Vẽ ý đồ, để máy thử phần còn lạiDraw the intent, let the machine try the rest

The architect draws and locks the intent · the machine tries the open variables · each option comes with what it trades off

01 · Locked · the architect’s sketched intent

02 · Open · the machine tries tower height, green space and podium coverage · one run, three corners of the trade-off triangle

Best viewtallest 191 m · coverage 43 % · green 38 %

viewshadeexposure

GFA

260,763 m²

good view

52.9 %

daylight

77.3 %

gap to GFA

36.4 %

Most facade daylighttallest 40 m · coverage 42 % · green 21 %

viewshadeexposure

GFA

156,455 m²

good view

41.3 %

daylight

87.1 %

gap to GFA

61.8 %

Closest to the brief GFAtallest 158 m · coverage 62 % · green 17 %

viewshadeexposure

GFA

409,438 m²

good view

37.4 %

daylight

71.8 %

gap to GFA

0.1 %

In urban design, many fundamental decisions are hard to quantify: the orientation of the street network, the shape of the skyline, the position of green spaces. Manual methods produce few options, and view, shade and solar radiation are usually assessed only at a late stage. Automated optimisation tools, by contrast, reduce every factor to a variable, which leaves the design intent hard to control. This post proposes an approach in between: the architect keeps the intent, and the machine explores only what is left open.

What we propose

The approach separates two kinds of decision. Design intent covers the urban grid, the skyline and the position of green spaces; the architect draws and locks them. The design variables are the height of the tallest tower, the green-space ratio and the site coverage of the podiums; the architect sets a range for each, and the machine searches within those ranges. Floor area ratio follows from these variables and is not itself a variable. The architect also decides which objectives the options are measured by, and in which sense, maximise or minimise; some objectives depend on the climate, so their sense is chosen for the region of the project: direct sun on facades should be reduced in a hot, humid climate but may be sought in a cold one.

The machine's part runs through four stages: generation, evaluation, search and selection. Each stage is a slot with a defined input and output, so it can be replaced by another tool, from a hand-written script or a plug-in for design software to a simulation service. Selection presents the trade-off options, those that no other option outperforms on every objective, for the architect to decide.

The illustration

To illustrate the approach, we built a simulator that runs in the browser, on a hypothetical site. It is not a working tool for a particular project but a small model that lets readers try the division of roles described above. In Design, the user draws the intent: the site boundary, the urban grid type, green spaces, the skyline and the good-view arc. In Set bounds, the user sets ranges for the three variables and the gross floor area of the brief; the intent stays fixed, and the senses of the objectives are preset. In Choose, four cards set side by side the three options nearest each corner of the objective triangle and one balanced option; all trade-off options remain available in a three-axis chart and as triangles, and up to four options can be pinned and compared side by side.

Evaluation comprises three objectives:

• Good view: the share of the facades that sees the good-view arc the architect marks on a ring around the site. Even the criterion of "good view" is thus part of the design intent.
• Facade daylight: the share of the sky in front of each facade sample point that no building blocks, averaged over the sample points. It is a relative index, close to the vertical sky component, not a daylight factor. The taller the towers, the wider the view, but the more the facades block one another's sky; this objective therefore pulls against the view.
• Gap to the brief floor area: the shortfall or excess against the gross floor area the architect sets for the brief. Floor area is thus a requirement to meet, not a quantity to maximise.

View and facade daylight are computed by ray casting on the massing. To read each option, the illustration also displays shade over the whole ground and direct sun exposure on facades (a relative exposure index, not converted to kWh/m²); these two layers take no part in the search. A particular project may need other indicators; then only the evaluation stage changes.

An AI coding agent (Claude Code) wrote the code for grid generation, massing, evaluation, optimisation and automated testing; the architect posed the question, chose the grid forms, decided which factors belong to the intent and approved each step. The division of roles during development therefore mirrors the division the tool proposes.

What it helps decide

The approach does not produce a single answer; it shows what each option achieves and what it trades off within the limits set. In the illustration, tall towers on sparse podiums widen the view but leave less daylight on the facades; low massing gives the facades more daylight but a narrower view; the trade-off options range from those that meet the brief floor area to those with far less floor area but a better view or more daylight, and it is the architect who decides what shortfall is acceptable. Running the same design with the same ranges again gives a slightly different set of options, since the search involves chance; the results therefore serve for comparison, not as an answer. When an objective keeps pushing a variable to the edge of its range, that limit decides the outcome; setting the limits is therefore a design decision, not part of the machine's work.

Limitations

The illustration uses a hypothetical site and few building types; daylight is computed for a uniform sky, sun and shade for a clear sky, with no weather data; only buildings within the site are considered, on sparse sample points. All values serve only to compare options with one another. Applying the approach to real projects, with real data and professional tools, will be the subject of later posts.

Try the illustration

Suggestions: raise the skyline towards the good-view arc and run the search; change the grid type while keeping the ranges; change the brief floor area and see how the cards change; pin two options from different corners of the triangle and compare them side by side.