one of the first architectural pre-visualisations: an art museum rendered before construction

Pinhole camera / camera obscura

Camera model where a pixel's colour corresponds to the unique ray passing through that pixel and the pinhole

Conductors vs. dielectrics

Metals reflect light via free electrons (opaque); dielectrics absorb and re-emit light internally, allowing transparency

AI milestones

ImageNet (2015, surpassed human benchmark), AlphaGo (2016), and generative models (2022–present: DALL-E, Stable Diffusion, Sora)

GenAI in architectural practice (examples)

Foster + Partners' The Forestias, Zaha Hadid Architects (Midjourney/Stable Diffusion), and Hickok Cole's Aitopia (ChatGPT as "creative partner")

Biological vs. artificial neuron

Biological: dendrites, soma, axon. Artificial: weighted inputs, summation, and a nonlinear activation function

Training via backpropagation

Forward pass computes output, loss is measured, backward pass computes gradients via the chain rule, weights update via SGD, repeated many times

Why architects need 3D scanning

As-built modelling, high-fidelity site context, preservation/adaptive reuse, and digital fabrication integration

Computer vision as inverse of computer graphics

Graphics goes 3D scene to image; computer vision goes image to 3D scene