← SCENT COMMONS WIP
01:00 / 20:00
Settings
Study details

The field spreads as simulation time advances. Concentration and the mass ledger interpolate between adjacent 10-second solver samples. Path opacity crossfades between those same samples; route geometry stays fixed and arrow phase persists. Pause freezes both. Turn off EVOLVE TIME to inspect a fixed field while the arrows keep moving.
Drag to orbit · Ctrl + scroll to zoom. Field = height-averaged concentration, extruded for display. Trail colour = destination outlet; arrow speed is schematic. Not CFD or a calibrated odour map.

01 / WHAT IS THIS SPACE?

A commons where you can choose
your distance from a scent activity.

One source, in Room 01. Rooms 02 and 03 receive transfer from that source. They do not contain three different fragrances. Ventilation and boundaries create an exposure gradient; the architecture gives visitors somewhere else to go.

01

6 × 10 m / 34 SEATS

Encounter

Intent: approach the scent activity and participate. The only release is here, beside local capture.

Tested: highest exposure of the three rooms under the default calculation.

02

6 × 10 m / 33 SEATS

Buffer

Intent: share the commons at a lower exposure. A proposed place to stay without approaching the source.

Tested: a small transfer from Room 01. No independent scent source.

03

6 × 10 m / 33 SEATS

Withdraw

Intent: choose more distance and lower exposure. Separate vestibule access avoids a route through Room 01.

Tested: further reduction. “Lower exposure” does not mean odour-free or medically safe.

Does the third room earn its place? The current scent-only test has not demonstrated that. In the earlier 4,680 m³/h visitor study, a third exposure level changed the pooled lower-tail score by −0.026 against two levels. The social, acoustic and activity roles above remain design proposals to test.

02 / WHAT DATA DO WE ACTUALLY HAVE?

Real odour ratings.
Simulated exposure. Assumed utility.

These are different kinds of evidence. Public datasets tell us how people rated chemicals. The spatial model estimates transport. The current link from exposure and liking to comfort is an assumption, not a measured dose–response curve.

REAL HUMAN RATINGS / BENCHMARK COMPLETE

DREAM olfaction challenge

Local training subset: 49 people · 338 molecules.
30 designated calibration odours; 19–30 valid ratings per person. 13,210 scored target responses.

20.24 → 19.23

PERSON-AVERAGED MAE / LOWER IS BETTER

A nested-selected personal offset reduced rating error by about 5% against the odour mean. It did not improve average ranking correlation.

Nested 5-fold person holdout; inner 3-fold selection. Known odours, new people with calibration responses. Not an external test.

Original data & challenge →
REAL HUMAN RATINGS / PERSONALISATION GATE FAILED

Bierling / CROWN

Public release: 1,227 participants. Local training: 894 people; 874 scored · 6,915 responses.
Only two shared calibration odours.

22.75 → 22.87

PERSON-AVERAGED MAE / LOWER IS BETTER

Personalisation did not beat the odour-mean baseline. Two responses are not enough evidence for this method to recommend a room reliably.

Nested person holdout. DREAM and Bierling use different rating scales and protocols; their absolute errors are not pooled.

Public dataset, version 3 →

What “30,000 visitors” means

300 synthetic sessions of 100 visits, resampled from the observed ratings of 874 real respondents. Each session uses an observed odour condition and respects 34 / 33 / 33 seat capacities.

The repeated visits test a design under varied arrival and preference combinations. They are not 30,000 independent people, AI-generated personalities, or extra ground-truth training labels.

874 → 30,000

OBSERVED RESPONDENTS → RESAMPLED VISITS

+7.00POOLED TAIL VS OLD B
+1.22CHOICE VS ASSIGNMENT

Default 4,680 m³/h assumptions only. Tail = mean score of the lowest 10% of visits. Choice lost to assignment in 95 of 300 session tails; there is no every-session guarantee.

WHY THE ANIMATION IS NOT THE VISITOR SCORE

The animation is a clean-start tracer pulse: release 0–10 minutes, then purge until 20 minutes. The visitor study used its earlier steady-initialised, constant-source exposure protocol. Its default lower-layer means were 0.144233 / 0.003856 / 0.000121. The two experiments answer different questions; this presentation does not feed animated fields into old scores.

The current graph contains assumed diffusion, internal flow reconstruction, door mixing and local capture. It does not resolve furniture, thermal jets or buoyant plumes. Both source position and mixing change implicit capture. No field measurement has calibrated the release, transport or perception.

Visitor results JSON · Spatial model method · Benchmark evidence & provenance

03 / HOW DOES THIS BECOME MACHINE LEARNING?

Two models to learn.
One human decision to support.

Learning what someone likes and learning how air carries a release are separate problems. A future design tool can combine them, but public smell ratings alone cannot train an automatic building controller.

A / PERSON → PREFERENCE

TESTED / LIMITED GAIN

Predict liking

Input: odour identity, descriptors and a person's calibration ratings.

Target: their held-out pleasantness ratings, with uncertainty.

Gate: beat the odour mean on held-out people. Add an odour holdout for any claim about unseen odours. Two-rating personalisation currently fails.

B / DESIGN → EXPOSURE

NEXT / NOT YET TRAINED

Emulate transport

Input: ventilation, capture position, boundary state, door schedule, source position and time.

Target: simulated zone exposure and removal fractions.

Gate: hold out entire operating conditions; compare against direct simulation and a simple interpolation baseline. This learns the simulator, not reality.

A × B / OPTIONS → CHOICE

SCENARIO TEST / ASSUMED UTILITY

Compare designs

Combine preference and exposure distributions with capacities, travel and an explicit comfort assumption.

Report uncertainty, the lower 10% and people made worse off.

Gate: remain acceptable across dose-anchor, arrival-order and utility sensitivity. A person selects the design; moving boundaries stay manual.

LATER / LINKED OBSERVATIONS

MISSING / REQUIRES NEW DATA

Learn actual choice

Link measured exposure to intensity, comfort, chosen room, dwell and exits, with consent.

Gate: validate on held-out sessions and people. Check lower-tail protection in new observations before considering any control system.

No autonomous system is trained or validated here.

The next experiment without fieldwork: expand simulation runs across ventilation, hood offsets and door schedules; reserve complete conditions for testing; train a small exposure emulator and compare it to interpolation. Separately test calibration counts on public ratings. Do not train a comfort model on invented comfort labels.
TRAINING TABLE CONTRACT / WHAT IS AVAILABLE, WHAT IS MISSING
FIELDSEVIDENCE STATUSUSE / SPLIT RULE
dataset_id, person_id, odour_id,
pleasantness, calibration_flag
Observed public ratingsPreference task. Hold out people; preserve repeated-response groups. Add odour holdout for unseen-odour claims.
simulation_case_id, time_s,
flows, source_xyz, door_schedule
Design inputs / simulatedExposure task. Keep every timestamp of a held-out condition out of training.
zone_e, local_removed,
other_removed, stored
Simulated targetsEmulator targets only. Not ppm, measured intensity or preference labels.
bootstrap_visit_id, session_id,
source_person_id, arrival_rank
Resampled scenariosDesign stress test. Repeated visitors cannot become independent ML subjects.
measured_exposure, intensity,
comfort, room_choice, dwell, exit
Missing linked field observationsNeeded for validated dose–response and actual choice learning. No fabricated substitutes.

04 / WHAT THE MOVEMENT MEANS

Every trail has a connected route.

The exporter computes net tracer transfer every 10 simulation seconds. Between saved samples, the display linearly interpolates concentration and the mass ledger, and crossfades adjacent path drawings without bending them. Intermediate frames are a presentation approximation, including intervals containing door changes; they are not extra solver output or an instantaneous velocity solution. It extracts routes that use positive transfer edges and finish on a real extract stencil. Source- and cell-to-terminal links are dashed; they are bindings, not resolved jets. Moving markers include these short connectors, so capture remains visible even when source and hood share a cell.

The translucent field and fixed spatial grain show a height-averaged concentration. Raster smoothing stays inside each compartment. The grain stays fixed and changes opacity continuously with interpolated concentration. Each route has a stable identity, so its arrows do not reset when the route list changes. During a crossfade, paths from adjacent solver snapshots can overlap. Source status switches exactly at 10 minutes; no release-origin route is drawn after that switch. On purge snapshots, routes begin in cells with retained tracer. The source stops emitting at 10 minutes. Bright arrowheads travel continuously along the paths; they do not blink at fixed locations.

Direction is data. Speed is graphic.

Weak cross-room routes are deliberately visible. Equal line widths, arrow speed and arrow count do not encode mass shares or travel time. Opacity uses a fixed logarithmic route-flux scale with a visibility floor, so weak paths remain visible without appearing equally strong. The mass ledger separately reports cumulative local capture, other removal and stored tracer.

The three rooms, seats, boundaries and ventilation terminals use the existing Rhino geometry. The source offset is a test assumption. The model does not prove an odour-free room or a safe exposure.