A designer analyses the past monthly energy-meter readings of a building to estimate how much electricity it will use next summer. Which kind of analytics is this?

Predictive analytics, because it uses past data to estimate what is likely to happen next

Descriptive analytics, because it summarises the readings that have already been recorded

Predictive analytics, because it uses past data to estimate what is likely to happen next

Prescriptive analytics, because it tells the designer exactly which action to take

Diagnostic analytics, because it explains why the earlier readings were high or low

An energy simulation of a new office block returns a very low annual cooling load. Which step best establishes that this result can be trusted?

Compare the model with measured data or a hand calculation for a similar case, and check its inputs

Re-run the same model with a finer time step until the printed numbers stop changing at all

Render the building at a much higher resolution so that the geometry looks accurate and complete on screen

Compare the model with measured data or a hand calculation for a similar case, and check its inputs

Export the model to an open exchange format and import it back into the same program to compare

Which of the following statements about pattern recognition, simulation and data analytics are correct?

In pattern recognition each item is described by features, and similarity is measured as a distance between them

Repeating a simulation many times with randomly varied inputs gives a spread of outcomes rather than a single value

In pattern recognition each item is described by features, and similarity is measured as a distance between them

A simulation whose equations are solved correctly is thereby validated against the real building

Repeating a simulation many times with randomly varied inputs gives a spread of outcomes rather than a single value

A strong correlation found in building data proves that one variable causes the change in the other

IFC (Industry Foundation Classes) is best described as

an open, vendor-neutral schema for exchanging BIM objects and their information between programs

a rendering engine that produces photorealistic walk-through animations of a BIM model for clients

a proprietary file format that can be opened only by the software of the company that created it, not by others

a scale that grades how reliable a model element is, from conceptual design to the as-built stage

an open, vendor-neutral schema for exchanging BIM objects and their information between programs

In a federated BIM model, a duct passes beside a beam without touching it, but it runs through the zone that the project reserves around the beam for insulation and maintenance access. What will clash detection report?

A soft clash, because a reserved clearance is infringed though the solids do not touch

A hard clash, because the two elements overlap in volume and cannot both be built in place

A soft clash, because a reserved clearance is infringed though the solids do not touch

A 4D clash, because two trades want the same area of the site in the same week of the programme

No clash, because clash detection reports only elements that physically intersect one another

Which of the following statements about BIM are correct?

At LOD 300 the geometry is precise enough for documentation, and at LOD 400 elements carry fabrication and assembly detail

In 4D BIM the model elements are linked to the construction schedule, so the build sequence can be simulated

At LOD 300 the geometry is precise enough for documentation, and at LOD 400 elements carry fabrication and assembly detail

In 4D BIM the model elements are linked to the construction schedule, so the build sequence can be simulated

5D BIM adds the energy performance of the building to the 3D geometry

An IFC file can be opened only in the same program that wrote it

Which statement correctly distinguishes augmented reality (AR) from virtual reality (VR)?

AR keeps the real surroundings visible and overlays digital content, whereas VR replaces the user’s view with a computer-generated scene

AR replaces the user’s view with a computer-generated scene, whereas VR keeps the real surroundings visible and overlays digital content

AR always needs a head-mounted display with head tracking, whereas VR can be viewed on any ordinary phone screen without special hardware at all

AR is used only for printed plans and sections, whereas VR is used only for finished photographs of completed buildings

A path-traced interior render shows visible noise at 100 samples per pixel. If the noise is proportional to 1/√N, about how many samples per pixel are needed to halve it?

400, because noise falls as 1/√N, so four times the samples halve it

200, because halving the noise needs twice the samples, as with most linear quantities

400, because noise falls as 1/√N, so four times the samples halve it

141, because the samples must grow by the square root of 2 to halve the noise

10,000, because halving the noise needs the sample count to be squared (100 × 100)

What most clearly distinguishes a shape grammar from a parametric model?

A shape grammar applies rules A → B to the shapes it finds, so the structure can change; a parametric model keeps fixed relationships and varies values

A shape grammar can be run only on a computer, whereas every parametric model has to be drawn out by hand on paper before it is used

A shape grammar optimises a numerical objective function over many candidates, whereas a parametric model contains no rules or relationships of any kind

A shape grammar applies rules A → B to the shapes it finds, so the structure can change; a parametric model keeps fixed relationships and varies values

A shape grammar rewrites only strings of text symbols that stand for rooms, whereas a parametric model rewrites geometric shapes directly

A generative study of a school block returns 40 candidate designs scored on two conflicting goals: maximum daylight and minimum facade area. Eight candidates are non-dominated. What is the significance of these eight?

None can improve on one goal without losing on the other: they form the Pareto front, and the designer must choose among trade-offs

Each of the eight is better than all the other 32 designs on both goals at once, so any one of them may simply be chosen and built

They are the last eight designs that the genetic algorithm produced before it stopped running at the final generation

They are the only designs that satisfy the planning constraints, so the other 32 were illegal and must be discarded from the study

A 6 mm diameter round cutter is used to mill a rectangular pocket in a timber panel. Which feature will the finished pocket necessarily show?

Internal corners rounded to a radius of at least 3 mm

Internal corners rounded to a radius of at least 6 mm

Sharp internal corners but external corners rounded to 3 mm

Visible layer lines on the floor of the pocket

Which of the following statements about digital fabrication methods are correct?

CNC milling is a subtractive process in which a round cutter limits the smallest internal corner radius

Laser cutting removes a thin strip of material, the kerf, so slot-together joints must be drawn allowing for it

CNC milling is a subtractive process in which a round cutter limits the smallest internal corner radius

An overhanging balcony printed by filament extrusion needs no supports, because the filament hardens instantly in mid-air

Laser cutting removes a thin strip of material, the kerf, so slot-together joints must be drawn allowing for it

A 3D-printed part is equally strong in every direction because its layers fuse completely

An architectural firm trains a neural network on 5,000 daylight simulations of office layouts so that designers see a predicted daylight score while they sketch. This network is best described as

a surrogate model, a fast learned approximation of a slower simulation, traded against some accuracy

a shape grammar that rewrites office plans by applying rules to the shapes it finds in them

a federated model that combines the architectural and services models so that clashes can be detected

a reinforcement-learning agent that is rewarded each time a wall is drawn in a legal position

a surrogate model, a fast learned approximation of a slower simulation, traded against some accuracy

In usability tests of a BIM-viewer app, most participants need nine steps to find a section drawing and several give up, although all of them rate the screen design as attractive. Which diagnosis is most accurate?

The UI is attractive but the UX is poor, because the task flow, not the surface styling, is the obstacle

The UX is satisfactory because the visual design was rated well, so only a change of colour remains to be done

The problem is the raster resolution of the interface icons, which should be raised to the print standard

The UI is attractive but the UX is poor, because the task flow, not the surface styling, is the obstacle

Neither is at fault, since users who give up are always outside the intended audience of the app

Which of the following statements about the use of AI in architectural design are correct?

A model trained on data from one climate may perform poorly when applied to a very different climate

Overfitting means a model has memorised its training examples and fails on examples it has not seen

Supervised learning needs examples that carry labels, such as plans tagged with their room types

A model trained on data from one climate may perform poorly when applied to a very different climate

An image generated from a text prompt is a buildable design, because the model has already checked structure and code compliance

Overfitting means a model has memorised its training examples and fails on examples it has not seen

Supervised learning needs examples that carry labels, such as plans tagged with their room types

Computer Applications in Architecture: Computation, BIM and Visualisation, Parametric and Generative Design, Fabrication and AI

This chapter answers Section B1.5 of Part B1, “Computer Applications in Architecture”: “Fundamental principles of computational methods; Logic-based understanding of pattern recognition, simulation, and data analytics; Computer applications in building design and performance analysis; Architectural visualization: 3d Modelling and BIM, Rendering, Virtual Reality, Augmented Reality; Computational design: parametric design, shape grammar and generative design; Visual communication; User Interface and User Experience Design; Digital design and fabrication methods: CNC mining [milling] and 3d printing; Artificial Intelligence for architecture.” It teaches algorithms and loops, the logic of pattern recognition, simulation and analytics, the raster and vector arithmetic behind images, 3D representations and BIM (LOD, IFC, clash detection, 4D and 5D), rendering, VR and AR, visual communication and UI/UX, parametric, shape-grammar and generative design, subtractive and additive fabrication, and AI. The numericals are image file and print size, polygon and voxel counts, loop-count tracing, parametric combinations and grids, render time, and print and milling time from stated rates.

1. Computational methods: algorithms, pattern recognition, simulation and analytics

A computational method turns a problem into steps that a machine can carry out without judgement: inputs, a finite list of unambiguous rules (an algorithm) and outputs. Every algorithm is built from three control structures: sequence (one step after another), selection (if a condition holds do this, otherwise that) and iteration (repeat for a count, or while a condition holds). The “if” that flags a room as too deep for daylight and the loop that places 40 columns on a grid are the same logic that runs all architectural software. A computer is fast and exact but literal: it handles what has been formalised as numbers, rules and geometry, so qualities that cannot be formalised, such as the sense of a place, stay with the designer.

Tracing pseudo-code is the standard numerical here. A loop “for i = a to n” runs n − a + 1 times (the usual slip is an off-by-one). Nested loops multiply: 5 × 4 gives 20 passes. When the inner bound depends on the outer variable, add instead: for i = 1 to 4 with an inner loop j = 1 to i, the body runs 1 + 2 + 3 + 4 = 10 times. A halving loop (while n > 1, n = n/2 with integer division) runs about log₂ n times: from n = 100 the values are 50, 25, 12, 6, 3, 1, which is 6 passes. Cost grows with the loop structure, proportional to n for one loop and to n² for two nested ones, which is why testing every pair among 10,000 elements for clashes needs a smarter search than brute force.

Pattern recognition finds regularity in data. Describe each item by features, measure similarity as a distance, then classify into known groups or cluster into unknown ones. The rules can be written by hand (a closed loop of walls enclosing a minimum area is a room) or learnt from labelled examples. Uses: recognising walls and openings in a scanned point cloud, grouping similar plans, reading hand sketches.
Simulation builds a model (simplified rules with parameters), splits space and time into small cells and steps, and applies an update rule again and again, so it is iteration at scale: sun position at hourly steps, heat flow through wall layers, people leaving a hall, airflow in a CFD mesh, stress in finite elements. It is only as good as its model and inputs. Verification asks whether the equations were solved correctly; validation asks whether they are the right equations for reality. Repeating a run with random inputs (Monte Carlo) gives a spread of outcomes, not one number.
Data analytics turns data into decisions: descriptive (what happened), diagnostic (why), predictive (what is likely next) and prescriptive (what to do). Sources include meter readings, occupancy sensors and post-occupancy surveys. Data must be cleaned, a correlation is not a cause, and a biased sample gives a biased answer.

⚠️ A simulation does not prove a design

A daylight or energy simulation answers the question of its model, not of the building. Change the weather file, the occupancy schedule or the material data and the answer changes. Use results to compare options under identical assumptions, and check a model against a hand calculation or a measurement before trusting it.

2. Computer applications in building design and performance analysis

Computers enter every stage: brief and site (GIS, survey and point-cloud data), concept (massing and sketch modelling), development (CAD and BIM), analysis (performance simulation), documentation (drawings, schedules, quantities), construction (4D and fabrication files) and operation (building management and facility data). The earlier a tool is used, the cheaper a change is, so quick massing-level checks of sun, daylight and energy belong to the concept stage, with detailed analysis once the form is stable.

Performance analyses and what they compute
Analysis	What it computes	Key inputs	Typical metric
Daylight	Illuminance on the working plane under a sky model	Geometry, glazing transmittance, reflectances, location	Daylight factor, % (indoor ÷ unobstructed outdoor illuminance under overcast sky)
Energy	Heating, cooling and lighting loads over a year	Envelope U-values, schedules, weather data, systems	Energy use intensity, kWh/m²·year
Airflow (CFD)	Air speed and pressure fields	Geometry, wind boundary, openings	Air speed, air changes per hour
Structure (finite elements)	Stress, deflection, vibration modes	Geometry, supports, loads, material	Stress, deflection
Egress	Movement of people to exits	Plan, population, exit widths	Evacuation time

Images are of two kinds. Vector drawings store lines and curves as mathematical objects, so they scale without loss and files stay small: CAD plans are vector. Raster images store a grid of pixels at a fixed resolution, so enlarging blurs them. Uncompressed raster size = width × height × bytes per pixel. 24-bit colour is 3 bytes per pixel (8 bits each for red, green and blue, giving 2²⁴ ≈ 16.7 million colours). A 4000 × 3000 pixel image holds 12 million pixels, so it is 12 × 3 = 36 MB uncompressed (1 MB = 10⁶ bytes). Print size = pixels ÷ pixels per inch (ppi): at 300 ppi the same image prints 13.3 × 10 inches, and 1 inch = 25.4 mm.

⚠️ Resolution is not file size, and dpi does not add detail

Compression (for example JPEG) shrinks the file, not the pixel count. Changing the dpi tag of an image does not add detail; it only changes the print size at which the same pixels are spread. Re-sampling upward invents pixels but no information.

3. 3D modelling and BIM: LOD, IFC, clash detection, 4D and 5D

Ways of representing a 3D form
Representation	What it stores	Strength	Limit
Wireframe	Edges and vertices only	Light and fast	No surfaces or volume; can be ambiguous
Surface (NURBS)	Smooth, mathematically defined skins	Exact free-form curvature in compact data	No inside or outside unless the skin is closed
Mesh	Vertices, edges and flat polygons	Fast to render; the format of scanning and printing	Smoothness needs many polygons; it is an approximation
Solid	Closed volumes with Boolean union, difference and intersection	Knows inside from outside; gives volume and mass	Heavier to edit
Voxel	A 3D grid of small cubes carrying values	Volumetric data: scans, density, analysis	Memory grows with the cube of the resolution

Two counting rules are examination favourites. A quad is two triangles, so a cube whose six faces are each divided into 4 × 4 quads has 6 × 16 = 96 quads, or 192 triangles. A voxel model of a volume L × W × H at voxel edge s has (L/s)(W/s)(H/s) voxels, so halving the edge multiplies the count by 8. A 12 m × 8 m × 6 m volume at 0.5 m has 24 × 16 × 12 = 4,608 voxels.

BIM (building information modelling) describes a building by intelligent, parametric objects such as walls, doors and slabs; each carries geometry and data (material, fire rating, cost, phase, manufacturer). Plans, sections, schedules and quantities are views generated from one database, so they cannot disagree, and a change is made once. Plain 3D CAD, by contrast, stores geometry only. Models from different disciplines are combined into a federated model held in a common data environment that all parties share.

LOD (level of development): how far a model element’s geometry and information can be relied on at a stage. On the common scale, 100 is conceptual (symbols and masses), 200 approximate generic elements, 300 precise geometry for documentation, 350 adds connections to other systems, 400 fabrication and assembly detail, and 500 the verified as-built condition. In computer graphics, LOD instead means level of detail: a simpler mesh substituted for a distant object.
IFC (Industry Foundation Classes): an open, vendor-neutral data schema, standardised internationally as ISO 16739, for exchanging BIM models between different software with the objects and their data intact, not just their shapes.
Clash detection automatically tests the federated model for conflicts. A hard clash is two elements occupying the same space (a duct through a beam); a soft clash violates a required clearance (no room to open a door or to service equipment); a 4D or workflow clash is a conflict in time and space on site. Finding them on screen costs far less than finding them after concrete is poured.

The dimensions of BIM
Dimension	What it adds to the model	Typical use
3D	Geometry and object properties	Design, coordination, visualisation
4D	Time: the schedule linked to elements	Construction sequence simulation, site logistics, progress comparison
5D	Cost: quantities linked to rates	Automatic quantity take-off and a cost estimate that updates with the design

ℹ️ Beyond 5D

Labels such as 6D and 7D are used differently by different organisations (sustainability, facility management), so they are not fixed facts to memorise. The firm ones are 3D geometry, 4D time and 5D cost.

4. Rendering, ray tracing, virtual reality and augmented reality

Rendering computes an image from a 3D scene: geometry, materials, lights and a camera. Two families dominate. Rasterisation projects each polygon onto the screen pixels and shades it with a local lighting formula; it is fast, so it drives real-time views, games and VR, but indirect light and true reflections are approximated. Ray tracing works backwards from the camera: for each pixel it sends a ray into the scene, finds the first surface hit, and spawns further rays toward the lights (shadows), along reflections and through glass (refraction). Path tracing is Monte Carlo ray tracing that follows many random bounces per pixel and so captures global illumination: colour bleeding, soft shadows, light reaching a room by bouncing off the floor. Its images are noisy at low sample counts; noise falls as 1/√N, so four times the samples halve the noise.

Render time scales with pixels and samples. Full HD (1920 × 1080) is 2,073,600 pixels; 4K UHD (3840 × 2160) is exactly four times that, so the same scene at the same quality takes about four times as long: a 25-minute full-HD render becomes about 100 minutes in 4K. Realism also depends on physically based materials (reflectance, roughness, transparency), textures mapped onto surfaces, accurate light sources and exposure. A non-photorealistic render (line, sketch or diagram style) is chosen when the aim is to explain, not to imitate a photograph.

Virtual reality and augmented reality compared
Feature	Virtual reality (VR)	Augmented reality (AR)
View of the real world	Replaced entirely by a virtual scene	Kept; virtual objects are overlaid on it
Typical device	Head-mounted display with head tracking	Phone, tablet or see-through glasses
Key technical need	Low latency and six-degree-of-freedom tracking	Registration: aligning virtual content with the real scene
Architectural use	Walk-through of an unbuilt design, client review, training	On-site overlay of a design or of hidden services; comparing progress with the model

ℹ️ Degrees of freedom and mixed reality

Three degrees of freedom (3DoF) track only head rotation, so you can look around; six (6DoF) add three translations, so you can walk around and lean in. Delay between head movement and the updated image causes discomfort, which is why VR needs fast real-time rendering. Mixed reality goes one step beyond AR: virtual objects are anchored to the surfaces of the real room and respond to them.

5. Visual communication, user interface and user experience

Visual communication is the deliberate design of drawings, diagrams, renders and presentations so that a specific audience (client, authority, contractor) understands the intent quickly. Its working principles are hierarchy (the main idea is seen first), contrast, alignment, proximity (related things grouped), repetition (consistent line weights, symbols and colours), balance and white space. Drawings need scale, north point, legend and readable text. Output also depends on colour model: screens mix light, so RGB is additive (red + green + blue = white), whereas print mixes inks, so CMYK is subtractive (cyan, magenta, yellow and black ink absorb light). A design built for screen in RGB can look duller in print, and print needs about 300 ppi where a screen is satisfied with far less.

User interface (UI) is what the user sees and touches: layout, controls, icons, typography, colour and feedback. User experience (UX) is the whole experience of reaching a goal: usefulness, ease, speed, trust, accessibility and how errors are handled. UI is a part of UX, and a handsome interface can still give a poor experience if the task flow is confusing. In architecture the same thinking applies to modelling software, client-facing VR walk-throughs, building dashboards and the wayfinding of the building itself.

User-centred process: understand users (interviews, observation, personas) → define tasks → sketch flows → build a wireframe (low-fidelity layout), then a mockup (visual design) and a clickable prototype → test with real users → revise and repeat.
Usability heuristics (Nielsen): visibility of system status, match with the real world, user control and freedom, consistency and standards, error prevention, recognition rather than recall, flexibility and efficiency, minimalist design, help in recovering from errors, and documentation.
Two laws: Fitts’s law says the time to reach a target grows with its distance and falls as it gets larger, so frequent controls should be big and close; Hick’s law says decision time grows with the number of choices, so reduce options or reveal them progressively.
Accessibility: adequate colour contrast, resizable text, keyboard and screen-reader support, and never conveying meaning by colour alone (colour-blind users).

⚠️ UI is not UX

Restyling buttons and colours changes the UI; it does not fix a task that needs nine steps or an error message that says nothing. UX problems are found by watching users attempt real tasks, not by asking whether the screen looks nice.

6. Computational design: parametric design, shape grammar and generative design

Parametric design defines a design by parameters (variables such as span, height, number of divisions, angle) and relationships (rules, equations, constraints) that tie the geometry to them. Change a parameter and the whole model regenerates. The logic is usually drawn as a dependency graph or visual program: input nodes feed function nodes, which feed the output geometry. It is excellent for exploring variants, keeping many repeated parts consistent and rationalising a complex surface into a few panel types. Its limits: a design can only vary in the ways the rules allow, and a tangled graph is hard to edit. A BIM door family, resizable by width and height, is already parametric.

Parametric counts follow simple grid arithmetic. A facade W wide and H high divided into nx × ny panels has nx × ny panels and (nx + 1)(ny + 1) grid nodes. An 18 m × 12 m facade in 1.5 m square panels gives 12 × 8 = 96 panels and 13 × 9 = 117 nodes. With k settings for each of n independent parameters, the design space holds kⁿ combinations: 4 settings of 5 parameters give 4⁵ = 1,024, which shows why exhaustive search soon becomes impossible.

A shape grammar, introduced by Stiny and Gips in the early 1970s, generates designs by shape rules of the form A → B: wherever the shape A can be found in the current design (under translation, rotation, scaling or reflection), it may be replaced by B. Generation starts from an initial shape, and each sequence of rule applications is a derivation; all designs a grammar can produce form its language. Rules act on shapes seen as drawings, not on labels, so new shapes can be seen that were never drawn, such as a new shape appearing in the overlap of two squares (emergence). Markers or labels steer the order of rules, and a design is finished when no marked, non-terminal part remains. Grammars have been written to describe an existing body of work (analytical grammars, the best known being for Palladian villas) and to invent new designs (original grammars).

Generative design reverses the usual order. The designer states goals (maximise daylight, minimise facade area), constraints (site, FAR, structure) and variables; the software generates many options, scores each with a performance simulation and improves them by optimisation. A genetic algorithm keeps a population of candidates, scores each with a fitness function, selects the fitter ones, combines them (crossover), perturbs them at random (mutation) and repeats for many generations. With several conflicting goals no single best exists; the result is a Pareto front, the set of solutions where no goal can be improved without worsening another, and the designer chooses among these trade-offs. The danger is a badly stated goal: the algorithm optimises exactly what is measured, so a result can be optimal and still wrong.

Three computational design approaches
Approach	Core idea	Designer supplies	Output
Parametric	Fixed relationships; values vary	The rules and the parameter values	One design family that updates with every change
Shape grammar	Rewriting rules applied to shapes	Initial shape and the shape rules	A language of designs, each with a derivation
Generative	Search guided by goals and scoring	Goals, constraints and variables	Many ranked options, often a Pareto front

7. Digital design and fabrication: CNC milling, 3D printing, laser cutting

Digital fabrication carries a model to a machine with no drawing in between: CAD for the geometry, CAM to plan the toolpaths, and a controller that moves the machine, usually from G-code. CNC milling spins a cutter along programmed toolpaths and removes material from a block, so it is subtractive. Three-axis machines move in X, Y and Z; five-axis machines add two rotations so the tool can reach undercuts and curved surfaces. A round cutter cannot cut a sharp internal corner: the smallest internal radius equals the cutter radius, though external corners can be sharp. Milling suits timber, foam, metal and stone, and is used for moulds and formwork, joinery, models and facade panels. A useful rate is material removal rate = cut width × cut depth × feed rate: 10 mm × 5 mm × 600 mm/min removes 30,000 mm³ (30 cm³) per minute, so clearing 9,000 cm³ takes 300 minutes.

3D printing is additive: the model is sliced into thin layers and the machine deposits or cures material layer by layer. Common processes are FDM/FFF (melted filament extruded), SLA (liquid resin cured by light), SLS (powder fused by a laser, the unfused powder supporting the part) and, at building scale, extrusion of concrete or earth for walls. Geometry that is costly to cut, such as internal voids and organic forms, is almost free to print and waste is small; but printing is slow, build size is limited, layer lines show, parts are weaker between layers, and overhangs need supports. Counting is simple: layers = height ÷ layer height, so a 120 mm part at 0.2 mm layers has 600 layers, and at 90 s per layer prints in 54,000 s = 15 hours. Where a deposition rate is given, time = volume ÷ rate.

Laser cutting is a 2D process: a focused laser follows vector lines through a sheet of card, acrylic, plywood or thin board, and can also engrave. The beam removes a strip of material, the kerf, so slotted joints must be drawn allowing for it. It is the usual tool for models, slot-together sections and waffle structures. Cutting stacked 2D profiles gives a stepped 3D form; milling gives a true 3D surface; printing builds the form up.

Subtractive, additive and 2D cutting compared
Feature	Subtractive (CNC milling)	Additive (3D printing)	Laser cutting
Principle	Remove material from a block	Build up layer by layer	Cut 2D profiles from a sheet
Waste	High: chips and offcuts	Low; supports may be discarded	Sheet offcuts; the kerf strip
Geometry limits	Tool access and cutter radius	Overhangs need support; build volume	Flat profiles only
Typical use	Moulds, joinery, panels, models	Complex models, voids, building-scale extrusion	Models, slot-together sections

⚠️ Additive does not mean faster

Printing wastes less material but is usually the slowest of the three for a large solid part, and the printed part is anisotropic: weaker across layer lines than along them. “Digital” also does not remove tolerances: stock sizes, tool wear, kerf and material movement still decide whether parts fit.

8. Artificial intelligence for architecture

Machine learning lets a computer learn patterns from data instead of following hand-written rules. In supervised learning the examples are labelled (plans tagged by room type; a building’s features paired with its measured energy use) and the model learns to predict the label. Unsupervised learning finds structure without labels (clustering floor plans by similarity). In reinforcement learning an agent learns by trial and reward. A neural network is layers of weighted nodes; deep learning uses many layers; convolutional networks are suited to images. Training uses one set of data, and a separate unseen set tests the result, because a model can overfit, memorising its training examples and failing on new ones.

Generative models (diffusion models and GANs for images, language models for text and scripts) produce new images or text from learned patterns; in design they serve ideation, mood boards and early massing studies, and language models can write scripts for parametric tools.
Surrogate models: a network trained on thousands of simulation results predicts daylight or energy in milliseconds instead of minutes or hours, giving live feedback while the designer sketches, at some loss of accuracy.
Computer vision extracts geometry and meaning from photographs, drone imagery and point clouds (scan-to-BIM), and monitors construction progress against the model.
Other uses: site and context analysis, plan-layout suggestion, cost and schedule prediction, automated code-compliance checking and smart-building control.

⚠️ What AI cannot do for you

A generated image is not a buildable design: it has no structure, services, code compliance or cost behind it. Models inherit the bias and gaps of their training data, and one trained on one climate or building type can fail in another. Many are black boxes that cannot explain a result, and questions of copyright and credit remain open. The architect stays responsible for the outcome, so AI is used with a human in the loop.

Key takeaways

Every algorithm is sequence, selection and iteration: a loop from a to n runs n − a + 1 times and nested loops multiply. Simulation is iteration over a model, only as valid as its model and inputs; pattern recognition classifies or clusters by features; analytics is descriptive, diagnostic, predictive or prescriptive.
Raster size = width × height × bytes per pixel (24-bit = 3 bytes) and print size = pixels ÷ ppi; vector drawings scale without loss. A voxel count is (L/s)(W/s)(H/s), so halving the voxel edge gives 8 times as many.
BIM is one database of parametric objects carrying data: LOD grades reliability (300 documentation, 400 fabrication), IFC is the open exchange schema, clash detection finds hard and soft conflicts, 4D adds time and 5D adds cost.
Ray tracing follows rays per pixel for shadows, reflections and refraction; path tracing adds global illumination and its noise falls as 1/√N; render time scales with pixel count. VR replaces the real view, AR overlays on it; UI is the surface and UX the whole experience.
Parametric = fixed relationships with varying values (kⁿ combinations); shape grammar = rules A → B on shapes from an initial shape, with emergence; generative = goals, constraints, scoring and a Pareto front.
CNC milling and laser cutting remove or cut material, 3D printing adds it layer by layer (layers = height ÷ layer height); a round cutter leaves an internal radius and a laser leaves a kerf. AI learns from data, can act as a surrogate for slow simulations, but overfits, inherits bias and leaves responsibility with the architect.

Practice questions (20)

Attempt each one before opening the answer. Every explanation names the tempting wrong option as well as the right one, because that is where marks are lost.

A designer analyses the past monthly energy-meter readings of a building to estimate how much electricity it will use next summer. Which kind of analytics is this?
1. Descriptive analytics, because it summarises the readings that have already been recorded
2. Predictive analytics, because it uses past data to estimate what is likely to happen next
3. Prescriptive analytics, because it tells the designer exactly which action to take
4. Diagnostic analytics, because it explains why the earlier readings were high or low
Show answer
Answer: B — Predictive analytics, because it uses past data to estimate what is likely to happen next
Using past readings to estimate a future quantity is predictive analytics. Descriptive analytics only summarises what has already happened (totals, averages, charts) and is the tempting answer because the input is historical data. Diagnostic analytics explains causes, and prescriptive analytics recommends an action.
An energy simulation of a new office block returns a very low annual cooling load. Which step best establishes that this result can be trusted?
1. Re-run the same model with a finer time step until the printed numbers stop changing at all
2. Render the building at a much higher resolution so that the geometry looks accurate and complete on screen
3. Compare the model with measured data or a hand calculation for a similar case, and check its inputs
4. Export the model to an open exchange format and import it back into the same program to compare
Show answer
Answer: C — Compare the model with measured data or a hand calculation for a similar case, and check its inputs
Trust comes from validation: checking the model against reality (measurements or an independent hand calculation) and checking that its inputs, such as weather and occupancy, are realistic. A finer time step only verifies that the equations are solved consistently, which is the tempting answer; a model can be solved perfectly and still be the wrong model. Rendering resolution and file export say nothing about the physics.
Trace the pseudo-code.
count = 0
for i = 1 to 6
for j = 1 to 6
if (i + j) is divisible by 4 then count = count + 1
The final value of count is ____.
Numerical answer — type the value.
Show answer
Answer: 9
The loops try all 36 pairs (i, j) with i and j from 1 to 6. The sum i + j lies between 2 and 12, so it is divisible by 4 only when it equals 4, 8 or 12. Sum 4: (1,3), (2,2), (3,1) = 3 pairs. Sum 8: (2,6), (3,5), (4,4), (5,3), (6,2) = 5 pairs. Sum 12: (6,6) = 1 pair. Total = 3 + 5 + 1 = 9. Counting all 36 pairs, or forgetting sum 12, are the usual slips.
Which of the following statements about pattern recognition, simulation and data analytics are correct?
1. In pattern recognition each item is described by features, and similarity is measured as a distance between them
2. A simulation whose equations are solved correctly is thereby validated against the real building
3. Repeating a simulation many times with randomly varied inputs gives a spread of outcomes rather than a single value
4. A strong correlation found in building data proves that one variable causes the change in the other
Show answer
Answer: A — In pattern recognition each item is described by features, and similarity is measured as a distance between them; C — Repeating a simulation many times with randomly varied inputs gives a spread of outcomes rather than a single value
Feature description followed by a distance measure is the standard recipe of pattern recognition, and Monte Carlo repetition with random inputs produces a distribution of results. Correct solution of the equations is verification, not validation, which needs comparison with reality. A correlation, however strong, does not prove causation: both variables may follow a third one, such as occupancy.
IFC (Industry Foundation Classes) is best described as
1. a rendering engine that produces photorealistic walk-through animations of a BIM model for clients
2. a proprietary file format that can be opened only by the software of the company that created it, not by others
3. a scale that grades how reliable a model element is, from conceptual design to the as-built stage
4. an open, vendor-neutral schema for exchanging BIM objects and their information between programs
Show answer
Answer: D — an open, vendor-neutral schema for exchanging BIM objects and their information between programs
IFC is an open, internationally standardised schema, so a model can move between programs with objects such as walls and doors, and their data, intact. The grading scale of reliability is LOD, which is the most tempting wrong answer. IFC is neither a renderer nor a proprietary format; being non-proprietary is its whole purpose.
In a federated BIM model, a duct passes beside a beam without touching it, but it runs through the zone that the project reserves around the beam for insulation and maintenance access. What will clash detection report?
1. A hard clash, because the two elements overlap in volume and cannot both be built in place
2. A soft clash, because a reserved clearance is infringed though the solids do not touch
3. A 4D clash, because two trades want the same area of the site in the same week of the programme
4. No clash, because clash detection reports only elements that physically intersect one another
Show answer
Answer: B — A soft clash, because a reserved clearance is infringed though the solids do not touch
A soft clash is a violation of a required clearance or access zone, even when the solid geometry does not touch. A hard clash needs actual overlap of volumes, which is the tempting choice only if the reserved zone is ignored. A 4D clash is a conflict of sequence in time and space on site. Modern clash tools can test clearances, so the claim that only physically intersecting elements are reported is false.
A site volume 24 m × 18 m × 10 m is represented by cubic voxels of 0.5 m edge. The voxel edge is then refined to 0.25 m. The number of additional voxels needed after refinement is ____.
Numerical answer — type the value.
Show answer
Answer: 241920
At 0.5 m the grid is 48 × 36 × 20 = 34,560 voxels. At 0.25 m it is 96 × 72 × 40 = 276,480 voxels, which is 8 times as many because halving the edge doubles the count along each of three axes. Additional voxels = 276,480 − 34,560 = 241,920. Answering 276,480 gives the new total, not the additional count, and doubling the count (a factor of 2) forgets the three dimensions.
Which of the following statements about BIM are correct?
1. At LOD 300 the geometry is precise enough for documentation, and at LOD 400 elements carry fabrication and assembly detail
2. In 4D BIM the model elements are linked to the construction schedule, so the build sequence can be simulated
3. 5D BIM adds the energy performance of the building to the 3D geometry
4. An IFC file can be opened only in the same program that wrote it
Show answer
Answer: A — At LOD 300 the geometry is precise enough for documentation, and at LOD 400 elements carry fabrication and assembly detail; B — In 4D BIM the model elements are linked to the construction schedule, so the build sequence can be simulated
LOD 300 is precise geometry for documentation and LOD 400 adds fabrication and assembly detail; 4D means time, the schedule linked to elements. 5D adds cost (quantities linked to rates), not energy, which is the tempting slip. IFC exists precisely so that other programs can open the model, so the claim that only the writing program can open it is false.
Which statement correctly distinguishes augmented reality (AR) from virtual reality (VR)?
1. AR keeps the real surroundings visible and overlays digital content, whereas VR replaces the user’s view with a computer-generated scene
2. AR replaces the user’s view with a computer-generated scene, whereas VR keeps the real surroundings visible and overlays digital content
3. AR always needs a head-mounted display with head tracking, whereas VR can be viewed on any ordinary phone screen without special hardware at all
4. AR is used only for printed plans and sections, whereas VR is used only for finished photographs of completed buildings
Show answer
Answer: A — AR keeps the real surroundings visible and overlays digital content, whereas VR replaces the user’s view with a computer-generated scene
AR overlays digital content on the real view, which is why it needs registration to align virtual objects with the real scene; VR substitutes the entire view. The claim that AR replaces the view merely swaps the two definitions, a common slip under time pressure. AR is typically viewed on phones, tablets or see-through glasses, not necessarily a head-mounted display.
A rendered image of 6000 × 4000 pixels is stored uncompressed at 32 bits per pixel (24-bit colour plus an 8-bit transparency channel). Its size in megabytes (1 MB = 10⁶ bytes) is ____.
Numerical answer — type the value.
Show answer
Answer: 96
Pixels = 6000 × 4000 = 24,000,000. 32 bits per pixel is 4 bytes, so size = 24,000,000 × 4 = 96,000,000 bytes = 96 MB. Using 3 bytes (colour only) gives 72 MB, and forgetting to divide the bits by 8 gives a figure eight times too large. Compression or a different dpi tag would not change the uncompressed size.
A path-traced interior render shows visible noise at 100 samples per pixel. If the noise is proportional to 1/√N, about how many samples per pixel are needed to halve it?
1. 200, because halving the noise needs twice the samples, as with most linear quantities
2. 400, because noise falls as 1/√N, so four times the samples halve it
3. 141, because the samples must grow by the square root of 2 to halve the noise
4. 10,000, because halving the noise needs the sample count to be squared (100 × 100)
Show answer
Answer: B — 400, because noise falls as 1/√N, so four times the samples halve it
Noise ∝ 1/√N, so halving it needs √N to double, hence N to become four times: 100 → 400 samples, and the render takes about four times as long. The linear answer of 200 is the tempting slip, giving only 1/√2 of the noise; 141 would reduce noise by a factor of about 1.19, and 10,000 confuses the square root with a square.
What most clearly distinguishes a shape grammar from a parametric model?
1. A shape grammar can be run only on a computer, whereas every parametric model has to be drawn out by hand on paper before it is used
2. A shape grammar optimises a numerical objective function over many candidates, whereas a parametric model contains no rules or relationships of any kind
3. A shape grammar applies rules A → B to the shapes it finds, so the structure can change; a parametric model keeps fixed relationships and varies values
4. A shape grammar rewrites only strings of text symbols that stand for rooms, whereas a parametric model rewrites geometric shapes directly
Show answer
Answer: C — A shape grammar applies rules A → B to the shapes it finds, so the structure can change; a parametric model keeps fixed relationships and varies values
The defining feature of a shape grammar is rule application on shapes seen as drawings: A → B wherever A is found, starting from an initial shape, so the topology of the result changes during the derivation. A parametric model keeps its relationships fixed and only the values change. Optimising an objective belongs to generative design, and a shape grammar works on shapes, not strings, which is the inversion in the claim about text strings.
A generative study of a school block returns 40 candidate designs scored on two conflicting goals: maximum daylight and minimum facade area. Eight candidates are non-dominated. What is the significance of these eight?
1. None can improve on one goal without losing on the other: they form the Pareto front, and the designer must choose among trade-offs
2. Each of the eight is better than all the other 32 designs on both goals at once, so any one of them may simply be chosen and built
3. They are the last eight designs that the genetic algorithm produced before it stopped running at the final generation
4. They are the only designs that satisfy the planning constraints, so the other 32 were illegal and must be discarded from the study
Show answer
Answer: A — None can improve on one goal without losing on the other: they form the Pareto front, and the designer must choose among trade-offs
A non-dominated solution cannot be improved on one goal without a loss on the other, so these eight form the Pareto front and choosing among them is a judgement about trade-offs. They are not better than every other design on both goals; that would make one clear winner, which conflicting goals rule out. Non-dominance concerns scores, not the order of generation or legality.
A parametric facade study varies six independent parameters: three have 4 allowed settings each, two have 3 settings each, and one is a two-way switch. The number of distinct design combinations is ____.
Numerical answer — type the value.
Show answer
Answer: 1152
For independent parameters the number of combinations is the product of the settings: 4 × 4 × 4 × 3 × 3 × 2 = 64 × 9 × 2 = 1,152. Adding the settings (3×4 + 2×3 + 2 = 20) is the usual slip; it confuses a product with a sum, which is also why exhaustive search becomes impossible so quickly as parameters are added.
A 6 mm diameter round cutter is used to mill a rectangular pocket in a timber panel. Which feature will the finished pocket necessarily show?
1. Internal corners rounded to a radius of at least 3 mm
2. Internal corners rounded to a radius of at least 6 mm
3. Sharp internal corners but external corners rounded to 3 mm
4. Visible layer lines on the floor of the pocket
Show answer
Answer: A — Internal corners rounded to a radius of at least 3 mm
A round cutter cannot reach into a sharp internal corner, so the smallest internal radius equals the cutter radius, which is half of 6 mm = 3 mm. The 6 mm answer confuses diameter with radius. External corners of the part can be sharp, so the claim of sharp internal but rounded external corners is reversed, and layer lines belong to additive printing, not milling.
A 3D-printed concrete wall segment is 1.35 m tall and is built in layers 9 mm thick. Every layer, including the last, takes 3 minutes to print and is followed by a 1-minute pause for the material to stiffen. Ignoring set-up time, the total time to complete the segment is ____ hours.
Numerical answer — type the value.
Show answer
Answer: 10
Layers = 1350 mm ÷ 9 mm = 150. Each layer cycle is 3 + 1 = 4 minutes, so total = 150 × 4 = 600 minutes = 10 hours. Leaving out the pause gives 450 minutes = 7.5 hours, and dividing 1.35 by 9 without converting metres to millimetres gives a nonsense layer count.
Which of the following statements about digital fabrication methods are correct?
1. CNC milling is a subtractive process in which a round cutter limits the smallest internal corner radius
2. An overhanging balcony printed by filament extrusion needs no supports, because the filament hardens instantly in mid-air
3. Laser cutting removes a thin strip of material, the kerf, so slot-together joints must be drawn allowing for it
4. A 3D-printed part is equally strong in every direction because its layers fuse completely
Show answer
Answer: A — CNC milling is a subtractive process in which a round cutter limits the smallest internal corner radius; C — Laser cutting removes a thin strip of material, the kerf, so slot-together joints must be drawn allowing for it
Milling is subtractive and a round cutter leaves an internal radius; a laser cutter removes a kerf that slotted joints must allow for. Filament printing of overhangs normally needs support structures because the material does not harden in mid-air, which is why SLS, with its unfused powder, is often preferred for complex overhangs. Printed parts are anisotropic: weaker across layer lines than along them.
An architectural firm trains a neural network on 5,000 daylight simulations of office layouts so that designers see a predicted daylight score while they sketch. This network is best described as
1. a shape grammar that rewrites office plans by applying rules to the shapes it finds in them
2. a federated model that combines the architectural and services models so that clashes can be detected
3. a reinforcement-learning agent that is rewarded each time a wall is drawn in a legal position
4. a surrogate model, a fast learned approximation of a slower simulation, traded against some accuracy
Show answer
Answer: D — a surrogate model, a fast learned approximation of a slower simulation, traded against some accuracy
Training on simulation outputs to predict them quickly is a surrogate model: it trades some accuracy for speed, which gives live feedback. A shape grammar uses explicit rewriting rules and learns nothing from data, a federated model is a combination of discipline models, and reinforcement learning needs a reward signal from interaction rather than a labelled set of simulation results.
In usability tests of a BIM-viewer app, most participants need nine steps to find a section drawing and several give up, although all of them rate the screen design as attractive. Which diagnosis is most accurate?
1. The UX is satisfactory because the visual design was rated well, so only a change of colour remains to be done
2. The problem is the raster resolution of the interface icons, which should be raised to the print standard
3. The UI is attractive but the UX is poor, because the task flow, not the surface styling, is the obstacle
4. Neither is at fault, since users who give up are always outside the intended audience of the app
Show answer
Answer: C — The UI is attractive but the UX is poor, because the task flow, not the surface styling, is the obstacle
The attractive screen shows that the UI (the surface) is fine, but nine steps and abandonment show that the experience of completing the task is poor. That is a UX problem, found by watching users do a real task, and a restyle cannot fix it; the flow must be shortened. Rating the look well is the tempting reason to declare the UX satisfactory, which confuses UI with UX.
Which of the following statements about the use of AI in architectural design are correct?
1. A model trained on data from one climate may perform poorly when applied to a very different climate
2. An image generated from a text prompt is a buildable design, because the model has already checked structure and code compliance
3. Overfitting means a model has memorised its training examples and fails on examples it has not seen
4. Supervised learning needs examples that carry labels, such as plans tagged with their room types
Show answer
Answer: A — A model trained on data from one climate may perform poorly when applied to a very different climate; C — Overfitting means a model has memorised its training examples and fails on examples it has not seen; D — Supervised learning needs examples that carry labels, such as plans tagged with their room types
Models inherit the range of their training data, so a change of climate or building type can degrade them; overfitting is memorising rather than generalising; supervised learning learns from labelled examples. A generated image has no structure, services, code compliance or cost behind it, so calling it buildable is the tempting overstatement, and responsibility stays with the architect.