Maciej Sachse
/work/mission-driven-habitability

/08 — 2026

Mission-Driven Habitability

A TU Delft studio response to mission-driven confinement at Antarctica's Troll Station: a self-supporting Voronoi interior fit-out inside the standard living container, 3D-printed from the station's plastic waste, with AI-driven circadian lighting for the polar night.

role
Group work (7 members)
location
TU Delft
org
TU Delft — MSc Architecture, Urbanism and Building Sciences
tutors
Henriette Bier (course), Arwin Hidding, Vera Laszlo, Lisa-Marie Mueller
tools
Rhino · Grasshopper · Karamba3D · Python (scikit-learn) · Arduino / ESP32 · HTML/JS · 3D printing
Mission-Driven Habitability

Overview

Troll Station sits in the Norwegian sector of Antarctica; during the polar winter, occupants spend long stretches inside a standard container-based shell, cut off from daylight and from resupply. The studio brief asked how the interior — not the envelope — could carry habitability through this period. Our team framed a worst-case stress test: two researchers remain continuously inside a single container for seven days. The scenario isn't a literal prediction — it is a way to surface the spatial, environmental and psychological demands that a normal-operation brief would hide.

Three criteria structured the proposal: agency (users can modify the space in response to changing routines), privacy (retreat remains possible under constant co-presence), and sleep (rest is treated as a spatial and environmental condition, not just a timetable). A self-supporting Voronoi fit-out with integrated foldable furniture absorbs routine changes; panels are 3D-printed from the station's own plastic waste; an AI model adjusts illuminance and correlated colour temperature in response to circadian logic, weather data, and selected physiological signals.

Geometry was developed in Grasshopper — the orthogonal container order preserved, the Voronoi infill used to articulate zones at room scale and acoustic / structural texture at panel scale. The lighting model was trained in Python with scikit-learn on a 70 / 30 split (~93% accuracy for illuminance, ~81% for CCT). Its predictions feed a Grasshopper visualisation and an ESP32 / Arduino prototype, driven either by a pre-computed 24-hour sequence or live from an HTML PhysioApp interface that lets the environment either mirror a detected state or compensate for it.

Context — Troll Station in Queen Maud Land, Antarctica: staff and housing profile (peak 35, visitor max 70, 1,500 m² under roof, 24 h showers, diesel / oil / gas / solar / wind) and annual weather (mean −25 °C; summer 9 Nov–1 Feb midnight sun; winter 15 May–27 Jul polar night)
Context — Troll Station in Queen Maud Land, Antarctica: staff and housing profile (peak 35, visitor max 70, 1,500 m² under roof, 24 h showers, diesel / oil / gas / solar / wind) and annual weather (mean −25 °C; summer 9 Nov–1 Feb midnight sun; winter 15 May–27 Jul polar night)
Annual cycle — sunlight hours (left) and number of occupants (right) per month. Mission-Driven Habitability targets the May–August window: 6 people, minimum daylight, no resupply, extreme temperatures
Annual cycle — sunlight hours (left) and number of occupants (right) per month. Mission-Driven Habitability targets the May–August window: 6 people, minimum daylight, no resupply, extreme temperatures
Voronoi logic in 2D — from a point distribution, to the Voronoi space, to chosen surfaces, to organic cells
Voronoi logic in 2D — from a point distribution, to the Voronoi space, to chosen surfaces, to organic cells
Two zones with merging schedules — the Voronoi infill articulates Zone 1 and Zone 2 inside the orthogonal container
Two zones with merging schedules — the Voronoi infill articulates Zone 1 and Zone 2 inside the orthogonal container
Process of panel planarisation — extruded edges are planarised (kangaroo / ngon planarization), then pairs are projected to create flat surfaces that maintain flat edges
Process of panel planarisation — extruded edges are planarised (kangaroo / ngon planarization), then pairs are projected to create flat surfaces that maintain flat edges
Joint creation — parameter studies (offset and tolerance) comparing connected vs disconnected plates across the panel joints
Joint creation — parameter studies (offset and tolerance) comparing connected vs disconnected plates across the panel joints
Light-to-panel ratio — density of light panels is greatest on the ceiling and decreases toward the lower levels, mimicking natural light distribution. Below: correlation between panel aspect ratio and structural efficiency
Light-to-panel ratio — density of light panels is greatest on the ceiling and decreases toward the lower levels, mimicking natural light distribution. Below: correlation between panel aspect ratio and structural efficiency
Acoustic performance — absorption coefficient α vs frequency for three Voronoi-cell region sizes: small (d ≈ 15 mm, high freq.), medium (d ≈ 35 mm, mid), large (d ≈ 60 mm, low)
Acoustic performance — absorption coefficient α vs frequency for three Voronoi-cell region sizes: small (d ≈ 15 mm, high freq.), medium (d ≈ 35 mm, mid), large (d ≈ 60 mm, low)
Plate design — exploded view of the prototype element: point light with sensor, LED strip, finishing plate, and the 3D-printed acoustic infill combined with structural corner optimisation
Plate design — exploded view of the prototype element: point light with sensor, LED strip, finishing plate, and the 3D-printed acoustic infill combined with structural corner optimisation
Moving through scales — structural patterns. Whole panel: a dense Voronoi pattern follows stress lines; overlapping lines may lose directional stiffness. Panel with cut-out: a looser pattern with reduced infill material
Moving through scales — structural patterns. Whole panel: a dense Voronoi pattern follows stress lines; overlapping lines may lose directional stiffness. Panel with cut-out: a looser pattern with reduced infill material
Stress-line analysis — a central hole in the panel pushes stress lines toward the edges, stiffening the joint with neighbouring panels and reducing the infill material needed
Stress-line analysis — a central hole in the panel pushes stress lines toward the edges, stiffening the joint with neighbouring panels and reducing the infill material needed
Prototype description — annotated detail of the ceiling element: finishing plate, pocket for the ring light, Voronoi outer finish structurally optimised, back-element finish
Prototype description — annotated detail of the ceiling element: finishing plate, pocket for the ring light, Voronoi outer finish structurally optimised, back-element finish
Geometry evolution — General Geometry → Structure + Light Points → Version 3.0: a seamless fusion of form and texture where Voronoi cells double as acoustic cushions, providing visual depth and enhanced auditory comfort
Geometry evolution — General Geometry → Structure + Light Points → Version 3.0: a seamless fusion of form and texture where Voronoi cells double as acoustic cushions, providing visual depth and enhanced auditory comfort
Artificial neural network — feed-forward network with input, hidden and output layers; the dataset is split 70 / 30 into training and testing sets, drawing on weather features (outdoor temperature, direct / diffuse normal radiation, relative humidity, illumination, total sky cover) and physiological features (heart rate, pupil diameter, blink rate per minute, skin conductance, respiratory rate)
Artificial neural network — feed-forward network with input, hidden and output layers; the dataset is split 70 / 30 into training and testing sets, drawing on weather features (outdoor temperature, direct / diffuse normal radiation, relative humidity, illumination, total sky cover) and physiological features (heart rate, pupil diameter, blink rate per minute, skin conductance, respiratory rate)
AI: final input data — weather inputs (direct normal illumination, total sky cover, infrared radiation) and physiological inputs (heart rate BPM, skin conductance, respiratory rate) selected as the model's operating features
AI: final input data — weather inputs (direct normal illumination, total sky cover, infrared radiation) and physiological inputs (heart rate BPM, skin conductance, respiratory rate) selected as the model's operating features
AI: plotting input data against CCT — scatter plots of each input feature against correlated colour temperature, used to identify redundancies and isolate the variables most relevant for prediction
AI: plotting input data against CCT — scatter plots of each input feature against correlated colour temperature, used to identify redundancies and isolate the variables most relevant for prediction
AI: implementation in the project — Python pipeline: normalise data → train the model on existing data → save predicted illuminance and CCT to CSV for downstream Grasshopper and Arduino use
AI: implementation in the project — Python pipeline: normalise data → train the model on existing data → save predicted illuminance and CCT to CSV for downstream Grasshopper and Arduino use
Implementation of the data in Arduino code — Setting A (24 h simulation looping a hard-coded CSV of R, G, B and brightness values; full day compressed into ≈4 min 27 s) and Setting B (live input from the HTML interface via an ESP32 Wi-Fi server exposing HTTP endpoints; RGB is scaled by brightness before reaching the LED strip)
Implementation of the data in Arduino code — Setting A (24 h simulation looping a hard-coded CSV of R, G, B and brightness values; full day compressed into ≈4 min 27 s) and Setting B (live input from the HTML interface via an ESP32 Wi-Fi server exposing HTTP endpoints; RGB is scaled by brightness before reaching the LED strip)
Physical prototype — LED strip driven from a breadboard-mounted microcontroller during testing
Physical prototype — LED strip driven from a breadboard-mounted microcontroller during testing
The team with the lit 1:1 Voronoi panel prototype
The team with the lit 1:1 Voronoi panel prototype

Scenario — seven days indoors

The design targets the most severe winter interval: continuous darkness, restricted external mobility, and the longest gap between resupply missions. Within that frame, two occupants share a single container for seven days without leaving. The scenario is not a prediction — it's a stress test that exposes the spatial, environmental and psychological demands of confinement that a 'normal-operation' brief would hide.

From AI output to LED input

The trained model emits correlated colour temperature (kelvin) and illuminance (lux) — environmentally meaningful, but not directly consumable by an LED strip. A translation chain converts CCT → RGB and illuminance → brightness, bounds CCT to the 2700–6500 K window, and couples the two channels so colour temperature and intensity stay correlated. Across a 24-hour cycle the red channel stays structurally high; what changes is the balance of blue and green.

  • Setting A — scripted 24-hour cycle

    A precomputed CSV of R, G, B, brightness rows is hard-coded into the Arduino sketch; the loop steps through one row every 3 s, compressing a full day into roughly 4 min 27 s.

  • Setting B — live control

    The ESP32 exposes HTTP endpoints on a local Wi-Fi network; the HTML PhysioApp posts JSON (preset, activity, response mode), the sketch parses it, scales RGB by brightness, and updates the strip.

  • Mirror or compensate

    Presets (calm, focus, stress, overload) × activities (sleep, eat, leisure, work) feed a single luminous output. Mirror reflects the detected state; compensate counterbalances it — a corrective rather than mimetic response.

Limitations & ethics

The seven-day scenario is a stress test, not an in-situ validation; the predictive model runs on a curated dataset rather than live Antarctic data; the prototype demonstrates translation at panel scale, not the full habitat over long-duration occupation. More importantly, a system that watches the body to adjust the environment is not ethically neutral — even when it reduces manual control, it normalises continuous physiological monitoring inside a domestic setting. Consent, transparency, data residency, and a clear boundary between support and regulation are open questions the design doesn't yet answer.

Demo videos

Adaptive interior — Voronoi fit-out walkthrough
AI-driven circadian lighting — 24-hour cycle + live PhysioApp control

Team

  • Giorgia VercelloniGroup member
  • Maciej SachseGroup member
  • Floruț RuxandraGroup member
  • Zuzanna SchleiferGroup member
  • Brendan ExterkateGroup member
  • Gabriel MarksGroup member
  • Wong Long KiGroup member