CapillAir

Instrumenting passive cooling in a three-storey Italian house, to find out what actually moves heat out of it — and to drive a window open/close advisor from the answer.

Short version: the mechanical heat-recovery ventilation unit everyone assumes is doing the work contributes about 100 W. The unpowered chimney the house already had, and which nobody had measured, contributes about ten times that.

What was measured

The intake shaft is a thermal flywheel. The VMC draws from a shaded west cavedio at interrato level. Across a day when outdoor air swung 24.0 → 29.9 °C, the shaft moved 0.49 °C — 75% damping, and it sat 5.6 °C below ambient at the afternoon peak. It is a large shaded mass, not a tracker, which makes it an excellent intake precisely when it matters.

So the bypass stays open all day. A heat exchanger tempers incoming air toward indoor temperature, so when the intake is already colder than indoors, running it through the core destroys 75–80% of the benefit. The intuitive "close the bypass when it's hot outside" is backwards here, because the intake never sees outdoor air.

The passive stack dominates. cavedio → interrato → stairwell → piano primo → out runs on a 2.97 °C gradient, ≈0.67 Pa, ≈1160 m³/h, ≈1090 W — against the VMC's ≈100 W. The whole thing needs no fan; the only operational lever is whether a high exit is open.

But it is a channel, not a house-wide sweep. This is the correction that matters most. Over one full night the house mean fell 1.19 °C, and two thirds of that arrived only after a piano terra window was opened — a second inlet, not a bigger gradient. The control that proves it: taverna, same level as the interrato but with no opening of its own, lost 0.3 °C in eleven hours while cucina lost 2.0, and stayed flat within 0.2 °C across 18 h. Rooms hanging off the corridor get essentially nothing. The warmest rooms were the ones the stack missed.

The window rule

Three attempts; the first two failed in ways worth keeping:

1. Closing on the live coolest room does not work. Open windows couple rooms to outdoor air, so as outdoor climbs the rooms climb with it and the threshold runs away from the number chasing it. The call ran ~84 min late and gave back ~0.8 °C.
2. Latching the threshold alone reopens into the morning ramp, because the open side still keyed off live rooms, which warm too.
3. A 1 h trend guard is fooled by a plateau in that ramp.

Final rule, verified by replaying the real CSV before deploying:

• OPEN when outdoor < min(warmest room − 1.0, coolest room − 0.3) and falling ≥ 0.3 °C over a sustained 2 h. Over 2 h the morning ramp is unambiguously up and the evening fall unambiguously down.
• CLOSE when outdoor rises above the banked minimum — the lowest temperature actually reached this cycle, persisted to disk, so the target cannot drift upward.
• The interrato and cavedio never close. The cavedio is below the house all day.
• Rain gates the open transition only, and never forces a close: the best cooling gradient of the whole campaign (−6.6 °C) was measured during rain.

It fails closed: without trend history it stays shut. A missed open costs one evening; a wrong open imports heat into a house that spent all night cooling.

Layout

path	what
log_climate.py	polls Home Assistant every 5 min → data/climate.csv
bin/vmcwatch.py	the live advisor; notifies open/close through HA
analyse.py	rolling 24 h analysis (analyse.py 48 for 48 h)
esphome/	ESPHome nodes: ambient, cavedio/VMC intake, north-face outdoor, plus two SHT41 device targets (battery Seeed XIAO C3, USB SuperMini C3)
HANDOFF.md	running session log, including the errors
RECIPE.md	build notes for the sensor nodes

Analysis windows are rolling, not calendar-day: the night purge straddles midnight and a startswith(today) filter silently discarded half of it.

Two traps worth stealing

Reference frames. Three different "outdoor" numbers exist in the CSV and confusing them caused six separate errors: north_t is true measured outdoor, out_real_t is the cavedio despite the name, and out_t is the forecast (coarse, ran +1.4 °C high). Damping was reported as 82–93% for two sessions because it had been computed against the forecast rather than the real north node; it is 75%. Before quoting any "outdoor" figure, resolve the column back to the entity behind it.

Watts at altitude. Every figure here is corrected for 290 m — ×0.942 against the sea-level 0.335 constant.

Note on the published data

data/climate.csv carries the thermal and humidity channels and the window state. The occupancy covariates the study logged for confound-masking — motion, presence, garage door — were removed from this repository and from its history, along with the WiFi SSIDs, because they describe a particular house rather than the physics. Nothing in the analysis reads them. Sensor credentials live in esphome/secrets.yaml, which was never committed.