Reading Swiss Apartments: From Floor Plans to Design Parameters

Apartment plans contain more information than their room areas. The relationship between a hall, a living space and a balcony is partly geometric, partly semantic, and partly dependent on how openings are represented. Extracting useful design parameters therefore requires an account of what each source record means, which records belong together, and which spatial relationships can actually be established.

This study examines Swiss Dwellings v3.0.0 through a complete element census, verified simulation joins, conditional apartment statistics and a bounded circulation analysis. Its objective is to develop an auditable reference for computational design. The result is a set of measured distributions and source examples; using them to generate and evaluate new apartments remains a subsequent research stage.

Download the research kit Explore the measured families and plans

1. Source and measurement method

Swiss Dwellings v3.0.0, published by Standfest and colleagues at Archilyse, combines semantic two-dimensional geometry, room-level simulations and building-level location data. The publisher describes a commercial-client plan collection with manual quality assurance. That provenance does not establish a probability sample of Swiss housing. The study uses the supplied release as a defined collection and does not extrapolate its frequencies to the country.

Four complete CSV files form the input: geometries.csv, simulations.csv, locations.csv and location_ratings.csv. Their sizes and CRC32 checksums match the saved release ZIP directory; SHA-256 hashes identify the analysed copies. The complete original ZIP was not assembled or checked against its published MD5. Earlier partial downloads were excluded from the analysis.

All 3,255,905 geometry records parsed as valid, nonempty polygons under Shapely's individual-geometry checks. Of these, 522,183 are semantic areas; the remaining records represent elements such as separators, openings and fixtures. Per-polygon validity establishes neither a consistent assembled floor plan nor a closed three-dimensional model.

The geometry uses site-local metres, with +x east and +y north. Different sites do not share a common origin suitable for direct geographic overlay. Area and perimeter are calculated from the source polygons, retaining holes. Minimum rotated enclosing rectangles provide orientation-aware short and long dimensions, aspect ratio and fill: polygon area divided by enclosing-rectangle area. These are descriptive shape measures.

2. Establishing apartment identity

The most consequential preprocessing decision is the join key. Area identifiers repeat across plans and floors: the geometry contains 207,147 duplicate area_id values globally. Joining on that column alone would multiply records and contaminate subsequent statistics. The verified join instead uses the full hierarchy:

(site_id, building_id, plan_id, floor_id, unit_id, area_id)

After normalising the numeric .0 suffix in geometry area IDs, both sides are unique on this composite key. All 367,678 simulation rows join exactly once, with zero apartment-ID or usage conflicts. The 154,505 area records without simulations remain visible. Much of that difference reflects semantic scope, including shafts and public areas, but it also includes 81 habitable-type records across 17 residential keys. Missing simulations are not zero-valued performance observations.

Apartment labels also repeat across sites. Grouping therefore uses (site_id, apartment_id). Even this grouping requires diagnosis: among matched residential keys, 1,628 span floors, 1,662 span units and five span buildings. Some multi-floor groups may be genuine multi-level homes; that distinction was not resolved case by case.

One concrete exception is the label d41d8cd98f00b204e9800998ecf8427e, the MD5 hash of an empty string. At site 11863 it combines 65 units across seven floors, producing 219.5 source rooms and 5,730.4 m². Its occurrence at 11 residential sites strongly suggests a placeholder-label problem, although the original encoding intent is unconfirmed. Those keys are retained in the audit evidence and excluded from the design-reference cohort.

The resulting cohort contains 42,242 site/apartment keys at 1,401 sites: one building, floor and unit per key; source net area between 20 and 250 m²; source room count between one and seven; no unmatched habitable areas; and no empty-string-hash label. These are research eligibility conditions. Excluded apartments are not thereby classified as poor designs.

3. Apartment programs and area budgets

Across all 44,947 matched residential keys, median source net area is 78.9 m², with the 10th and 90th percentiles at 47.1 and 113.9 m². That pooled statistic mixes different programs. Conditioning on source room count produces a more useful reference for design questions.

A 2.5-source-room family has a median net area of 61.7 m² and a p10–p90 band of 49.6–74.0 m². For 3.5 source rooms, the corresponding values are 82.2 m² and 69.7–98.7 m², across 9,464 reference apartments. For 4.5 source rooms, they are 100.7 m² and 85.8–119.7 m². Percentiles describe this collection's variation; they do not prescribe minimum, target or maximum areas.

The source's room-count weights are part of the measurement definition. A ROOM or BEDROOM contributes one, while LIVING_DINING contributes 1.5. Kitchens, bathrooms and corridors contribute zero. A common 3.5-room program therefore combines a living/dining zone with two generic/private-room slots, plus supporting spaces. A generic ROOM is not automatically a verified bedroom, and a separate kitchen polygon can still be openly connected to living/dining.

Median corridor share in the 3.5-room family is 12.0% of source net area. This is useful evidence that circulation deserves an explicit area budget. It does not follow that an 82.2 m² apartment should be assembled by adding independently sampled median room sizes: the medians need not sum to a coherent program. A future generator should start from complete observed programs and retain joint relationships before modifying dimensions.

4. Room shape and exterior space

Among matched residential room instances, generic ROOM polygons have a median area of 14.4 m² and BEDROOM polygons 14.0 m². Their median enclosing-rectangle aspect ratios are approximately 1.4, and median fills are close to one. Living/dining areas have a median area of 28.4 m² but a lower median fill of 0.870; corridors have a median fill of 0.854. These statistics support different shape families rather than a universal rectangular-room model.

An enclosing rectangle also cannot establish passage clearance. The median corridor rectangle's short side is 2.11 m, compared with 1.44 m for the source's biggest-inscribed-rectangle width. The two measurements describe different geometric objects. Neither establishes the narrowest unobstructed route through a turning hall.

Exterior-space accounting introduces another semantic distinction. Among matched apartment keys, 77.9% include a BALCONY polygon and 8.8% a LOGGIA; the flags can overlap. Approximately 89.0% of the stricter reference cohort includes at least one of the five explicit exterior labels: balcony, loggia, terrace, patio or garden. These are collection-specific frequencies.

Zero source-net-area contribution does not mean outdoors. Both balconies and staircases can contribute zero, while loggias can contribute positively. Terraces, gardens and patios lack simulation rows in this release. The analysis therefore keeps semantic exterior area separate from source net area, using full area geometry for exterior labels. Adding the two totals would double-count positively contributing loggias.

5. Circulation as a sensitivity study

Spatial adjacency is weaker evidence than a usable connection. To investigate circulation, the study selects 400 eligible source plans with seed 42 and one apartment per selected plan. This is a bounded sample, separate from the stricter design-reference cohort. Nearby floor context supplies doors, windows and walls.

A door creates a candidate edge only when it is sufficiently close to exactly two room polygons. Open-passage candidates require at least 0.5 m of nearby boundary after subtracting buffered source walls. Ambiguous doors remain flagged rather than connecting every nearby room indiscriminately.

The inferred graph is connected in 217 of 400 cases at 1 cm, 387 at 5 cm, and 358 at 10 cm. The decline at the largest tolerance matters: increasing proximity can introduce additional ambiguous associations and reduce accepted edges. The 5 cm setting is an empirical sensitivity choice for this analysis, not a universal snapping tolerance or a claim about source precision.

At 5 cm, the sample contains 2,751 door edges and 386 open-passage candidates. Frequent pairs include hall–generic-room, hall–bathroom and living/dining–balcony. These edge counts suggest a hall-centred family with connected living, kitchen and outdoor zones as one possible design starting point. They are not apartment percentages or a trained graph grammar.

No pairwise overlap between simulated room areas was found in the sampled cases. That check does not cover every wall or fixture, prove a gap-free apartment shell, or establish an entrance-rooted accessible route. Door operation, furniture, clear widths and physical obstacles require additional tests.

6. Environmental and location context

The 369-column simulation file contains eight identity/usage fields and 361 measures. Those include layout, sun, view, noise and connectivity descriptors. The source's solar fields describe specified times on three dates in 2018. They cannot be treated as measured contemporary daylight or annual energy performance, and they cannot be transferred unchanged to a newly generated plan.

Building-level context also requires its own unit of analysis. The study excludes the five cross-building apartment keys and compares aggregates at building level, retaining missing context. In the examined set, preselected Spearman rank correlations between context and median apartment size or corridor share all have absolute values below 0.15. Overall location rating versus median net area is approximately +0.04. The results provide little support for a simple rule that a higher-rated location implies a larger apartment.

Blank amenity fields remain missing because their meaning is not settled for every record. Likewise, source window annotations describe incidence rather than daylight adequacy. About 96.3% of matched room records have a height of exactly 2.6 m; alongside the publisher's warning about assumed heights, this repetition supports using height as an adjustable schematic default rather than an empirical ceiling-height distribution.

7. Implications for computational design

The strongest result is a usable measurement framework: coherent identities, explicit eligibility, conditional programs and geometric descriptors that retain their definitions. A subsequent apartment-design experiment could use these as descriptive priors while evaluating newly constructed geometry separately.

1. Choose a brief and coherent program. Select a source-room family and an area budget, retaining observed relationships among rooms and circulation. Resolve intended bedroom use explicitly.
2. Construct circulation before detailed packing. Provide an entrance-rooted graph, then place rooms, real portals and wall allowances. Allow open layouts and more complex living spaces rather than forcing every plan into one corridor template.
3. Evaluate the actual proposal. Test valid room geometry, overlap, shell coverage, openings, unobstructed routes and fixture clearances; run environmental evaluation in the proposed site context.
4. Separate repeated plans during evaluation. Retain site and plan groups when making training or validation splits. Compare against straightforward baselines and inspect resemblance to source examples.

The repetition issue is measurable: 44,947 matched apartment keys reduce to 26,389 conservative signatures of site, source plan and sorted simulated area IDs. Median net area changes from 78.9 m² per apartment instance to 80.6 m² per signature. These signatures remove some repeated-floor instances but do not globally deduplicate architectural designs. Randomly splitting room rows could therefore give a misleading impression of generalisation.

No apartment generator was trained or validated in this study. The proposed pipeline is a research direction, not a demonstrated generation result. National representativeness, structural feasibility, construction detail and legal compliance also remain outside the evidence.

8. Reproduction, evidence and limits

The research kit contains the four original analysis scripts, a compact cohort CSV, sampled topology records, six exact source examples, summary JSON files, methods and attribution. The README distinguishes checking the included snapshot from rebuilding the census from the publisher's four source CSVs. Original multi-gigabyte input files and large derived caches are excluded.

For a lightweight check, extract the kit, open a terminal in its folder and run:

python check_snapshot.py

This standard-library check verifies packaged-file hashes, recomputes the reference cohort's conditional net-area quantiles and corridor medians, and recounts sampled graph connectivity. It does not repeat the original full-geometry extraction. The original study used Python 3.13.9, pandas 2.3.3, NumPy 2.2.6, PyArrow 21.0.0, Shapely 2.2.0 and Matplotlib; full reproduction commands and cache conditions are in the kit.

The saved study records complete simulation-join checks and agreement between polygon and source simulation areas to a maximum relative difference of approximately 1.97e-8. An independent shoelace check on 512 seeded area polygons records a maximum relative difference below 6e-16. These checks support the extraction method; they do not establish the performance of a future generator. This article reuses that dated research run, with fresh checks of the compact snapshot rather than another full census.

The practical contribution is a traceable path from semantic source plans to conditional design parameters, with explicit boundaries around each inference. The next research question is whether those parameters improve a small set of generated plans under direct geometry and circulation tests.