MICROCOSM — Technical and Educational Plan

Project: MICROCOSM — nine student-built instrumented ecosystems at Ioanid school, Bucharest
Funder: ARCUB (approved equipment list, Annex 2) · Technical partner: Alacrity
Document: v1.0, 2026-10-20 · Status: proposal for discussion with the school and reconciliation with procurement
Replaces: "MICROCOSM — Alacrity Education, Ecosystems & Approved Technical Equipment" (5 pages, undated)

How to read this document: Chapter 0 is the two-page summary. Chapter 1 explains what was wrong with the proposal and what changed. Chapters 2–4 fix the platform every ecosystem shares. Chapter 5 is the core: one detailed design per ecosystem, written so that a team of students with a teacher-coach can build it. Chapter 6 is the priced budget and the line-by-line reconciliation with the approved list. Chapter 7 is the schedule to 2026-12-18 and beyond, the people, the workshops and the risks. Chapter 8 answers the proposal's open questions. Diagrams are in Mermaid and render on GitHub, GitLab, Obsidian, VS Code and most Markdown viewers.

0. Executive summary

Nine living, instrumented ecosystems on one shared platform (an ESP32 node per ecosystem running ESPHome, a Raspberry Pi 5 hub running Home Assistant on the school's existing Wi-Fi), each with a written science track (question, hypotheses, variables, controls, replication, analysis) and an engineering track (requirements, prototype, test, iterate), at least two weeks of data, one chart per team at the showcase, and a Time Capsule designed to run for ten years without depending on its electronics.

Safety is designed in. No student or teacher ever builds, wires or switches 230 V. Mains exists only inside certified products on a shelf, on RCD-protected strips, switched only through smart plugs. Everything in water, substrate or inside an enclosure is 12 V DC or less. Safety interlocks (heater cut-offs, pump limits, dry-run guards) live in the node firmware and in hardware thermal fuses, not in the hub.

Budget. The recommended plan costs about 20 989 RON (4 115 EUR) for equipment, enclosures, hub, network and biology, or 24 137 RON with a 15 % contingency. Of this, about 7 263 RON is biology, substrates and consumables that the proposal did not budget at all, and about 961 RON is the hub and network it did not list. The approved list, priced as written, would have cost about 8 500 RON and would not have produced a working project.

Schedule. Orders go out 2026-10-21 to 10-23. Hub and standard node are built during the autumn break. Teams assemble nodes 2026-11-10, enclosures 11-09 to 11-20, biology is established by 11-20 to 12-04, treatments run 12-01 to 12-18, Time Capsule is built 12-01 to 12-11, showcase 12-15 to 12-17, ARCUB report 12-18. The interesting biology (decomposition, blooms, restoration, harvests) is expected January to June 2027; the plan schedules those rounds too.

1. Review of the current proposal

The proposal ("MICROCOSM — Alacrity Education, Ecosystems & Approved Technical Equipment", 5 pages) has a good skeleton: nine distinct ecosystems, a clear science/engineering split per ecosystem, a paired 2026/2056 experiment, a ten-year Time Capsule, and the right priority (educational outcomes before scientific results). Everything below the skeleton is missing or inconsistent. The equipment list reads as a generic "IoT terrarium" shopping list that was never checked against the nine ecosystems it is supposed to serve.

1.1 What is sound and is kept

1.2 Structural problems

1.3 Component-level problems

1.4 Net effect of the fixes

The nine ecosystems keep their identity. The equipment list changes in three directions: fewer of the wrong sensors (BME680, SD modules), more of the cheap right ones (soil probes, DS18B20, floats, LED strip), and the addition of everything the proposal forgot (hub, network, power, protection, heaters, air pumps, enclosures as COTS tanks, and the entire biology budget). The full reconciliation, line by line against the approved list, is in Chapter 6.

2. Design principles and hard rules

Every decision in this plan follows from seven rules. Where a later chapter deviates, it says so and why.

2.1 Hard safety rules (non-negotiable)

1. Mains (230 V AC) is never built, wired, modified or switched by students or teachers. All 230 V exists only inside certified COTS products with their own cord and plug: 12 V power bricks, the Raspberry Pi power supply, the hydroponics LED panel, the aquarium air pumps, the mister adapters.
2. Nothing mains-powered is ever inside an enclosure, in water, or in substrate. Everything that touches water, substrate or the inside of an enclosure is SELV: sensors at 3.3 V or 5 V; pumps, fans, LED strips and heat pads at 12 V DC; misters at 24 V DC from their own adapter.
3. Mains devices live on the shelf or wall, above and away from water, with a drip loop on every cable, on an RCD-protected (30 mA) power strip. Mains is switched only through certified COTS smart plugs, never through relay modules.
4. Node boxes mount outside the enclosure, above the water line. Cables enter through glands. Every 12 V brick is a certified adapter.
5. Simulated contamination only. Salinity is table salt, pH is vinegar or baking soda, disturbance is compaction, paving and heat. No heavy metals, no real pollutants, no vertebrates, no toxic plants. Gloves for soil and compost work, FFP2 masks when sieving mineral dust.

2.2 Engineering principles

6. COTS first. Complete, cased, cabled modules over loose parts: waterproof DS18B20, cased SHT31, Gravity-style sensors with cables, pre-made MOSFET and buck modules, ESP32 screw-terminal breakouts, pre-crimped cables, COTS glass aquariums, COTS pumps, COTS LED strips and profiles, COTS smart plugs, COTS junction boxes. Breadboards are for the 3.3 V / 5 V sensor side during prototyping only. Final assemblies use screw terminals or JST connectors, with no loose jumper wires.
7. Safety logic lives in the node, not in the hub. Heat-pad cut-offs, pump run-time limits and reservoir-low interlocks are implemented in ESPHome firmware on the ESP32 and backed by hardware thermal fuses. Home Assistant adds schedules, dashboards and alerts on top, so a hub outage never creates a hazard.
8. One platform, nine configurations. Every node is the same box, same pin map, same base firmware. A team configures, it does not re-invent. Spare nodes are interchangeable with any ecosystem.
9. Education before results. A well-documented failed experiment is a full success. Each ecosystem has a Science track (question, hypothesis, variables, control, replication, analysis) and an Engineering track (requirements, prototype, test, iterate). Adults set up the mains side and cut glass; students do everything else.
10. Design for December first, for spring second, for 2036 last. By 2026-12-18 all nine ecosystems must be alive, instrumented and reporting with at least two weeks of data. The interesting biology (decomposition, blooms, restoration, harvests) is expected in January–June 2027. The Time Capsule is designed so that the living system never depends on the electronics.

2.3 Fixed platform decisions

2.4 Standard pin map

Every node uses the same map, so any node YAML differs only in which packages are enabled and which setpoints apply.

3. Common platform

Architecture overview

Nine identical ESP32 nodes ("MICROCOSM node v1") talk over the project's own 2.4 GHz Wi-Fi to one Raspberry Pi 5 running Home Assistant OS (HAOS). The Pi stores the data, serves the dashboards and builds/flashes the firmware. The school network is optional: it gives Internet for NTP, remote access and off-site backup, and the lab keeps working without it.

flowchart LR
  subgraph NODES["9 ESP32 nodes - ESPHome - Wi-Fi 2.4 GHz"]
    N1["mc-01-rot"]
    N2["mc-02-ecocolumn"]
    N3["mc-03-brownfield"]
    N4["mc-04-future2026"]
    N5["mc-05-future2056"]
    N6["mc-06-regolith"]
    N7["mc-07-hydro"]
    N8["mc-08-soil"]
    N9["mc-09-capsule"]
  end
  SP["Smart plugs - hydro LED panel, air pump"]
  R["School Wi-Fi - 2.4 GHz WPA2-PSK SSID for the project"]
  subgraph PI["Raspberry Pi 5 - Home Assistant OS - 192.168.50.10"]
    HA["HA Core - recorder SQLite 30 d"]
    EH["ESPHome Device Builder"]
    MQ["Mosquitto MQTT"]
    IDB["InfluxDB 1.8 - keep forever"]
    GF["Grafana"]
  end
  C["Student laptops and phones on MICROCOSM Wi-Fi"]
  W["School network and Internet"]
  X["NTP - Tailscale or Nabu Casa - Google Drive backup"]
  N1 & N2 & N3 & N4 & N5 & N6 & N7 & N8 & N9 -->|"native API"| R
  SP -->|"API or MQTT"| R
  R ---|"Ethernet"| PI
  HA --> IDB
  IDB --> GF
  EH -->|"OTA firmware"| R
  MQ -.->|"advanced track"| R
  C -->|"HA and Grafana dashboards"| R
  R -->|"WAN optional"| W
  W --> X

One measurement, end to end:

sequenceDiagram
  participant S as Sensor (SHT31)
  participant E as ESPHome node
  participant H as HA Core
  participant R as Recorder (SQLite)
  participant I as InfluxDB
  participant G as Grafana or CSV export
  E->>S: I2C read every 60 s
  S-->>E: raw temperature and RH
  E->>E: filters - average, calibrate_linear, range check
  E->>H: native API push (Noise-encrypted protobuf)
  H->>R: state row, purged after 30 days
  H->>I: point - measurement = unit, tag entity_id, field value
  G->>I: InfluxQL query per node
  I-->>G: time series to panel, CSV download or nightly file

Why the ESPHome native API for the base, not raw MQTT

Mosquitto stays installed for three uses: (1) the advanced track (Arduino/PlatformIO on the spare boards with PubSubClient, topics microcosm/<node>/<quantity>, HA MQTT discovery as a learning task); (2) Tasmota-flashed smart plugs, if those are bought instead of ESPHome-flashable ones; (3) third-party devices. ESPHome nodes may also add an mqtt: block as an Engineering-track exercise. The base firmware does not depend on it.

The standard node ("MICROCOSM node v1")

Specification

Power budget (worst case per node)

Rules: the 5 V rail via the LM2596 stays ≤ 2 A (no heatsink), so at most 2 servos per node. The sum of 12 V loads that can be on at the same time is ≤ 4 A (80 % of the 5 A brick). If a node exceeds this (e.g. a 25 W pad plus 2 m of LED), split it: the LED strip gets its own PSU12, with grounds joined at the box. Misters run from their own 24 V adapter through the relay. Wire: 0.75 mm² for the 12 V input, 0.5 mm² for 12 V loads ≤ 2 A, 0.25 mm² (stock sensor leads) for signals.

Block / wiring diagram

flowchart TB
  PSU["PSU12 brick 12 V 5 A - 230 V side on the RCD shelf"] -->|"P1 DC jack 5.5/2.1, centre +"| F
  subgraph BOX["IP65 node box - outside the enclosure, above water line, glands down"]
    F["Inline fuse 5 A"] --> B12["12 V terminal bus + GND bus"]
    B12 --> BK["LM2596 set to 5.1 V"]
    BK -->|"5 V to VIN"| ESP["ESP32 DevKitC on TERMBOARD"]
    ESP -->|"3.3 V"| BUS["I2C GPIO21/22 and 1-Wire GPIO4 with 4.7k"]
    OLED["OLED 0x3C behind lid window"]
    DIV["Dividers - 1k/2k ECHO, 10k/20k flow"]
    M1["MOSFET 1 - GPIO16"]
    M2["MOSFET 2 - GPIO17"]
    M3["MOSFET 3 - GPIO18"]
    M4["MOSFET 4 - GPIO19"]
    RL["Relay 1/2 - GPIO23/25"]
    B12 --> M1
    B12 --> M2
    B12 --> M3
    B12 --> M4
    ESP --> M1
    ESP --> M2
    ESP --> M3
    ESP --> M4
    ESP --> RL
    ESP --> DIV
  end
  BUS --> OLED
  BUS -->|"S1 JST-XH 4p"| I2CS["SHT31 / SCD41 / BH1750 / BME680"]
  BUS -->|"S2 JST-XH 3p"| DS["DS18B20 probes"]
  ESP -->|"S3-S6 JST-XH 3p, GPIO32-35"| SOIL["Soil probes A-D"]
  ESP -->|"S7 JST-XH 3p, GPIO36/39"| AN["pH / TDS boards"]
  ESP -->|"S8 JST-XH 2p, GPIO13/14"| FL["Float switches"]
  DIV -->|"S9 JST-XH 4p"| JSN["JSN-SR04T level"]
  DIV -->|"S10 JST-XH 3p"| FLOW["YF-S401 flow"]
  ESP -->|"S11/S12 JST-XH 3p, GPIO26/27"| SV["SG90 servos"]
  M1 -->|"A1 screw terminal"| LED["12 V LED strip in profile"]
  M2 -->|"A2"| FAN["12 V fan"]
  M3 -->|"A3"| PUMP["12 V pump"]
  M4 -->|"A4, via 72 C thermal fuse"| HEAT["12 V heat pad"]
  RL -->|"A5"| MIST["24 V mister - own adapter"]

Connectors, colours, labels

Wire colours (all nodes, all harnesses): red = 12 V, black = GND, orange = 5 V, yellow = 3.3 V, blue = signal (second signal in the same cable: green, e.g. SCL). Loads switched by a MOSFET: red to +12 V, black to LOAD−; the load's black wire is never connected to the GND bus directly.

Labels: printed heat-shrink or flag labels at both ends of every cable, format <node#>-<port> <device> <position>, e.g. 04-S3 SOIL A, 04-A3 PUMP. The box lid carries a port map sticker (copy of the table above with the node's actual devices) and the node name + IP.

BOM for one node (Future 2026 example, 3 PWM loads)

Assembly session (one team, 90 min, checklist)

Prerequisites (done by Alacrity/teacher beforehand): kits bagged per node; glands holes pre-marked from a paper template; firmware compiled and flashed once via USB to each ESP32 (so on the day the board only needs OTA); multimeter per team.

QA checklist (before the ESP32 goes in, then with it)

Hard safety rules restated for the node

1. The node and everything connected to it is SELV ≤ 12 V DC (24 V for the mister loop only). No mains conductor ever enters a node box.
2. 230 V exists only inside certified bricks/adapters, the Pi PSU, the COTS LED panel and the air pump, all plugged into the RCD-protected strip on the shelf. Mains loads are switched only by certified smart plugs, never by a relay module.
3. The box hangs outside the enclosure, above the water line, glands down, with a drip loop on every cable.
4. The heat pad is wired through MOSFET 4 and a 72 °C thermal fuse taped to the pad, and its firmware interlock is enabled (see Firmware).
5. Unplug the brick before opening the box or (dis)connecting anything.

Firmware (ESPHome)

File layout (in /config/esphome/ on the Pi, edited from the ESPHome Device Builder or Studio Code Server)

/config/esphome/
  secrets.yaml                 # Wi-Fi, OTA, one API key per node; admin-only
  mc-01-rot.yaml ... mc-09-capsule.yaml   # one ~25-line file per node, owned by the team
  common/
    mc-base.yaml               # Wi-Fi, API, OTA, time, I2C, 1-Wire, OLED, diagnostics
    sens-sht31.yaml  sens-bme680.yaml  sens-scd41.yaml  sens-bh1750.yaml
    sens-ds18b20.yaml          # vars: sid, label, addr, offset   (one include per probe)
    sens-soil.yaml             # vars: pos, pin, dry_v, wet_v     (one include per probe)
    sens-analog.yaml           # vars: aid, pin, label, unit, v1, y1, v2, y2 (pH, TDS)
    act-light.yaml             # PWM1 GPIO16 + local photoperiod
    act-fan.yaml               # PWM2 GPIO17
    act-pump.yaml              # vars: pid, pin, label, max_s  (PWM3 GPIO18 by default)
    act-heater.yaml            # PWM4 GPIO19 + thermostat + latched interlock
    act-relay.yaml             # vars: rid, pin, label, max_s  (GPIO23/25)
    act-servo.yaml             # vars: sid, pin, label         (GPIO26/27)
    in-floats.yaml  guard-dryrun.yaml  in-flow.yaml  in-level.yaml

The common/ files are owned by Alacrity and changed only by pull request/teacher. Teams edit only their node file. Every file below passed esphome config, and both the example node and a test node with every package enabled compiled cleanly (ESPHome 2026.9.1, ESP-IDF 5.5.5). Node-level substitutions override the defaults in the packages.

ESPHome components used: sht3xd, scd4x, bh1750, bme680, one_wire + dallas_temp, adc, copy, ledc, monochromatic light, speed fan, output/template switch, servo, pulse_counter, ultrasonic, gpio/template/status binary_sensor, ssd1306_i2c + font, sntp + homeassistant time, wifi_signal, uptime, script, interval, globals, template number/button.

secrets.yaml (template; the real file stays on the Pi)

wifi_ssid: "MICROCOSM"
wifi_password: "<WPA2 key, 16+ chars>"
fallback_ap_password: "<12+ chars>"
ota_password: "<random 20 chars>"
mc00_api_key: "<32-byte base64>"   # generate: openssl rand -base64 32
mc01_api_key: "<...>"
mc04_api_key: "<...>"
# ... one per node incl. spares mc-10..mc-12

Base firmware and packages

The full listing of common/mc-base.yaml, the sensor packages, the actuator packages and the heater safety interlock is in Appendix B. All files compiled cleanly with esphome config and esphome compile on ESPHome 2026.9.1 and are delivered alongside this document in the microcosm-firmware/ folder.

Example node: mc-04-future2026.yaml

# mc-04-future2026.yaml — Future 2026 (paired with mc-05). Only this file is edited by the team.
substitutions:
  node_name: mc-04-future2026
  friendly_name: "MC 04"
  area: "Future 2026"
  api_key: !secret mc04_api_key
  user1_src: soil_a
  user1_label: "Soil A pct"
  user2_src: co2
  user2_label: "CO2 ppm"
  light_on_min: "420"              # 07:00 (identical in mc-05!)
  light_off_min: "1260"            # 21:00

packages:
  base: !include common/mc-base.yaml
  sht31: !include common/sens-sht31.yaml
  scd41: !include common/sens-scd41.yaml
  bh1750: !include common/sens-bh1750.yaml
  t_sub_a: !include
    file: common/sens-ds18b20.yaml
    vars: {sid: t_sub_a, label: "Substrate temp A", addr: "0x5a3c01f0956e2b28", offset: "0.0"}
  soil_a: !include
    file: common/sens-soil.yaml
    vars: {pos: a, pin: GPIO32, dry_v: "2.62", wet_v: "1.31"}
  soil_b: !include
    file: common/sens-soil.yaml
    vars: {pos: b, pin: GPIO33, dry_v: "2.58", wet_v: "1.27"}
  light: !include common/act-light.yaml
  fan: !include common/act-fan.yaml
  pump: !include
    file: common/act-pump.yaml
    vars: {pid: pump, pin: GPIO18, label: "Irrigation pump", max_s: "60"}

The resulting HA entities are sensor.mc_04_air_temp, sensor.mc_04_air_rh, sensor.mc_04_co2, sensor.mc_04_illuminance, sensor.mc_04_substrate_temp_a, sensor.mc_04_soil_moisture_a/_b (+ diagnostic sensor.mc_04_soil_raw_a/_b), light.mc_04_grow_light, switch.mc_04_grow_light_schedule, fan.mc_04_fan, fan.mc_04_irrigation_pump, sensor.mc_04_wifi_rssi and binary_sensor.mc_04_status. HA builds the id from the device's friendly name MC 04 plus the entity name. Check the ids once after adoption: they are frozen at first registration, and renaming later breaks the InfluxDB series.

mc-05-future2056 uses the same file plus act-heater.yaml (heat_probe: t_sub_a) and act-relay.yaml (mister, max_s: "30"), with identical light_on_min/light_off_min, so the only differences between 04 and 05 are the experimental variables.

Firmware rules

Hub (Raspberry Pi 5 + Home Assistant OS)

Install (≈ 90 min, teacher + Alacrity)

Storage: HAOS boots from the 64 GB A2 microSD card. SD wear is the usual hub failure, so the nightly backup below is mandatory and a second card with a restored image is kept in the hub box. An NVMe drive was considered and dropped as a cost cut; it can be added later without reinstalling.

Add-ons and roles

configuration.yaml blocks

recorder:
  purge_keep_days: 30          # short-term SQLite for HA history graphs
  commit_interval: 5
  exclude:
    entity_globs:
      - sensor.*_uptime

influxdb:
  host: a0d7b954-influxdb      # InfluxDB add-on hostname
  port: 8086
  database: homeassistant
  username: homeassistant
  password: !secret influxdb_password
  max_retries: 3
  default_measurement: state   # unit-less entities (switches, binary sensors)
  tags:
    project: microcosm
  include:
    entity_globs:
      - sensor.mc_*
      - binary_sensor.mc_*
      - switch.mc_*
      - light.mc_*
      - fan.mc_*
      - number.mc_*
      - input_boolean.mc_*     # mc_XX_maintenance flags -> quality = MAINT
  exclude:
    entity_globs:
      - sensor.mc_*_uptime

input_boolean:                 # one per node; set before hands-on work
  mc_04_maintenance:
    name: "MC 04 maintenance"
    icon: mdi:wrench
  # ... mc_01 ... mc_09

shell_command:
  mc_export_csv: "python3 /config/scripts/mc_export.py"

automation:
  - id: mc_nightly_csv
    alias: "MC nightly CSV export 02:15"
    triggers:
      - trigger: time
        at: "02:15:00"
    actions:
      - action: shell_command.mc_export_csv

InfluxDB writes one measurement per unit (°C, %, ppm, lx, L/min, m, state, …), tag entity_id (without the domain), field value.

Nightly CSV export: /config/scripts/mc_export.py

Writes yesterday (UTC) as microcosm_YYYY-MM-DD.csv to the USB stick (label it MCBACKUP; HAOS mounts it under /media/MCBACKUP) and to /share/microcosm/csv (Samba share, rclone). The InfluxDB password goes in /config/.influx_pass. Re-export any day with python3 /config/scripts/mc_export.py 2026-11-03 from the Studio Code Server terminal. Tested against a mock InfluxDB response.

#!/usr/bin/env python3
"""MICROCOSM nightly export: one UTC day from InfluxDB 1.8 -> long-format CSV.
Usage: python3 /config/scripts/mc_export.py [YYYY-MM-DD]   (default: yesterday UTC)
Columns: timestamp_utc,node,entity,value,unit,quality
"""
import base64, csv, datetime as dt, json, os, re, sys, urllib.parse, urllib.request

INFLUX = "http://a0d7b954-influxdb:8086/query"
DB, USER = "homeassistant", "homeassistant"
PASS = open("/config/.influx_pass").read().strip()
OUTDIRS = [("/media/MCBACKUP", "microcosm/csv"), ("/share", "microcosm/csv")]  # (mount, subdir)
# physical plausibility per unit -> quality RANGE outside it
LIMITS = {"°C": (-10, 60), "%": (0, 100), "ppm": (300, 10000), "lx": (0, 70000),
          "L/min": (0, 10), "m": (-0.05, 1.0), "pH": (2, 12), "µS/cm": (0, 5000)}

day = dt.date.fromisoformat(sys.argv[1]) if len(sys.argv) > 1 else \
    dt.datetime.now(dt.timezone.utc).date() - dt.timedelta(days=1)
t0, t1 = f"{day}T00:00:00Z", f"{day + dt.timedelta(days=1)}T00:00:00Z"

def query(q):
    url = INFLUX + "?" + urllib.parse.urlencode({"db": DB, "q": q, "epoch": "s"})
    req = urllib.request.Request(url)
    tok = base64.b64encode(f"{USER}:{PASS}".encode()).decode()
    req.add_header("Authorization", "Basic " + tok)
    with urllib.request.urlopen(req, timeout=120) as r:
        return json.load(r)["results"][0].get("series", [])

series = query(f"SELECT value FROM /.*/ WHERE entity_id =~ /^mc_/ "
               f"AND time >= '{t0}' AND time < '{t1}' GROUP BY entity_id")

def node_of(ent):
    return re.match(r"mc_\d+b?", ent).group(0)   # mc_04, mc_08b

# maintenance windows: input_boolean.mc_XX_maintenance (1 = on)
maint = {}
for s in series:
    e = s["tags"]["entity_id"]
    if e.endswith("_maintenance"):
        maint[node_of(e)] = sorted((int(t), v) for t, v in s["values"])

def in_maint(node, t):
    state = 0
    for ts, v in maint.get(node, []):
        if ts > t:
            break
        state = v
    return state == 1

rows = []
for s in series:
    unit, ent = s["name"], s["tags"]["entity_id"]
    node = node_of(ent)
    for t, v in s["values"]:
        if v is None:
            continue
        lo, hi = LIMITS.get(unit, (float("-inf"), float("inf")))
        q = "MAINT" if in_maint(node, t) else ("RANGE" if not lo <= v <= hi else "OK")
        ts = dt.datetime.fromtimestamp(t, dt.timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
        rows.append((ts, node, ent, v, "" if unit == "state" else unit, q))
rows.sort()

for root, sub in OUTDIRS:
    if not os.path.isdir(root):
        continue                          # USB stick not mounted -> skip
    d = os.path.join(root, sub)
    os.makedirs(d, exist_ok=True)
    with open(os.path.join(d, f"microcosm_{day}.csv"), "w", newline="", encoding="utf-8") as f:
        w = csv.writer(f)
        w.writerow(["timestamp_utc", "node", "entity", "value", "unit", "quality"])
        w.writerows(rows)
print(f"{day}: {len(rows)} rows")

Grafana (one template dashboard for all nine)

• Data source: InfluxDB, URL http://a0d7b954-influxdb:8086, database homeassistant, user grafana, InfluxQL.
• Dashboard MICROCOSM node, variable node (Custom: mc_01,mc_02,mc_03,mc_04,mc_05,mc_06,mc_07,mc_08,mc_09), time range default 7 days.
• Panels (one query each, legend $tag_entity_id):

• An extra dashboard Future 2026 vs 2056 overlays mc_04_* and mc_05_* for the same quantity (one panel per quantity, two queries, plus a difference panel using a Grafana Transform → Add field from calculation → Binary operation).

CSV export methods for students

Naming conventions

Home Assistant dashboards

Experiment-mode buttons are HA scripts. The node firmware still enforces the limits (pump max-run, heater interlock):

script:
  mc_05_rain_event:
    alias: "MC 05 rain event (2 min at 60 %)"
    sequence:
      - action: fan.turn_on
        target:
          entity_id: fan.mc_05_rain_pump
        data:
          percentage: 60
      - delay: "00:02:00"
      - action: fan.turn_off
        target:
          entity_id: fan.mc_05_rain_pump
  mc_05_heatwave_on:
    alias: "MC 05 heatwave (+5 C setpoint)"
    sequence:
      - action: number.set_value
        target:
          entity_id: number.mc_05_heater_setpoint
        data:
          value: "{{ states('number.mc_05_heater_setpoint') | float(25) + 5 }}"
      - action: switch.turn_on
        target:
          entity_id: switch.mc_05_heater_enable

Add a logbook.log step (name, message) to every experiment-mode script so interventions appear in the event log that teams export with their data.

Alerting (Telegram group "MICROCOSM staff"; e-mail via the SMTP integration as fallback)

Set up Settings → Integrations → Telegram bot (polling), with the bot added to the staff group. Alerts go to staff, not to students' private phones.

automation:
  - id: mc_node_offline
    alias: "MC node offline > 15 min"
    triggers:
      - trigger: state
        entity_id:
          - binary_sensor.mc_01_status
          - binary_sensor.mc_02_status
          - binary_sensor.mc_03_status
          - binary_sensor.mc_04_status
          - binary_sensor.mc_05_status
          - binary_sensor.mc_06_status
          - binary_sensor.mc_07_status
          - binary_sensor.mc_08_status
          - binary_sensor.mc_09_status
        to:
          - "off"
          - "unavailable"
        for: "00:15:00"
    actions:
      - action: telegram_bot.send_message
        data:
          message: "MICROCOSM: {{ trigger.to_state.name }} offline for 15 min"

  - id: mc_sensor_nan
    alias: "MC sensor unknown > 10 min"
    triggers:
      - trigger: template
        value_template: >
          {{ states.sensor | selectattr('entity_id', 'match', 'sensor\.mc_')
             | selectattr('state', 'in', ['unknown', 'unavailable']) | list | count > 0 }}
        for: "00:10:00"
    actions:
      - action: telegram_bot.send_message
        data:
          message: >
            MICROCOSM sensors without data:
            {{ states.sensor | selectattr('entity_id', 'match', 'sensor\.mc_')
               | selectattr('state', 'in', ['unknown', 'unavailable'])
               | map(attribute='entity_id') | join(', ') }}

  - id: mc_thresholds
    alias: "MC threshold / interlock"
    triggers:
      - trigger: state
        entity_id:
          - binary_sensor.mc_03_heater_interlock
          - binary_sensor.mc_05_heater_interlock
          - binary_sensor.mc_07_level_low_alarm
        to: "on"
      - trigger: numeric_state
        entity_id:
          - sensor.mc_04_air_temp
          - sensor.mc_05_air_temp
        above: 38
        for: "00:10:00"
      - trigger: numeric_state
        entity_id: sensor.mc_07_water_temp
        above: 28
        for: "00:30:00"
    actions:
      - action: telegram_bot.send_message
        data:
          message: "MICROCOSM alert: {{ trigger.to_state.name }} = {{ trigger.to_state.state }}"

User accounts

HA has no per-entity permissions, so a team account can technically switch another team's pump. This is acceptable because every safety limit is in node firmware; misuse is a classroom-rule matter. Add-on panels (ESPHome, InfluxDB, terminals) are admin-only.

Data pipeline, retention and backup

flowchart LR
  N["9 nodes - 60 s values"] -->|"native API"| HA["HA Core"]
  HA --> REC["Recorder SQLite - 30 days"]
  HA --> IDB["InfluxDB - autogen, keep forever"]
  IDB --> GF["Grafana dashboards"]
  IDB --> EXP["02:15 mc_export.py - yesterday UTC"]
  EXP --> USB["USB stick MCBACKUP - CSV per day"]
  EXP --> SH["/share/microcosm/csv - Samba"]
  SH --> GD["Google Drive folder - rclone or Drive for Desktop"]
  HA --> BK["03:00 HA full backup, encrypted - incl. InfluxDB add-on data"]
  BK --> GDB["Google Drive backups - keep 14"]
  BK --> LOC["Local copy on the Pi"]

Restore test: once per term, restore the latest backup onto the spare SD64 in a second Pi (or the same Pi with a swapped card), check that dashboards and the last 24 h of InfluxDB data are present, then swap back.

Hub BOM

Network

The nodes use the school's existing Wi-Fi. No router is bought. What the project needs from school IT, in writing, by 2026-10-23:

Setup checklist (30 min): put the SSID and password in secrets.yaml; set manual_ip in each node YAML from the reservation list; check RSSI on every node after placement (acceptance: better than −75 dBm; move the node box or add a 2 m antenna extension if not); label each node box with its name and address.

What happens offline: nodes keep running their own automations and safety logic (heater interlock, pump limits, dry-run guard, light schedule all live in the firmware). Home Assistant schedules and alerts stop until the network returns. Design rule: safety logic is never in the hub. ESPHome's default reboot-on-no-connection is disabled so that a long outage does not restart the nodes.

Fallback: if IT cannot provide the SSID by 2026-10-27, buy a cheap 2.4 GHz router (≈ 160 RON) the same day and run an isolated "MICROCOSM" network with the Pi on its LAN port; the rest of this chapter is unchanged. Remote access, if wanted, is Tailscale (free) on the Pi.

Security basics: the Wi-Fi password may be given to students; the Home Assistant admin password and the ESPHome API keys are not. Students have non-admin Home Assistant accounts. Passwords are rotated at the start of each school year.

Calibration & metrology

General policy

M1. Capacitive soil probes (per probe, per substrate)

1. Seal first: conformal coating or clear nail varnish over the PCB edge cut and components, heat-shrink over the electronics, mark the max insertion line.
2. Flash the node; watch sensor.mc_XX_soil_raw_<pos> (V, 60 s average).
3. Air: probe dry, held in air, 3 min → record V_air (sanity check only, ≈ 2.6–2.8 V at 3.3 V supply).
4. Dry substrate: 500 mL of the actual substrate, air-dried 3 days on a tray (or 105 °C for 24 h in an oven where available), lightly packed to the planned density in a beaker. Insert to the line, 3 min → V_dry.
5. Saturated substrate: add water until free water appears at the surface, wait 30 min, re-insert at the same depth, 3 min → V_wet.
6. Enter dry_v/wet_v in the node YAML (sens-soil.yaml include) → OTA.
7. Optional (Science track, gravimetric curve): from the dry sample, add water in 4 steps (e.g. 10/20/30/40 % of dry mass), mix, record V and the gravimetric water content (g water/g dry substrate). Fit a curve; this turns "% of saturation" into a real water content.

Output meaning: soil_moisture = relative saturation, 0 % = air-dry, 100 % = saturated, for that substrate. It is not volumetric water content and cannot be compared across substrates unless the gravimetric curve is done.

Record sheet M1

M2. SHT31 cross-check: saturated salt (NaCl) at 75.3 % RH

1. Small cup with ≈ 30 g table salt (NaCl, no iodine/anti-caking is best) + water dropwise until it is a wet slurry, like wet sand, with undissolved crystals always present.
2. Cup + sensor (on its cable through the lid, sealed with Blu-Tack) in a 0.5–1 L airtight glass jar. The sensor must not touch the slurry.
3. Stable room temperature (± 1 °C), out of sunlight, 12–24 h (overnight).
4. Read air_rh: expected 75.3 % at 20–25 °C. SHT31 spec ±2 % RH → accept 73–78 %. Offset > 3 % → replace or note an offset filter. Also note T vs the reference thermometer.
5. Optional second point: potassium carbonate K₂CO₃ slurry = 43 % RH.

Record sheet M2

M3. DS18B20 ice-water 0 °C check

1. 1 L insulated cup packed full of crushed ice, filled with cold tap water to just below the ice top, stirred, 5 min.
2. All probes in together (tips mid-depth, not touching the walls), plus the reference thermometer. 5 min, stirring every minute.
3. Expected 0.0 °C; spec ±0.5 °C → pass. Write the offset -reading into the probe's offset var only if |error| > 0.2 °C.
4. Second point: all probes + reference in a stirred room-temperature water bucket, to cross-check that probes agree within 0.3 °C.

Record sheet M3

M4. SCD41 CO₂: forced recalibration (FRC), and why ASC is off

• ASC (automatic self-calibration) assumes that the sensor sees fresh air (≈ 420 ppm) at least once every 7 days and treats the lowest reading of the week as 400 ppm. In a closed tank with soil, compost or plants, CO₂ rarely drops to outdoor level. ASC then silently shifts the baseline, readings drift low over weeks, and respiration data become meaningless. ASC is disabled in firmware (automatic_self_calibration: false) on every node.
• FRC procedure: take the node (on a USB power bank via PSU5 cable, or the whole box on a long USB lead) outdoors to the school courtyard or a park, ≥ 10 m from traffic and people (your breath is ≈ 40 000 ppm), sensor in shade, 5 min stable readings (periodic mode). Then press button.mc_XX_co2_frc_425_ppm. Bucharest urban background is ≈ 425–450 ppm; we use 425 ppm and accept ±30 ppm residual uncertainty.
• Inter-sensor check (important for paired 04/05): put all four SCD41 nodes side by side in one closed box for 1 h. The four readings should agree within ±(40 ppm + 5 %). Record the offsets and repeat each term.

Record sheet M4

M5. pH probe: 2-point (7, then 4), slope check at 10

1. Rinse the probe in distilled water, blot (don't wipe). Buffers at room temperature, note T.
2. pH 7.00 buffer, stir gently, wait for stable V (2 min) → V7. Rinse.
3. pH 4.00 buffer → V4. Rinse. (pH 10 → V10 for a slope check only.)
4. Set v1: V7, y1: "7.0", v2: V4, y2: "4.0" in the sens-analog.yaml include. Slope should be ≈ 0.17–0.18 V/pH for a Gravity v2 board (≈ 59 mV/pH at the electrode, amplified).
5. Store the probe in KCl storage solution, never in distilled water. Recalibrate monthly. If pH and TDS probes sit in the same water, read them alternately or fit a galvanic isolator: they interfere.

M6. TDS/EC: 1413 µS/cm standard

1. 1413 µS/cm solution at 25 °C (or note T); rinse, immerse, 2 min.
2. The Gravity TDS conversion is EC = (133.42·V³ − 255.86·V² + 857.39·V) × K, with temperature compensation EC25 = EC / (1 + 0.02·(T − 25)) using the water DS18B20 (template sensor in the Hydro node). Adjust K until the reading is 1413 ± 2 %. TDS (ppm) = EC25 × 0.5.
3. Monthly; probe rinsed and dried after use when not permanently installed.

M7. BH1750 plausibility vs phone lux app

1. Same position, same angle: BH1750 vs phone app (front camera/ambient light sensor) under (a) room light, (b) the grow light at canopy height, (c) daylight at the window.
2. Phones are uncalibrated: accept ±30 %. Purpose: catch a dead, saturated (> 54 000 lx default range) or shaded sensor.
3. Limitation to teach: lux is weighted to the human eye, not PAR. For a white full-spectrum LED, PPFD (µmol·m⁻²·s⁻¹) ≈ lux / 70 is a rough guide only, and only valid for comparing the same light type.

Record sheets M5–M7 use the same layout: node, probe label, date, standard/condition, raw V or reading, set value, pass, by.

Dataset structure

Nightly files microcosm_YYYY-MM-DD.csv, long format, UTF-8, comma-separated:

Companion files in the same shared folder: sensors.csv (entity, probe label, ROM/serial, position, install date, calibration ref), calibration_log.csv (sheet rows M1–M7), events_log.csv (timestamp_utc, node, event, by: watering, feeding, salt added, experiment start/stop. HA logbook entries are exported here as well).

Data dictionary (base entities)

Commissioning & acceptance

Per-node checklist (sign-off by teacher + team)

Per-hub checklist

"Day-one data review" session (45 min, the day after each 24 h soak; repeat weekly in the first month)

Spare-parts policy

Safety

Lab wall sheet: "MICROCOSM electrical & water safety rules" (print A4, laminate)

1. Only 12 V (24 V for misters) inside or near the ecosystems. Nothing that plugs into a wall socket goes into an enclosure, into water or into soil.
2. 230 V devices are only certified, closed products (power bricks, Pi charger, LED panel, air pump, smart plugs). Never open, cut, extend or repair them. No relay modules on 230 V, ever.
3. All 230 V plugs live on the RCD-protected strip on the high shelf. Only the teacher plugs or unplugs at that strip.
4. Every cable has a drip loop (a U-shaped dip below the socket/box) so water runs off the bottom of the loop, not into the plug or box.
5. Before hands go into water or substrate: turn on maintenance mode in HA, unplug that node's 12 V DC jack at the box, switch the smart-plugged devices off at the plug, and hang the "OFF – working" tag on it.
6. Dry hands, dry bench. No drinks on the electronics bench. Wipe spills before touching anything electrical.
7. Node boxes stay closed, glands tight, above the water line. Open a box only after unplugging its brick.
8. Power off before connecting or disconnecting anything. Check polarity, and check 5 V before inserting an ESP32.
9. Heaters only through the node (interlock on, thermal fuse fitted). Never cover a heat pad or run one outside the enclosure.
10. Smell, heat, smoke, a tingle, a warm brick or a cracked cable: stop, step back, call the teacher. The teacher switches off the RCD strip and logs the incident.

RCD and power distribution

• One portable RCD adapter, 30 mA, type A (certified, 16 A) between the wall socket and each power strip, even if the room circuit has an RCD. Press TEST every Monday and tick the log.
• Max. two strips, each fed directly from a wall socket (no daisy-chaining), mounted on the shelf, switch facing the front. Estimated total load ≈ 300 W typical, < 700 W worst case (9 × 12 V bricks, 2 mister adapters, Pi, router, LED panel 45 W, air pump 5 W).

Mains device list and placement

Shelf: wall-mounted, top surface ≥ 40 cm above the tops of the tanks, front edge set back so that no open water is directly beneath it, bricks fixed with velcro or cable ties so they cannot fall. Cables run along the wall in a trunking/cable tray, never across walkways or over tank openings, and are fixed every 30 cm.

Who may plug and unplug what

Platform BOM

The platform parts (nodes, hub, power, spares, tools) are priced in Chapter 6 together with everything else; the quantities there are the ones to order. One node core (ESP32, breakout, box, glands, buck, OLED, cable kit) costs about 250 RON; the hub (Pi 5 kit, SD cards, backup stick) about 700 RON.

4. Sensor and actuator policy

The proposal asked: "useful vs unnecessary sensors". The answer depends on what each ecosystem is trying to measure, so this chapter allocates by need, not by headcount.

4.1 Useful, unnecessary, and wrong

4.2 Allocation matrix

The matrix below is the result of the nine ecosystem designs in Chapter 5 after the cost-cut pass of 2026-10-20. "Spare" is the project-wide spare pool.

Notes: LEDSTRIP is in 5 m reels (fractions are shares of a reel; about 14 m are actually cut, so 4 reels suffice); ACRYLIC in 50×100 cm panels (1 polycarbonate sheet gives 4, plus 2 acrylic panels); MESH in m²; TUBING in m; GLAND 80 covers 82 planned holes because several designs over-counted spares. The Carnivorous Bog alternative (Section 5.8b) is not counted; if chosen instead of Living Earth, its sub-BOM replaces that column. The Time Capsule's SCD41 is deferred to 2027 and not in the Buy column.

4.3 Automation policy

Automation is applied only where it serves a hypothesis or protects the system. Everything else stays manual and measured, because a student pouring 150 mL from a measuring cylinder and writing it down learns more than a pump that runs on a schedule.

4.4 Spare capacity

After the cost-cut pass the spare pool is deliberately thin. It covers the failures that are common and cheap; anything else is a same-week purchase from a Romanian shop.

5. Ecosystem designs

Each design follows the same template: concept, science track, engineering track, enclosure and build, biology, instrumentation, automation, operations, safety, educational outcomes, sub-BOM, timeline. Catalog IDs refer to the parts catalog in Chapter 6.

1. The Orchestra of Rot

Concept (refined)

A 45 L glass tank (TANK45) holds a living decomposer community: compost worms, two woodlouse species, springtails, oyster-mushroom mycelium and a compost/leaf-mould microbial inoculum. It sits on a moist, layered bed with a randomised grid of labelled litter bags buried in it. Each bag holds 3.00 g of one dry material (lime leaves, cardboard, cotton, bread, orange peel, or PE plastic as the control). Students measure how fast each material disappears, whether the animals speed it up (coarse- vs fine-mesh bags), and how the community "breathes" (CO2 production rate from a closed-lid test) and "smells" (BME680 gas resistance before and after an activated-carbon filter). Over the 8 weeks they will see bread and orange peel moulder and collapse, white mycelium spreading through cardboard, worms and isopods moving into coarse bags, and a measurable daily respiration curve. Leaves, cardboard and cotton are slow, so the clear differences between those materials and between mesh sizes only show up in the spring 2027 runs.

Node: mc-01-rot. Team: 3–4 students (one each for biology lead, data lead, engineering lead and logbook/QA).

Science track

Driving question: Which everyday materials does a decomposer community recycle fastest, which members of the "orchestra" do the work, and how do moisture and temperature change the tempo?

Hypotheses (testable)

• H1 – Material: mass-loss rate ranks bread > orange peel > lime leaves (Tilia) > cotton ≈ cardboard > PE plastic (≈ 0 %). Reason: more sugar/starch and less lignin means faster loss.
• H2 – Macrofauna access: for leaves, cardboard and cotton, coarse-mesh bags (5 mm, which worms and isopods can enter) lose ≥ 1.5× more mass than fine-mesh bags (≈ 1 mm, which only microbes, fungi, springtails and mites can enter) over the same period.
• H3 – Respiration & conditions: (a) adding 100 g of fresh food scraps raises the CO2 production rate by ≥ 50 % within 48 h and lowers headspace gas resistance (more VOCs), with both returning to baseline within 7 days; (b) the nightly CO2 production rate rises with core temperature, giving Q10 ≈ 2 over the natural Dec–Feb temperature swings.

Variables

Replication & controls (litter-bag method)

• Bags: 6×6 cm finished, sewn with nylon fishing line (no metal staples). Each bag is tethered by 20 cm of line to a numbered plastic tag at the surface. Material is dried at 60 °C for 24 h (kitchen oven, teacher-supervised) or air-dried on a radiator to constant mass (± 0.02 g on 2 consecutive days), then weighed to 0.01 g.
• Run 1 (16 Nov → 14 Dec 2026, 28 days), 27 bags: 6 materials × coarse mesh × 3 replicates = 18, plus 3 slow materials (leaves, cardboard, cotton) × fine mesh × 3 = 9. Positions are randomised on a 7 × 4 grid (row pitch 6.5 cm, column pitch 7 cm). Students roll dice or use random.org and record the map in the logbook.
• Handling-loss ("travel") bags: 1 per material (6). Each is buried and dug out immediately, then dried and reweighed. The mean loss is subtracted from every result in that material.
• PE plastic is the negative control. It is expected to show 0 ± 2 % loss, which measures the method's error (soil left stuck to the bag, weighing error).
• Window bags: 6 extra coarse bags (one per material) pressed against the front glass at 5 cm depth in a 3 cm strip in front of the grid. Students photograph them through the glass for VDI scoring without disturbing the bed. They are not weighed.
• Run 2 (15 Dec 2026 → 15 Mar 2027, 90 days): same 27-bag design in fresh bags, adding a 7th treatment: Pleurotus-pre-colonised cardboard (coarse, ×3). Run 3 (22 Mar → 7 Jun 2027): students design it (e.g. a wet vs dry half using the divider, or PLA "compostable" plastic vs PE).

Sampling schedule

Visual Decomposition Index (VDI): 0 = intact; 1 = discoloured or mould visible; 2 = softened, < 25 % fragmented; 3 = 25–50 % fragmented or holes; 4 = > 50 % gone, shape lost; 5 = unrecognisable or gone.

Analysis

• % mass loss = (M0 − Mt − handling loss) / M0 × 100. Decay constant (Olson): k = −ln(Mt/M0) / t, in day⁻¹. Compare Run 1 k with Run 2 k.
• H1: mean ± SD per material (n = 3), bar chart with error bars. Kruskal–Wallis test (or one-way ANOVA for older students) in Google Sheets/LibreOffice. Rank order is compared with the predicted one.
• H2: coarse vs fine per material, ratio of means. Run 1 alone gives n = 3, which is low power, so it is reported as an effect size. Pool Run 1 and Run 2 for a t-test.
• Organism preference: animals counted per coarse bag at harvest (worms/isopods per bag by material), plus the isopod choice arena: a 2 L food box with 4 food discs (leaf, cardboard, carrot, peel), 10 Porcellio, counted every 15 min for 1 h, 3 trials. χ² against an equal split.
• H3: respiration rate (method below) is plotted against days since feeding and against core temperature. Q10 = (R2/R1)^(10/(T2−T1)) is fitted from nightly tests. Gas ratio R/R0 is plotted in Grafana as a 24 h median, with feeding events marked as HA annotations.
• Carbon budget (liceu extension): mg C respired per day (CO2 × 12/44) compared with g C lost from the bags plus the food added.

Expected outcomes and when

These are planning estimates for 20–23 °C and moist substrate. By mid-December, H1 is supported at the fast end only (bread and peel versus the rest). H2 will probably not be significant after 28 days and needs Run 2. H3a should be visible within 2 weeks of starting feeding. H3b needs the winter temperature range, so it shows by February. Expected respiration: 5–60 ppm/min headspace rise, i.e. roughly 10–100 mg CO2/h for the whole tank.

Engineering track

Problem statement: Design an indoor biological recycling unit, no bigger than a 50×30 cm shelf footprint, that processes the most class fruit/vegetable scraps per week. Along the way it must not release noticeable odour, must not let fungus gnats escape, and must keep the community aerobic and alive with ≤ 10 min of student work per school day.

Requirements and constraints

Candidate solutions (students compare these, not just build them)

Test criteria and methods

• Carbon-filter efficiency: E = 1 − (1 − R_out/R_room) / (1 − R_in/R_room), from the two BME680s during a fan-on period after feeding. Repeat at bed depths of 5 cm and 10 cm.
• Moisture stability: SD of soil A over 7 days, plus the number of manual waterings needed.
• Throughput: weigh the food added; check the feeding pocket after 7 days (food recognisable: yes/no).
• Odour: blind sniff panel. Containment: count the sticky card. Energy: fan duty-cycle minutes per day from HA history × 12 V × 0.12 A.

Iteration plan

Enclosure & build

Tank: TANK45, glass 50×30×30 cm (≈ 45 L), internal floor ≈ 49×29 cm. It sits on a bench at 75 cm height. Directly above it, on a wall shelf at ≥ 120 cm, are the RCD (30 mA) power strip, PSU12 and the mister's 24 V adapter, with a drip loop on every cable.

Layers (bottom → top)

Two aeration tubes (Ø 20 mm PVC conduit, 25 cm long, 5 mm holes every 2 cm on the lower 15 cm, bottom capped, mesh-covered top) stand in the rear corners, touching the LECA. They double as leachate check and siphon ports: a 6 mm TUBING line and a 60 mL syringe go down one tube.

Lid (from ACRYLIC 3 mm): 50×30 cm in two pieces, sitting on 9×6 mm EPDM foam tape stuck around the rim (this seal makes the respiration test possible).

• Fixed rear strip, 50×10 cm, siliconed to two acrylic cleats inside the rim. It carries:
  • Exhaust port (right): Ø 50 mm hole → 15 cm PVC pipe with a perforated end cap at the bottom, a mesh disc, 75 g activated-carbon pellets (10 cm bed) and a top mesh disc. An 8×8 cm acrylic adapter plate holds the 12 V 80 mm FAN on top, blowing out. BME680 #2 (0x76) sits in a perforated cap 3 cm above the fan outlet (post-filter air).
  • Intake port (left): Ø 20 mm hole fed by a 16 mm ID hose from the mister reservoir, covered with stainless MESH. An SG90 servo mounted on top of the lid drives an acrylic flap over the port: open = 90°, closed = 0° onto a foam pad.
  • 4 × PG7 glands: soil A, soil B, DS18B20, and one I2C cable to the under-lid sensor pair.
  • Under the strip: SCD41 and BME680 #1 (0x77) on a 6×4 cm acrylic plate, 2 cm below the lid, sensor faces pointing sideways (not up, to avoid drips). Their PCBs are conformal-coated with the sensor openings masked.
• Hinged front panel, 50×20 cm: 2 × 40 mm hinges with M3 stainless bolts and nylock nuts, a knob, and a self-adhesive 0.5 mm stainless MESH patch over a 4×8 cm slot for passive-vent tests in v1 (taped shut from v2 on).
• Divider: acrylic 29×16 cm, perforated (see Engineering), pushed into the substrate at 20 cm from the left wall. Feeding zone = left 20 cm; bag field = right 29 cm. The bag grid is then 4 columns × 4 rows in the right zone plus 3 columns × 4 rows in the left zone, with feeding pockets kept to the top 3 cm and rotated among 4 spots in the left zone.

Node box: BOX (IP65, 150×110×70 mm) on an acrylic L-bracket screwed to the bench back wall, its bottom 10 cm above the tank rim, right side. Glands on the underside, cables with drip loops. The box holds the TERMBOARD + ESP32, BUCK (12 → 5 V), MOSFET (fan), RELAY (24 V mister) and the OLED behind a siliconed clear-acrylic window in the box lid.

Mister reservoir: RESERVOIR 10 L on the bench, left of the tank, below the node box. MIST float inside. Two holes in the reservoir lid: a 10 mm meshed fresh-air inlet and a 20 mm fog outlet → 60 cm of 16 mm hose → tank intake. Fog is drawn in by the exhaust fan running at 25–30 %.

flowchart TB
  subgraph SHELF["Wall shelf above, mains zone, RCD 30 mA strip"]
    PSU["PSU12 12 V 5 A brick"]
    AD24["Mister 24 V adapter"]
  end
  subgraph NODE["Node box mc-01-rot IP65"]
    ESP["ESP32 on TERMBOARD + OLED"]
    BUCK["BUCK 12 to 5 V"]
    MOS["MOSFET fan GPIO17"]
    REL["RELAY mister GPIO23"]
  end
  subgraph LID["Acrylic lid with EPDM gasket"]
    EXH["Exhaust tube with 75 g carbon"]
    FAN["FAN 80 mm 12 V"]
    BME2["BME680 no.2 post-filter 0x76"]
    SERVO["SG90 intake damper GPIO26"]
    UNDER["SCD41 and BME680 no.1 under lid"]
  end
  subgraph TANK["TANK45 50x30x30 cm"]
    HEAD["Headspace about 17 L"]
    LITTER["Leaf litter 2-3 cm, cork hides"]
    BAGS["Litter-bag plane at 5 cm, 27 + 6 window bags"]
    SOILA["SOIL A feeding zone GPIO32"]
    SOILB["SOIL B bag field GPIO33"]
    DS["DS18B20 core at 8 cm GPIO4"]
    LECA["LECA sump 3 cm"]
    AER["2 aeration and leachate tubes"]
  end
  RES["RESERVOIR 10 L with MIST float"]
  PSU --> BUCK --> ESP
  PSU --> MOS --> FAN
  AD24 --> REL --> RES
  RES -- "fog hose 16 mm" --> SERVO --> HEAD
  HEAD --> EXH --> FAN --> BME2
  UNDER --> ESP
  BME2 --> ESP
  SOILA --> ESP
  SOILB --> ESP
  DS --> ESP
  ESP --> SERVO

Step-by-step build (≈ 14 h of team time over weeks 1–3)

Biology

Organisms and Romanian sourcing

No vertebrates. Isopods and worms are invertebrates, which is compliant with the brief.

Substrate recipe (≈ 17 L, made in a 30–60 L tub; wear gloves)

• 6 L coco coir (one 650 g brick hydrated with 3 L water, squeezed)
• 6 L sieved (10 mm) school-garden soil and mature compost, 1:1. Commercial "compost/humus" only if it has no added fertiliser.
• 3 L brown corrugated cardboard (≈ 250 g dry), shredded into 1–2 cm pieces, no glossy or coloured print
• 2 L shredded dry leaves (lime/maple)
• 1 L mature compost + 0.5 L leaf mould (the inoculum)
• 100 g eggshell (boiled 5 min, dried, crushed) for calcium for isopods and as a pH buffer
• Water to "wrung-out sponge" moisture: a hard squeeze gives 1–3 drops (≈ 65–70 % water, wet basis). Record the total water added.
• Weigh a 200 g sample, dry it, and record the dry fraction to estimate the bed's dry mass (expected ≈ 4 kg) for the mass-specific respiration rate.

Establishment and acclimation

1. Day 0 (Nov 9): layer the bed, insert the probes, close the lid, run the fan base schedule. Watch CO2, T and soil for 3 days. The core must stay < 30 °C; if it heats up, the bed is too rich, so mix in 1 L dry cardboard.
2. Day 0: start a Pleurotus pocket in a separate 2 L lidded box. 200 g cardboard is soaked in boiling water for 1 h (pasteurisation), cooled and drained, then layered with 50 g spawn. It sits in the dark at room temperature for 2–3 weeks until white. Half goes into the tank on Dec 1 (rear-right corner, 5 cm deep); the rest makes the Run 2 "pre-colonised cardboard" bags.
3. Day 3 (Nov 12): add the worms on the surface under room light. They burrow in within 30 min; any that stay on the surface after 2 h mean conditions are wrong (check moisture and pH). Day 4: add the springtail culture and the isopods under the cork hides.
4. Days 3–7: no feeding beyond the bed's own cardboard and leaves. Weekly checks start.
5. Nov 16: bury the Run 1 bags in the randomised map, then start feeding.

Feeding (left feeding zone only, rotate among 4 pockets, bury 2–3 cm under the litter)

• Weeks 4–5: 50 g twice a week. Weeks 6–8: up to 100 g twice a week if the previous food has gone. Always add an equal volume of shredded cardboard ("browns").
• Freeze all scraps for ≥ 48 h, then thaw. This kills gnat and fruit-fly eggs and softens tissue. Chop to < 2 cm.
• OK: carrot, cucumber, lettuce, squash/pumpkin, apple/pear cores, banana peel, melon rind, coffee grounds (≤ 20 % of a feed), loose tea leaves, eggshell.
• NOT allowed: meat, fish, bones, dairy, eggs (other than shell), oily/fried/salty/cooked sauced food, bread and pasta in bulk (bread only inside test bags), citrus, onion, garlic and chilli beyond the 3 g orange-peel test bags, pet faeces, glossy or coloured paper, treated or painted wood, pesticide-sprayed plants, any "compostable" plastic outside test bags.

Hygiene

• Nitrile gloves for every hand contact; wash hands afterwards. Lid opened ≤ 10 min. No eating or drinking at the bench.
• Mould: white mycelium and pale mould are normal and tolerated. Green, black or pink fuzzy mould on the surface: cover it with 1 cm of dry leaves or remove it with tongs into a sealed bag. Never shake mouldy material. Students with asthma or a known mould allergy do not do harvests or sieving. Harvest and sieving are done with FFP2 masks, next to an open window, with no fans blowing toward people.
• Fungus gnats: frozen food, food always buried, 0.5 mm mesh on every vent, a yellow sticky card outside the lid and one inside on the litter. If more than 10 per week on the inside card: let the top 2 cm dry for 3 days and add BTI (Bacillus thuringiensis israelensis) granules to the misting reservoir/top watering at the label dose.
• Odour: sour or rotten-egg smell means anaerobic conditions. Stop feeding, turn the feeding zone, add 1 L dry cardboard, run the fan at 60 % for 24 h.

Instrumentation

What the BME680 does and does not measure. Its gas channel is a heated metal-oxide (MOX) film whose electrical resistance drops when reducing gases are present: ethanol, acetone, aldehydes, esters, terpenes (limonene from orange peel), H2S and amines from rotting food. It gives one number for the total of all of them. It does not identify which gas is present, does not give ppm, does not measure CO2 (the SCD41 does), and responds weakly to ammonia. It is affected by humidity and temperature and drifts over days. Therefore:

1. Run a 48 h burn-in first.
2. Use the ratio R/R0, where R0 is the 24 h median in the clean classroom before installation, re-checked monthly with the lid open for 30 min.
3. Compare the two sensors under identical conditions (#1 vs #2) rather than absolute values.
4. Validate against the blind sniff panel.

We use the raw bme680 ESPHome platform (gas resistance in Ω). Bosch's BSEC "IAQ" index is an optional extra for advanced students, with the caveat that it was tuned for living rooms, not compost. The BME680 RH reading is unreliable above 90 % RH, so the humidity loop uses the SCD41's RH.

Pin assignment (standard map only)

Entity ids: mc_01_co2, mc_01_air_temp, mc_01_air_rh, mc_01_gas_res_in, mc_01_gas_res_out, mc_01_gas_ratio_in, mc_01_core_temp, mc_01_soil_moisture_a, mc_01_soil_moisture_b, mc_01_resp_rate, mc_01_fan, mc_01_mister, mc_01_intake.

Calibration

• SOIL A/B: in the real substrate, record raw volts in (1) air, (2) oven-dried substrate in a 500 mL cup, (3) saturated substrate. Then make 3 intermediate mixes, measure them gravimetrically, and fit a line → calibrate_linear to % water (wet basis). Re-check monthly with the weekly gravimetric samples.
• DS18B20: ice-water slurry (0.0 °C) and room water against a reference glass/digital thermometer; set an offset if > 0.3 °C.
• SCD41: automatic_self_calibration: false, because the tank never sees fresh air for long enough. Forced recalibration to 420 ppm: hold it at an open window outdoors for 5 min after a 3 min warm-up, monthly. Cross-check side by side with the Living Earth SCD41 for 1 h (should agree within ± 50 ppm).
• BME680: 48 h burn-in; R0 baseline as above; check that #1 and #2 agree within 15 % in the same air before installing #2 in the exhaust.
• Scale: check with a 100 g calibration weight before each weighing session.

Respiration test protocol (30 min, manual version; the automatic nightly version is identical)

1. Do not open the lid or feed for ≥ 2 h beforehand. Note the time, core T and air T.
2. In HA, press "Start respiration test". The node turns the fan and mister off and the servo closes the intake. Cap the exhaust outlet with the rubber cap (manual tests only).
3. Log CO2 every 60 s for 30 min. Discard minutes 0–5 (mixing). Abort if CO2 > 4500 ppm (sensor limit 5000).
4. Slope s (ppm/min) = linear regression of minutes 5–30 (Sheets SLOPE). The node also publishes (C25 − C5)/20.
5. CO2 production = s × 10⁻⁶ × V / Vm, where V = headspace 17 L (measure the actual headspace height) and Vm = 24.2 L/mol at 22 °C (scale by (273 + T)/295). Example: s = 20 ppm/min → 1.4 × 10⁻⁵ mol/min = 14 µmol/min → × 44 = 0.62 mg CO2/min = 37 mg CO2/h → ÷ 4 kg dry bed = 9 mg CO2/h/kg.
6. Re-open the intake and run the fan at 60 % for 10 min. Record any leak notes (the gasket condition is checked monthly with a smoke/incense test held outside the gasket).
7. Known bias: some CO2 leaks through the gasket and carbon tube, so rates are slight underestimates. Use them for comparisons, not absolutes.

Automation

All logic runs on the ESP32 (it keeps working if Wi-Fi or HA is down). HA provides visibility, override and alerts.

Setpoints

stateDiagram-v2
  [*] --> Boot
  Boot --> Base: intake open, mister off, fan 40 pct 2 min
  Base --> Vent: gas ratio below 0.60 or CO2 above 2500
  Vent --> Base: ratio above 0.75 and CO2 below 1500
  Vent --> VentRest: 20 min at full speed
  VentRest --> Vent: after 10 min at 40 pct
  Base --> Humidify: RH below 82 and soil A below 70
  Humidify --> Base: RH at least 90 or 90 s elapsed
  Base --> RespTest: 02.00 schedule or HA button
  RespTest --> Base: 30 min done, slope published
  RespTest --> Vent: CO2 above 4500 or timeout 35 min
  Base --> FailSafe: SCD41 missing 5 min
  Vent --> FailSafe: SCD41 missing 5 min
  Humidify --> FailSafe: sensor missing
  FailSafe --> Base: sensor back for 5 min
  Base --> Manual: HA override
  Manual --> Base: Auto selected or 2 h timeout

ESPHome implementation outline (students fill in the full YAML)

• sensor: scd4x (update_interval 60s, automatic_self_calibration: false). sensor: bme680 ×2 (addresses 0x77/0x76, heater: {temperature: 320, duration: 150ms}). dallas_temp on GPIO4. adc on GPIO32/33 with attenuation: 12db, a sliding_window_moving_average (window 10) and calibrate_linear.
• output: ledc GPIO17 at 25 kHz → fan: speed "Rot fan". switch: gpio GPIO23 "Rot mister". servo on ledc GPIO26 at 50 Hz.
• A template R/R0 sensor (R0 is a number entity set from HA). select "Mode" with options Auto / Manual / Test. button "Start respiration test" → script (fan off, mister off, servo 0°, delays, publishes mc_01_resp_rate).
• Triggers via on_value_range, for example:

  sensor:
    - platform: scd4x
      co2:
        name: "mc_01_co2"
        on_value_range:
          - above: 2500
            then:
              - if:
                  condition:
                    lambda: 'return id(mode).state == "Auto";'
                  then:
                    - script.execute: vent_cycle
  

• Duty limits and timeouts use script with mode: single, delay, and an interval: 60s supervisor that checks has_state() on every sensor for the fail-safe.
• Manual override in Home Assistant: dashboard card with Mode select, fan speed slider (0–100 %), mister toggle, intake open/close, test button. Manual mode auto-reverts to Auto after 2 h (template script on the node). Mains is not involved anywhere: the fan and servo are 12/5 V, the mister is 24 V DC switched on its low-voltage side, and no smart plug is needed for this ecosystem.

Operations

Daily (school days, ≤ 10 min, task rota on the tank)

• Glance at the OLED and HA dashboard: Mode = Auto; CO2 < 2000; soil A/B 60–75 %; core T 15–28 °C; no alerts.
• Check the reservoir level (refill to 8 L with tap water that has stood 24 h; never above the MIST float limit).
• Look through the glass: worms on the glass or surface? sour smell? standing water in the LECA? Note it in the logbook.
• Count gnats on the outside sticky card (tally).

Weekly

• Mon: window-bag photos + VDI; pH 1:5 extract; gravimetric samples (2 depths).
• Tue/Fri: feed per plan (weigh food and cardboard), manual respiration test (Tue 12:00 is the pre-feed test, Fri 12:00 is ≈ 72 h after feeding).
• Thu: carrot bait slices out. Fri: count bait slices + cork hides; sniff panel; review the Grafana week and annotate events.
• Fri: empty the mister reservoir, rinse, wipe with 3 % hydrogen peroxide or dilute vinegar, rinse, refill (stops biofilm/Legionella-type growth).
• Check leachate: if > 1 cm of water stands in the LECA, siphon it out via an aeration tube; dilute 1:10 and use it on school plants (not on food crops).

Lab notebook (paper, numbered pages, plus the shared sheet)

• Date, time, initials; food type and mass in; cardboard mass in; water added (mL); lid-open minutes.
• Observations (mould colour and location, animals seen, smell), VDI scores, counts, pH, gravimetric values.
• Respiration test raw values (C5, C25, slope, T) and any deviations (lid opened, test aborted).
• Any firmware or setpoint change (version, reason). The litter-bag map and bag mass table, with M0 to 0.01 g.

Maintenance

• Activated carbon: replace (75 g) when R_out/R_room < 0.85 during venting or every 8 weeks.
• Mesh and sticky cards: replace cards every 2 weeks.
• Re-coat soil probe tops if bare copper or swelling is visible; recalibrate monthly.
• SCD41 FRC monthly; BME680 R0 monthly; gasket check monthly.
• Thin the worm population if it exceeds ≈ 300 g (estimated from the March harvest), giving the extras to the Living Earth team or the school compost.

End-of-term plan (14–18 Dec)

• Dec 14: harvest Run 1 (count animals in bags, rinse, dry). Dec 15: insert Run 2 bags (new randomised map). Dec 16–18: final weighing, first results poster.
• Before winter break (approx. Dec 19 – Jan 10): one feed of 150 g frozen scraps + 1 L cardboard on Dec 17. Reservoir full. Mode Auto. HA alerts to the teacher's phone (core T, CO2, soil, node offline).
• If the school heating drops below 10 °C, decomposition just slows. E. fetida tolerates ≥ 5 °C, so no heater is added (keeps the design SELV-simple). The data becomes a natural low-temperature treatment for the Q10 analysis.
• If the node is offline for > 24 h over the break, the teacher checks the room. The bed is safe for 3 weeks unattended.
• End of school year (June 2027): Run 3 harvest, vermicompost harvested by light migration (food on one side for 2 weeks, scoop the other side), organisms rehomed to the school compost bin, tank cleaned and stored, data archived (InfluxDB + CSV).

Safety & risks

The general rules still apply: no 230 V anywhere inside or near the tank. The fan and servo are 12/5 V. The mister is 24 V DC with its own certified adapter. All bricks sit on the shelf above, on the RCD strip, with drip loops. The node box is outside the tank, above the rim, with cables entering through glands. Students never open or modify mains devices.

Educational outcomes & assessment

Sub-BOM

Cost-cut 2026-10-20: BTI granules removed (buy only if fungus gnats appear). PPE, pens and calibration solutions are shared project items.
Price basis: Romanian retail incl. VAT, Oct 2026 estimates. Electronics: optimusdigital.ro / sigmanortec.ro / cleste.ro (SCD41 modules are listed at sigmanortec.ro). Hardware: dedeman.ro / leroymerlin.ro. Organisms: OLX / vermiculture and terrarium sellers. Fallback: AliExpress.

PHPEN, CALSOL, the hub (PI5, ROUTER) and the RCD power strip are project-wide shared items and are not counted here.

8-week timeline (Oct 26 – Dec 18, 2026)

Check the 2026/27 school calendar. If a holiday falls in this window, shift by one week; Dec 1 (Romania's national day) is a public holiday.

gantt
  title Orchestra of Rot - build and Run 1
  dateFormat YYYY-MM-DD
  axisFormat %d %b
  section Prepare
  Order parts and organisms          :p1, 2026-10-26, 5d
  Collect leaves and isopods         :p2, 2026-10-27, 8d
  Lid and vent design review         :p3, 2026-10-29, 2d
  section Build
  Cut and drill lid and vent tube    :b1, 2026-11-02, 4d
  Node box wiring and ESPHome        :b2, 2026-11-03, 4d
  BME680 burn-in and calibrations    :b3, 2026-11-06, 4d
  Sew and weigh litter bags          :b4, 2026-11-04, 6d
  section Establish
  Layer substrate and settle         :e1, 2026-11-09, 3d
  Pleurotus cardboard colonisation   :e2, 2026-11-09, 21d
  Add worms springtails and isopods  :e3, 2026-11-12, 2d
  section Run 1
  Bury Run 1 bags and start feeding  :milestone, r0, 2026-11-16, 0d
  v1 passive vent baseline           :r1, 2026-11-16, 7d
  v2 carbon fan AB test              :r2, 2026-11-23, 10d
  Isopod choice arena trials         :r3, 2026-11-24, 8d
  v3 mister loop and auto resp test  :r4, 2026-12-03, 9d
  Harvest Run 1 bags                 :milestone, r5, 2026-12-14, 0d
  Dry weigh analyse and poster       :r6, 2026-12-14, 5d
  Insert Run 2 bags                  :milestone, r7, 2026-12-15, 0d
  Winter break preparation           :r8, 2026-12-17, 2d

2. The Domino Effect — EcoColumn

Node: mc-02-ecocolumn · Team of 3–4 · Footprint 90×40 cm, 180 cm tall rack · Water on board ≤ 35 L, all 12 V DC.

Concept (refined)

A four-level column on a boltless metal rack: a tilting "field" of two removable soil cassettes (lanes A/B) receives metered artificial rain; runoff carrying sediment and fertiliser drops through swappable filter cartridges into a 25 L "pond" with plants, snails and algae. Rain comes from a separate clean reservoir, so the only nutrient input to the pond is what washes off the field (closed nutrient accounting). The A/B lanes and interchangeable cassettes give every erosion experiment its own side-by-side control. An array of twelve 0.5 L jars beside the pond runs a fast fertiliser-dose bioassay, so eutrophication results exist by mid-December, while the pond itself is the slow, whole-system experiment that runs to June 2027.

Science track

Driving question: How much soil and fertiliser leaves a field during a storm, and what does it do to the pond downstream?

Hypotheses (testable by the team)

Variables

Replication strategy

1. Bare-vs-bare calibration first (3 events, both lanes bare): measures the natural lane-to-lane difference. A later treatment effect only counts if it is ≥ 2× this calibration difference.
2. Four interchangeable cassettes (2 vegetated, 2 bare): every comparison is run once, then the cassettes are swapped left↔right and run again, cancelling position bias (rain bar uneven, light gradient).
3. Repeat rain events: ≥ 3 events per condition (2 per week, ≥ 48 h apart).
4. Before/after on the pond: 2-week baseline with unfertilised runoff, then fertilised runoff; compare weekly means.
5. Jars: 3 replicates per dose, positions assigned by dice, rotated one place every 2 days.

Sampling schedule

Manual measurements (methods)

• Sediment mass (TSS method): collect the whole event's runoff per lane in a 1 L graduated jug; read total volume V. Stir hard, immediately pour 100 mL into a measuring cylinder; filter through a pre-weighed coffee filter (weigh dry filter on 0.01 g scale). Dry 48 h on a radiator shelf or 2 h at 105 °C in an oven (teacher). Weigh. TSS (g/L) = (dry filter − filter) / 0.1 L; sediment yield = TSS × V; per area = yield / 0.048 m². For bare lanes with heavy loads, also let the remaining jug settle 24 h, decant and dry the whole sediment in a pre-weighed foil tray (cross-check).
• Turbidity: the second TURB probe in a black-sleeved 250 mL bench cup (see Instrumentation), converted to "g/L kaolin-equivalent" via the calibration curve. Backup with no electronics: a 60 cm clear transparency tube with a Secchi pattern on the bottom (depth at which pattern disappears, cm).
• pH: continuous PHKIT in pond; PHPEN (shared pool) for jars and runoff.
• Nitrate/phosphate: aquarium 6-in-1 strips for quick NO3 (0–250 mg/L, coarse); JBL-type NO3 and PO4 drop tests (colour card, ~0.05–1.8 mg/L PO4) for weekly pond and filter in/out samples. Always test tap water once as baseline (Bucharest tap water contains some nitrate).
• Algae: (a) Secchi card: a 5 cm black/white quadrant disc on a stick, lowered in the pond, depth of disappearance (cm); (b) photo green index: phone on a fixed stand, jar in front of a white card, LED light only, fixed exposure (pro mode or lock AE); in a free app (ImageJ/Fiji, or a spreadsheet of RGB values from a colour-picker app) compute G/(R+G+B) on a centre patch; (c) Lemna minor frond count inside a floating 10 cm PVC ring.

Analysis

• Per event: sediment yield (g/m²), runoff coefficient = runoff volume / rain volume, filter efficiency = (C_in − C_out)/C_in for turbidity and TSS.
• H1/H3: paired differences (A − B) per event; mean ± SD; ratio vegetated/bare; check effect ≥ 2× calibration difference. Liceu: paired t-test or sign test in a spreadsheet; gimnaziu: bar chart with min–max whiskers.
• H1b: scatter sediment vs angle, fit a line or power curve.
• H2: line chart turbidity and green index vs day per dose (mean of 3 jars); dose–response at day 14; rank order test.
• Pond: Grafana time series of pH (daily max − min = photosynthesis swing, a productivity proxy), turbidity, temperature, level; before/after weekly means.

Mid-December vs. spring

Design consequence: the pond is started in week 2 (not week 6) so it is biologically stable before the first nutrients arrive, and the jars carry the eutrophication result for the December presentation.

Engineering track

Problem statement: Design a field-edge system that keeps soil and fertiliser on the field and out of the pond, using only cheap natural materials, without flooding during a 20 mm storm, and that a farmer (a student) can service in under 2 minutes.

Requirements

Candidate solutions

• Rain bar: (a) PVC 16 mm pipe with 2 rows × 8 holes Ø 0.8 mm (pin vice) per lane; (b) same pipe with 23G blunt dispensing needles (more uniform drops); (c) holes + a fine mesh 5 cm below to break drops into a wider pattern.
• Filter cartridges (2.5 L PP boxes, 20×14×10 cm, outlet in bottom under mesh): gravel 4–8 mm; pool-filter sand 0.4–0.8 mm over gravel; layered gravel/sand/activated charcoal; plant biofilter = sand + 3-week ryegrass mat on top; student-invented (e.g. coir, straw, biochar).
• On-field measures (tested on cassettes): grass buffer strip, straw mulch, contour ridges (3 acrylic strips across the slope), coir mat, check dams of gravel.

Test criteria: sediment removal (%), turbidity removal (%), NO3/PO4 change, overflow (yes/no, volume in catch tray), clogging (outflow rate at event 1 vs event 5), swap time, cost.

Iteration plan

1. v0 (week 3): rain bar (a), measure CU; if < 80 %, build (b) and retest.
2. v1 filters (week 5): build 3 cartridges (gravel, sand, layered); run inline on bare-lane runoff; rank with the criteria above.
3. v2 (week 7): redesign the best one (e.g. thicker sand, plant mat on top, coarser layer against clogging); retest; compare v1 vs v2 on the same event type.
4. Spring: lifetime test (10 events), cleaning method, combine best on-field measure + best filter ("whole-catchment solution") and measure pond impact.
   Every iteration is logged in the design notebook: requirement → change → test result → decision.

Enclosure & build

Layout (boltless metal rack 180×90×40 cm, 5 shelves, anchored to the wall with an anti-tip strap; shelves lined with PP sheet)

Soil cassettes (×4): 3 mm ACRYLIC, inner 40 × 12 × 6 cm; base, two sides, back wall; front end open with a 1 cm sill and a 2 cm strip of MESH below the sill so seepage also exits. Fill 5 cm soil mix (2 parts garden topsoil : 1 part washed sand, sieved 5 mm) = 2.4 L per cassette. Lane area 0.048 m² → 1 mm of rain = 48 mL per lane.

Rain rate (computed): pump flow is trimmed (PWM + bypass valve) to Q = 0.40 L/min, split equally to both lanes (servo valve "both") = 0.20 L/min per lane = 4.2 mm/min instantaneous. Rain is pulsed (like oscillating-nozzle field simulators): each 15 s pulse = 100 mL total = ~1 mm per lane.

Pulsing keeps the FLOW sensor inside its accurate range (≥ 0.3 L/min). Limitation to discuss with students: 25 cm drops have far less impact energy than real rain, so absolute erosion is underestimated; comparisons remain valid.

Water path

flowchart TD
  R["RESERVOIR 15 L dechlorinated water"] --> P["PUMP 12 V via MOSFET PWM3"]
  P --> BV["Bypass ball valve back to reservoir"]
  BV --> R
  P --> F["FLOW YF-S401"]
  F --> NRV["Non-return valve"]
  NRV --> V["3-way valve on SERVO, A / both / B"]
  V --> RA["Rain bar A"]
  V --> RB["Rain bar B"]
  RA --> CA["Cassette A on tilt platform"]
  RB --> CB["Cassette B on tilt platform"]
  CA --> GA["Funnel A, hose"]
  CB --> GB["Funnel B, hose"]
  GA --> JA["Optional 1 L sample jug A"]
  GB --> JB["Optional 1 L sample jug B"]
  GA --> FA["Filter cartridge A"]
  GB --> FB["Filter cartridge B"]
  FA --> PD["Pond TANK25, 18 L"]
  FB --> PD
  FA --> CT["Catch tray overflow"]
  FB --> CT
  CT --> PD
  PD --> DD["Weekly drawdown to 5 L jug, max level 21 cm"]
  S1(["FLOAT reservoir low"]) -.-> R
  S2(["SOIL A and SOIL B"]) -.-> CA
  S2 -.-> CB
  S3(["BH1750 at water surface"]) -.-> PD
  S4(["JSN in stilling tube"]) -.-> PD
  S5(["FLOAT pond high"]) -.-> PD
  S6(["TURB, PHKIT, DS18B20"]) -.-> PD

Overflow policy: no siphon overflow on the glass tank (siphons can fail and drain or flood). Instead: 7 cm freeboard (18 → 25 cm = 7 L spare), FLOAT high at 21 cm blocks all events, JSN level > 20 cm blocks events, weekly manual drawdown to 18 cm (this is also the pond sample), spill box under the tank holds > 25 L. One S20 storm adds ≤ 2 L (≈ 2 cm).

Build steps

Total ≈ 22 student-hours + cure times.

Biology

Pond (TANK25, 18 L)

Do not buy Elodea nuttallii or Lagarosiphon major (EU invasive species of Union concern). Prefer Planorbarius over Physa acuta (invasive). No organism or water from this system is ever released outdoors: plant waste is dried and binned, water goes down the sink.

Field (cassettes): 2 vegetated cassettes sown densely with Lolium perenne (perennial ryegrass, 5 g per cassette) + 2 g wheat (Triticum aestivum, germinates in 3–5 days, fast cover) — one cassette may be ryegrass + white clover (Trifolium repens, 1 g) for spring comparison. 2 bare cassettes = control. Buffer strip: a 8 × 12 cm sod of ryegrass grown in a separate small tray, cut and inserted at the low end of a bare cassette for H3. Seeds from agricultural/garden shops (Dedeman, Agroland, eMAG).

Establishment timeline: sow week 1 (Oct 26–30) under the slope LED (or a windowsill), keep moist with a spray bottle; ≥ 80 % cover at day 18–21 (by Nov 16–20). Pond filled week 2, plants week 2, snails and inoculum week 3 (after 7 days of plant settling). Jars started Nov 18.

Fertiliser (measured doses)

• Field: granular NPK 15-15-15, 2.0 g per cassette (= 42 g/m², ≈ 63 kg N/ha, a realistic field rate), sprinkled evenly 24 h before a "fertilised" event.
• Jars: stock solution 10 g soluble NPK 20-20-20 in 1.00 L water (10 mg/mL; label, store dark, 4 weeks). Per 500 mL jar (450 mL dechlorinated water + 50 mL sieved pond water): 0 / 0.25 / 1.0 / 4.0 mL stock by syringe → 0 / 1 / 4 / 16 mg N/L and 0 / 0.44 / 1.7 / 7 mg P/L. Jars covered with cling film with pin holes; topped up to the mark with dechlorinated water.

Avoid: fish or any vertebrate (brief rule; also adds nutrients); mosquito risk — check weekly for larvae, cover the pond with a MESH frame lid if seen; peat-based potting compost on the slope (floats, unrealistic erosion); copper-containing products; pesticides.

Instrumentation

TURB notes: SEN0189-type sensors measure light transmitted across a 1 cm gap (IR LED → phototransistor); output 0–4.5 V, falls as turbidity rises. The "NTU" formula in the datasheet is approximate — we calibrate in g/L kaolin-equivalent instead. Limitations: sensitive to ambient light (hence the opaque sleeve and black cup), bubbles and algal film on the window (wipe weekly with soft cloth), not accurate below ~0.05 g/L, and the probe top is not waterproof (keep top 2 cm dry; seal cable entry with heat-shrink).

PHKIT notes: the glass probe must never dry out; out of water store it in its cap with 3 M KCl storage solution (never distilled water). Drift ~0.05–0.1 pH/week: check weekly in pH 7, recalibrate 2-point monthly or when drift > 0.1. Algal film on the bulb: rinse with tap water, soak 10 min in vinegar if fouled. Probe lifetime in continuous immersion 6–12 months: budget a replacement probe in spring. Ground-loop check: read pH with pump/LEDs on and off; if the reading jumps > 0.05, add an analog signal isolator.

Pin assignment (standard pin map only)

Power: PSU12 → MOSFET loads (LEDs ≤ 10 W/m × 3.2 m ≈ 32 W, pump ~4 W) and BUCK 5.0 V → ESP32 VIN, JSN, FLOW, TURB ×2, PHKIT, SERVO. Worst case ≈ 39 W = 3.3 A of 5 A. All ADC pins on ADC1, attenuation 11 dB (attenuation: 12db in current ESPHome).

Fail-safe float wiring: both floats are mounted so the contact is closed (to GND) in the safe state — reservoir float closed while water is above 3 L; pond float closed while level is below 21 cm. Open circuit (triggered or broken wire) = unsafe → pump blocked.

ESPHome flow excerpt (YF-S401 ≈ 5880 pulses/L; replace with your calibrated value):

sensor:
  - platform: pulse_counter
    pin: GPIO5
    name: "mc_02_rain_flow"
    unit_of_measurement: "L/min"
    update_interval: 5s
    filters:
      - multiply: 0.00017007   # 1/5880 L per pulse
    total:
      name: "mc_02_rain_total"
      unit_of_measurement: "L"
      filters:
        - multiply: 0.00017007

Calibration procedures

Automation

Rain-event controller runs inside the ESP32 (ESPHome script), so an event always finishes or aborts even if Wi-Fi or Home Assistant is down. HA only triggers it (dashboard button with preset S10/S20/D2 and valve position) or schedules D2 drizzle when both SOIL values fall below the target band.

stateDiagram-v2
  [*] --> Idle
  Idle --> Precheck : HA button or schedule
  Precheck --> Fault : interlock open
  Precheck --> SetValve : all checks pass
  SetValve --> Raining : servo settled 2 s
  Raining --> Pausing : pulse time elapsed
  Pausing --> Raining : pulses remaining
  Pausing --> Draining : all pulses delivered
  Raining --> Fault : float open or no flow or volume cap
  Pausing --> Fault : float open
  Draining --> Logging : 15 min runoff window
  Logging --> Idle : event record written
  Fault --> Idle : manual reset in HA

• Precheck: reservoir float closed; pond float closed; JSN level valid and < 20 cm; last event ≥ 2 h ago; for experiment events, SOIL A and B within the target band (warn, do not block, if outside).
• Raining: pump ON at calibrated PWM; flow pulses counted; pulse ends on time (15 s).
• Logging: event ID, preset, valve position, angle (entered in HA), pulses, litres, mean L/min, SOIL before/after, pond level before/+15 min. Stored in InfluxDB and appended to the nightly CSV.
• Fault: pump OFF, servo to "both", HA notification and OLED message; requires manual reset.

LED schedule

Winter break (Dec 21 – Jan 7): lights continue on schedule, rain events disabled, reservoir and pond topped up on the last school day, HA alerts sent to the teacher's phone.

Safety interlocks (summary)

1. Pump runs only while both floats are in the safe (closed) state and JSN < 20 cm.
2. Hard caps per event: ≤ 2.0 L (11,760 pulses) and ≤ 300 s cumulative pump time.
3. Flow fault: no pulses within 10 s of pump ON → abort (dry pump, kink or sensor failure).
4. restore_mode: ALWAYS_OFF on pump and MOSFET outputs; after any reboot the pump is off.
5. Minimum 2 h between events; maximum 4 events per day.
6. All loads are 12 V DC; the only mains items (PSU12 brick, RCD strip) are on the wall shelf beside the rack.

Operations

Per-event checklist (≈ 40 min)

1. HA dashboard: no faults; reservoir float OK; pond level < 20 cm; note SOIL A/B.
2. Set angle (prop block, check inclinometer); rake bare cassettes 1 cm; photo of both cassettes from the fixed point.
3. Place clean 1 L jugs under hoses A and B (erosion mode) or leave hoses in cartridges (filter mode).
4. Start preset in HA; enter event ID, angle, cassette IDs, treatment.
5. After the 15 min runoff window: read jug volumes; take 100 mL TSS subsamples and bench-turbidity samples; label (date_event_lane).
6. Filter mode: grab 100 mL at hose inlet and at pond-end outlet at minutes 2, 5, 10.
7. Pour remaining jug runoff slowly through its cartridge into the pond (keeps nutrient accounting closed).
8. Photo after; write the event log line.

Weekly (Monday, ≈ 30 min): pond Secchi, photo green index, Lemna count, NO3/PO4, pH pen vs PHKIT; drawdown pond to 18 cm into the 5 L jug (this is the weekly pond sample); top up reservoir to 15 L with dechlorinated water; wipe TURB window; check pH probe in pH 7; check for mosquito larvae; inspect tubing and drip loops; refill catch tray dry.

Manual logs (printed sheet + shared spreadsheet): event log (ID, date, preset, angle, cassettes, rain L, runoff A/B mL, TSS A/B, turbidity A/B, notes); pond log; jar log; maintenance log (calibrations, cleaning, faults).

Filter cleaning: after every 5 events or when outflow time doubles: tip the top 1 cm of sediment-laden sand into a labelled bag (keep as sample), backwash gravel/sand in a bucket, re-pack to the marked height. Charcoal is replaced, not washed, every 10 events.

Water changes: no routine pond water changes (they would remove the nutrients being studied); only weekly drawdown and evaporation top-up with dechlorinated water. Emergency only: if snails die or water smells of rotten eggs, replace 30 % and log it.

Sample storage: nutrient samples tested the same day, or stored ≤ 48 h at 4 °C in a labelled screw-cap bottle in a dedicated (non-food) fridge or cool box; dried sediment filters in labelled zip bags in a dry box (archive for the year); photos named mc02_YYYYMMDD_subject.jpg in the shared folder.

Safety & risks

• HARD RULES apply: no 230 V is built, wired, modified or switched by students; the only mains items (RCD-protected strip, certified PSU12 brick) sit on a wall shelf beside and above the rack, away from water, with drip loops. Nothing mains-powered is in water, in soil, or inside the column. Everything that touches water — pump, sensors, LED strips — is ≤ 12 V DC.
• Water near electronics: node box IP65 on the side upright at 110 cm, outside the shelf footprint, glands facing down, drip loop on every cable; PH-4502C and all boards inside the box; LED strips in aluminium profiles, never submerged.
• Spills: pond and reservoir each stand in a spill box larger than their volume; shelves lined; mop and towels kept at the rack; rack wall-anchored against tipping (≈ 35 kg of water at 40 cm and floor level only).
• Overflow: fail-safe floats, JSN limit, event volume caps, 7 cm freeboard, weekly drawdown (see Automation).
• Hygiene: nitrile gloves for soil, pond water and fertiliser; wash hands after every session; pond water is never tasted, pipetted by mouth or poured into food containers; fertiliser and stock solution labelled and locked in the teacher's cupboard; snails handled with a net, not hands.
• Biology: no vertebrates; no release of any organism or water outdoors; avoid invasive species listed in Biology.
• Glass: rimless tank handled by two people, empty only; no load on the glass rim.

Educational outcomes & assessment

Students will be able to:

1. Formulate a testable hypothesis on erosion/eutrophication and name independent, dependent and controlled variables.
2. Explain why a calibration (bare vs bare) run and a position swap are needed before claiming an effect.
3. Measure sediment yield (TSS × volume), turbidity, pH, NO3/PO4, and convert rain to mm and mm/h.
4. Interpret a dose–response and a time series (diurnal pH swing) from the dashboard.
5. Write requirements, build and test a filter against them, and justify a v2 from data.
6. Explain the chain field → runoff → nutrients → algae → light/oxygen in the pond, and link it to real Romanian cases (Danube Delta and agricultural lakes eutrophication, EU Nitrates Directive buffer strips).

Assessment artefacts: lab notebook (individual), event/pond/jar logs (team), engineering design log with requirements table and test results (team), 5-minute December presentation with one graph per hypothesis (team), short individual reflection.

Roles rotate every 2 weeks: hydrologist (rain, slope), chemist (water tests), engineer (filters, hardware), data lead (HA/Grafana, spreadsheet).

Sub-BOM

Cost-cut 2026-10-20: lighter shelving, strips plus one PO4 test instead of two drop tests, shared precision scale, electrical consumables moved to the project electronics lot.

Prices: estimates incl. VAT, RON (1 EUR ≈ 5.1 RON), October 2026, basis given per line.

Not included (shared pool): PHPEN, TDSPEN, syringes, labels, sample bags, small tools, hub/router. Spring item to budget: spare pH probe ≈ 70 RON.

Changes vs. draft allocation: SOIL 1 → 2 (one per lane, needed for antecedent-moisture control and A/B replication); TURB 1 → 2 (second as bench probe for runoff, filter in/out and jar samples — replaces buying a turbidimeter).

8-week timeline

Check the 2026–27 school calendar: if the autumn break falls in week 1, the teacher sows the cassettes. Nov 30 and Dec 1 are public holidays.

gantt
  title EcoColumn build and first experiments
  dateFormat YYYY-MM-DD
  axisFormat %d %b
  section Build
  Collect parts, soil mix, sow cassettes      :b1, 2026-10-26, 5d
  Rack, cassettes, platform, rain bars        :b2, 2026-11-02, 5d
  Node wiring, ESPHome, HA dashboard          :b3, 2026-11-04, 6d
  Supply line, filters, pond hardware         :b4, 2026-11-09, 5d
  Calibrations and interlock tests            :b5, 2026-11-12, 6d
  Commissioning sign-off                      :milestone, m1, 2026-12-15, 0d
  section Biology
  Grass establishment                         :g1, 2026-10-27, 21d
  Pond fill and plants                        :g2, 2026-11-03, 7d
  Snails and inoculum, pond baseline          :g3, 2026-11-10, 28d
  Jar dose bioassay                           :g4, 2026-11-18, 21d
  Fertilised runoff into pond                 :g5, 2026-12-07, 12d
  section Experiments
  Bare vs bare calibration events             :e1, 2026-11-17, 4d
  H1 vegetated vs bare with swap              :e2, 2026-11-23, 5d
  H1b angle series                            :e3, 2026-12-02, 3d
  H3 buffer strip with fertiliser             :e4, 2026-12-07, 5d
  section Engineering
  Rain bar uniformity v0 to v1                :n1, 2026-11-12, 5d
  Filter v1 build and tests                   :n2, 2026-11-23, 9d
  Filter v2 iteration                         :n3, 2026-12-07, 5d
  section Report
  Analysis and presentation                   :r1, 2026-12-14, 5d

3. Toxic Ground — Brownfield

Node: mc-03-brownfield · Team: 3–4 students · Enclosure: TANK63 (60×30×35 cm) used as a 5-zone landscape + a cup "assay bench" for fast, truly replicated tests.

Safety baseline (repeat of HARD SAFETY RULES 1–5): contamination is simulated only — table salt (NaCl), household vinegar (acetic acid), baking soda / garden lime, compaction, a concrete tile and a 12 V heat pad. No heavy metals, no motor oil, no fuel, no fertiliser "spikes", no soil dug from roadsides or old industrial sites, ever. Everything inside or under the tank is ≤ 12 V DC; the only mains item is the certified PSU12 brick on the shelf above, on the RCD strip.

Concept (refined)

A 60 × 30 cm "city block" in a glass tank, divided by siliconed acrylic walls into five isolated soil cells, each one a typical urban damage type:

The tank is the system-scale demonstration (one plot per treatment). The assay bench (48 × 400 mL clear cups) is where the real dose-response science and fast engineering screening happen, because cups can be truly replicated and cress gives a result in 7 days. Phase 1 (Nov–Dec): degrade, measure, understand. Phase 2 (Jan–May): restore zones A–D with the interventions that won on the bench, and track recovery against E.

Science track

Driving question: How do salt, pH change, heat and compaction — the typical damage of city soil — affect whether plants can germinate and grow, and which damage is worst?

Hypotheses (students write their own; these are the defaults):

Variables

Plot layout and replication

Every tank zone (11 × 28.8 cm inside) has the same three bands, front to back:

• Front band (y = 0–9 cm): cress — 3 rows × 5 seeds, rows 2.5 cm apart, seeds 2 cm apart, 5 mm deep (15 seeds/zone).
• Middle band (y = 9–19 cm): measurement band — SOIL probe, DS18B20 (A and E), extract cores, infiltration ring, penetrometer. In zone A this band is the tile.
• Back band (y = 19–29 cm): radish — 3 rows × 3 seeds, 4 cm apart (9 seeds/zone). Re-sown with barley / clover / orache in Phase 2.

Pseudo-replication warning (teach it explicitly): the 3 rows inside one zone are sub-samples of one treated plot, not independent replicates. Tank results show patterns; the bench (3 independent cups per level) is what supports a statistical claim.

Assay bench (true replication). Clear PP cups 400 mL, no drainage holes, 250 g dry soil mix each, brought to 60 % of water-holding capacity by weight, 10 seeds per cup, cups covered with cling film until germination, labelled with code (not treatment) so counting is blind.

Dosing calculations (students do these): salt dose per cup = 250 g × 3 g/kg = 0.75 g NaCl, dissolved in the water used to wet the cup. Tank zone: weigh the dry mix for that zone (≈ 3.2 L ≈ 2.2–2.6 kg), dose = mass × 3 g/kg (≈ 7 g). Expected EC1:5 increase: 1 g NaCl per L ≈ 2 000 µS/cm, and 1 g/kg in a 1:5 extract is 0.2 g/L → ≈ +400 µS/cm per g/kg (so 1 / 3 / 6 g/kg ≈ +400 / +1 200 / +2 400 µS/cm above the baseline). Students predict this before measuring, then compare.

pH shift: titrate a 100 g sample first — add 9 % vinegar diluted 1:10 in 5 mL steps, measure pH1:5 after each step, plot the curve, then scale to the zone mass. For +1.5 use garden lime (CaCO₃, slow, days) or baking soda (NaHCO₃, fast) the same way. Re-check pH weekly and re-dose zone D if it drifts back more than 0.5 units (record every dose).

Sampling schedule

Manual measurement protocols

1. Germination % = seeds with visible radicle ≥ 2 mm / seeds sown × 100, per row (tank) or per cup (bench).
2. Height: soil surface to tip of the highest leaf, 5 tallest plants per row; leaf count = true leaves only (cotyledons separately).
3. Soil EC and pH, 1:5 extract: take a core from the middle band with a cut-off 60 mL syringe (top 5 cm); air-dry 48 h on paper; crush, sieve 2 mm; weigh 10.0 g into a 100 mL screw-top jar; add 50 mL distilled water; shake 1 min, stand 30 min, shake again 30 s, let settle 30 min; dip TDSPEN (EC mode, auto-temperature-compensated to 25 °C) into the clear liquid, read when stable (30 s); rinse in distilled water; then PHPEN in the same extract. Record water temperature. Blank: distilled water alone must read < 10 µS/cm.
   • EC notes: report µS/cm, not "ppm". If a pen only shows TDS, convert with the pen's factor (NaCl scale: EC ≈ TDS × 2; "442" scale: EC ≈ TDS × 1.4 — check the manual). EC1:5 is a relative index: to estimate the saturated-paste ECe used in agronomy multiply by ≈ 8–14 for sandy mixes (texture-dependent; our peat-sand mix is uncertain, so we compare zones, not absolute agronomic classes). Rule of thumb: ECe > 4 dS/m = saline for most crops.
   • Calibrate TDSPEN with CALSOL EC 1413 µS/cm and PHPEN at pH 7 then 4 before every session; log the calibration.
4. Infiltration: push a 5 cm long ring of 50 mm PVC pipe 2 cm into the soil; pour 50 mL distilled water in one go; stopwatch from pouring until no water glistens on the surface. Report seconds; repeat once 10 min later (second value = near-saturated rate).
5. Penetrometer (static pencil test): an unsharpened HB pencil (flat 7 mm end) held vertically in a 20 mm PVC tube guide; place a 1.0 kg load (1 L water bottle) on top for 10 s; measure sinking depth in mm with a ruler. 3 readings per zone in the middle band, ≥ 3 cm apart. Calibrate the idea first on loose vs. tamped cups.
6. Bulk density: 60 mL syringe core (volume known), oven-dry at 105 °C (teacher, kitchen oven) or air-dry 72 h, weigh: ρ = dry mass / volume (g/cm³).

Analysis

• Bench: mean ± range of germination % per level (n = 3); plot germination and shoot mass vs. measured EC1:5 (not just nominal dose); estimate the EC at which germination is halved by linear interpolation; compare cress vs. barley vs. orache curves (H5).
• Tank: time series in Grafana of ΔT = mc_03_soil_temp_paved − mc_03_soil_temp_ref and of moisture per zone; bar charts of row means per zone; compare each zone with E as % of control.
• Simple statistics for liceu students: difference of means with range overlap; optional t-test / one-way ANOVA in a spreadsheet on bench data only.
• Ranking: "Which damage is worst?" — % of control for germination, final biomass, infiltration, penetration. Students must argue with data and name the confounders.

What shows by mid-December vs. spring

Engineering track

Problem statement: Our city block has four damaged plots (sealed + hot, compacted, salty, acidified). Design, build and test low-cost interventions that restore each plot so that plants grow at least 80 % as well as in the control soil, using only materials a school or a city could afford.

Requirements

Candidate interventions

Test criteria: each intervention is compared against (a) untreated damaged cups/zone and (b) control, with 3 cups per treatment; pass = meets R1 + its specific requirement. Cost is computed per m² from the actual grams used.

Iteration plan

1. R1 (16–23 Nov): screen 5 treatments on 3 g/kg salty soil: untreated, leach 2 PV, gypsum + leach, compost 20 %, biochar 5 % (15 cups, cress, 7 days + EC).
2. R2 (23–30 Nov): keep the best 2, vary dose (e.g. leach 1 vs. 3 PV; compost 10 vs. 30 %) and combine (gypsum + compost + leach).
3. R3 (30 Nov–7 Dec): test the final recipe on 6 g/kg soil and on compacted soil.
4. Tank (Jan–Mar): apply final recipes to zones A–D (one at a time, logged), re-sow, compare with E; second paved-zone design (permeable paving prototype v2) after first infiltration results.
5. Each iteration: design sheet (sketch, prediction, cost), build, test, one-paragraph "what we change and why".

Enclosure & build

Recommendation: TANK63 (60×30×35 cm glass, ~5 mm walls, inside ≈ 58.8×28.8 cm) used upright as a landscape: wide footprint, shallow 12 cm substrate, 23 cm of air above for plants, lighting and fan. Glass gives a side view of soil layers and roots. (The 60×40 BOXPLASTIC tray alternative is cheaper but opaque sides hide the soil profile and the walls flex when dividers are siliconed; use it only if TANK63 is unavailable.)

Internal layout (left → right, viewed from the front)

Profile of each zone (bottom → top): 2 cm LECA drainage → geotextile → 10 cm soil mix → (zone A: 1 cm sand + tile). Each zone has a sump tube (16 mm PVC, 15 cm long, slotted bottom 2 cm, in the back-left corner down to the LECA) — a 60 mL syringe with silicone tubing pulls out excess water / leachate for measurement. Dividers stand 4 cm above the soil so watering cannot splash salt across zones.

Lid & light: two LEDPROFILE bars (cut to 58 cm, each with 55 cm of LEDSTRIP) screwed to two acrylic cross-bars (5 × 30 cm) resting on the top rim → LEDs 23 cm above the soil, at y ≈ 9 and y ≈ 20 cm. Rear half of the top covered by a 60 × 15 cm acrylic strip with a FAN cut-out (Ø 76 mm) above zone A and MESH-covered intake slots above zone E; front half open for access.

Under the tank: whole tank stands on a 20 mm XPS board (cushion + insulation); a 10×20 cm pocket is cut in the XPS under zone A for the HEATPAD, adhesive side up against the glass; KSD301 45 °C NC thermostat and 72 °C one-shot thermal fuse taped (aluminium tape) to the pad, wired in series in the pad's + lead. A filled tank weighs ≈ 45–50 kg: build it in its final place on a sturdy table.

Node box mounting: BOX on the wall/shelf bracket at the right end, box bottom above the tank rim; all cables enter from below via GLANDs with drip loops. PSU12 on the shelf above, on the RCD strip.

flowchart LR
  subgraph TANK["TANK63 top view - front at bottom - LEDs 23 cm above soil"]
    direction LR
    A["Zone A Paved plus heat - tile 10x10 cm, HEATPAD under glass"]
    X["XPS buffer 2 cm"]
    B["Zone B Compacted"]
    C["Zone C Saline 3 g NaCl per kg"]
    D["Zone D Acidified pH minus 1.5"]
    E["Zone E Control"]
    A --- X --- B --- C --- D --- E
  end
  S1(["SOIL A plus DS18B20 paved at 2 cm under tile"]) -.-> A
  S2(["SOIL B"]) -.-> B
  S3(["SOIL C plus SHT31 on back glass plus BH1750 at canopy"]) -.-> C
  S5(["DS18B20 ref plus SOIL D in control"]) -.-> E
  F(["FAN 80 mm in lid above A"]) -.-> A
  N["Node box mc-03-brownfield outside right end"]
  E --- N

Soil probe allocation: SOIL A (paved), B (compacted), C (saline), D-channel probe placed in zone E (control). Zone D (acidified) has no continuous probe (only 4 ADC1 soil channels in the pin map): it is loose like E, so it is watered with the same volume as E, and its variable is pH, checked by extract. (Entity names follow the zone: mc_03_soil_moisture_paved, _compacted, _saline, _control.)

Build steps

Total ≈ 17 h of hands-on time + curing.

Biology

Indicator plants (all non-toxic, fast):

Seed sourcing (Bucharest): cress, radish, beet and orache seed packets (Agrosem, Florian, Opal, Agrosel brands, 3–7 RON/packet) at Dedeman, Hornbach, Leroy Merlin and agro shops (e.g. Piața Obor seed stalls); white clover in 100 g–1 kg lawn-seed bags (Dedeman/agro shops); barley as untreated feed grain from agro shops / pet shops ("orz", check it is not fungicide-treated seed — no coloured coating); Salicornia from online seed sellers (eMAG marketplace / EU shops). Use untreated seed only (no coloured fungicide coating).

Soil mix (one batch for everything, ~35 L): universal potting soil ("pământ universal", peat-based) : washed fine quartz sand = 1 : 1 by volume, mixed dry in a tub, sieved through 1 cm mesh. Measure baseline EC1:5 and pH1:5 of the batch (commercial potting soil often has 200–800 µS/cm of fertiliser salts and pH 5.5–6.5 — this is the "0" level). Do not use schoolyard or roadside soil (unknown contaminants).

Inoculants: optional mycorrhizal inoculant (garden-shop "micorize" product) dusted into the radish/clover rows in Phase 2 as an extra factor; clover nodules (pink = active) are a bonus indicator — if seed is not pre-inoculated, expect few nodules and record that.

Earthworms (optional bio-indicator, invertebrate): Eisenia fetida (red wigglers, "râme roșii", fishing shops, ~15 RON/cup). Use only in a separate avoidance test: a 1 L box split in half, control soil vs. test soil (salty / acid / restored), 10 worms placed on the line, cover dark, count on each side after 48 h; worms then return to the Orchestra of Rot compost. Never put worms in the salty tank zones.

Establishment steps

1. Mix and moisten batch to ~60 % of water-holding capacity (squeeze test: holds shape, no dripping).
2. Fill zones; apply treatments (salt dissolved in the wetting water, mixed in a bucket per zone before filling — homogeneous; vinegar the same way).
3. Let zones equilibrate 3 days, then measure baseline EC/pH/infiltration/penetrometer.
4. Sow with a template (cardboard strip with holes at 2 cm / 4 cm), 5 mm deep, mist with distilled water, cover the tank front half with cling film until day 3.
5. Remove film when > 50 % of control seeds are up; begin height counts.

Instrumentation

Pin assignment (standard pin map only)

MOSFET modules: use logic-level D4184 type (IRF520 boards do not fully switch at 3.3 V gate drive). Soil probes powered from 3.3 V so their output stays inside the ADC range; ADC attenuation 11 dB (written 12db in current ESPHome).

Power budget (PSU12 12 V 5 A = 60 W): LEDs 1.1 m × 14.4 W/m ≈ 16 W, HEATPAD 20 W, FAN 1.5 W, ESP32 + sensors via BUCK ≈ 2 W → ≈ 40 W peak (67 % of rating).

Soil probe calibration (per probe, per zone, in that zone's own soil):

1. Take ~400 mL of the zone's treated mix (salty mix for C, compacted core for B) in a cup.
2. Air reading (probe in air) and dry reading (air-dry mix) → 0 %.
3. Saturated reading (add distilled water until free water appears, wait 30 min) → 100 % of saturation.
4. A gravimetric mid-point: weigh cup, compute water content, record raw ADC → check linearity.
5. Enter dry/wet raw voltages in ESPHome calibrate_linear for that channel; re-check in week 6.

Why per zone, including salty soil: a resistive fork probe measures the electrical current through the soil, which depends on dissolved ions as much as on water — adding salt makes the soil read "wet" even when it is dry, and the DC current corrodes the forks. A capacitive probe measures the soil's dielectric permittivity (water ≈ 80, mineral/organic matter ≈ 3–5, air = 1), which depends mostly on water content, so salt affects it much less. But the v2.0 probe works at a relatively low frequency (~1.5 MHz), where high ionic conductivity still raises the reading somewhat, and peat vs. compacted sand changes the dry/wet endpoints. So zone C may read a few % "wetter" than reality, and B differs from E — without per-zone calibration a moisture difference could be a salt or density artefact. Students can test this directly: same water content, 0 vs. 6 g/kg cup, compare raw readings.

ESPHome excerpt (heater loop; full YAML in the repo):

one_wire:
  - platform: gpio
    pin: GPIO4
sensor:
  - platform: dallas_temp
    id: t_paved
    name: "mc_03_soil_temp_paved"
    address: 0x0000000000000000   # fill in from log
  - platform: dallas_temp
    id: t_ref
    name: "mc_03_soil_temp_ref"
    address: 0x0000000000000000
output:
  - platform: slow_pwm
    id: heat_out
    pin: GPIO19
    period: 10s
globals:
  - id: heat_lockout
    type: bool
    restore_value: true
    initial_value: "false"
number:
  - platform: template
    id: delta_t
    name: "mc_03_heat_delta_setpoint"
    min_value: 0
    max_value: 8
    step: 0.5
    initial_value: 6
    optimistic: true
    restore_value: true
switch:
  - platform: template
    id: heat_enable
    name: "mc_03_heat_enable"
    optimistic: true
    restore_mode: RESTORE_DEFAULT_OFF
button:
  - platform: template
    name: "mc_03_heat_reset"
    on_press:
      - lambda: 'if (id(t_paved).state < 35) id(heat_lockout) = false;'
interval:
  - interval: 30s
    then:
      - lambda: |-
          float tp = id(t_paved).state, tr = id(t_ref).state;
          bool bad = isnan(tp) || isnan(tr) || tp < -50 || tp > 84;
          if (!bad && tp >= 42.0f) id(heat_lockout) = true;
          if (bad || id(heat_lockout) || !id(heat_enable).state) { id(heat_out).set_level(0); return; }
          float target = std::min(tr + (float) id(delta_t).state, 38.0f);
          float duty = (target - tp) * 0.25f;          // 25 % per degC error
          duty = std::max(0.0f, std::min(duty, 0.70f)); // duty limit 70 %
          id(heat_out).set_level(duty);

Automation

Heat-island control (zone A). Goal: soil under the tile at 2 cm = reference (zone E) + 6 °C during the photoperiod. Layers of protection: (1) software target capped at 38 °C, P-control with duty limit 70 %; (2) software lockout at 42 °C paved-soil temperature or any sensor fault (needs a manual reset in HA, only accepted below 35 °C); (3) KSD301 45 °C NC bimetal thermostat on the pad (hardware, self-resetting) — the pad/glass never exceeds ~45 °C; (4) 72 °C one-shot thermal fuse in series as last resort. HA automation turns mc_03_heat_enable on at 07:30 and off at 21:00, and sends an alert on lockout. The heater is off during school holidays.

stateDiagram-v2
  [*] --> Off
  Off --> Regulating: heat_enable on and both probes valid
  Regulating --> Regulating: every 30 s duty = 0.25 x error, max 0.70
  Regulating --> Off: heat_enable off at 21.00
  Regulating --> Lockout: paved temp 42 or more, or probe fault
  Off --> Lockout: probe reads 42 or more
  Lockout --> Off: manual reset in HA and paved temp below 35
  note right of Regulating
    target = ref + 6, capped at 38
    KSD301 opens at 45 on the pad
  end note

Fan (GPIO17): air exchange 5 min every hour during the photoperiod; additionally ON at 100 % if mc_03_air_humidity > 85 % or mc_03_air_temp > 30 °C, OFF when < 80 % and < 28 °C (prevents mould on the cling-film days and condensation above the warm zone). PWM 25 kHz, minimum 40 % (below that 80 mm fans stall).

LED (GPIO16): 14 h/day, 07:00–21:00, ESPHome light with 15 min sunrise/sunset ramps, 80 % brightness default (adjust so BH1750 reads the same lux in every run; log the setting).

Irrigation stays manual, with measured volumes — why:

• Water volume is part of the experiment: it dilutes or concentrates salt and leaches it to the sump; every millilitre must be known for the salt mass balance.
• Five isolated cells would need five pumps/valves; one shared line would carry salt between zones.
• The tank has no drain; an automatic fault would waterlog a cell (anaerobic soil, ruined experiment).
• Students learn more by deciding when and how much to water from sensor data than by watching a pump.

Operations

Watering protocol: distilled water only (Bucharest tap water adds ~300–500 µS/cm of salts). Water a zone when its calibrated moisture < 40 % of saturation; add water with a 60 mL syringe / 100 mL cylinder to bring it to ~60 % (volume from the calibration: mL = (target − current) × zone pore volume; start with 50 mL steps); water the cress and radish bands, never the tile (zone A: pour 1/3 of the dose on the tile edge to simulate runoff). Log zone, date, mL, reason. Same volume rules on the bench, by weight (cup to target mass).

Weekly checklist (Monday, 20 min)

• Photos: phone on fixed tripod mark, 50 cm in front, full tank + one top-down per zone with a printed 1 cm grid card and zone label
• HA dashboard: all 8 sensor entities reporting, heater state, no lockout
• ΔT in Grafana last 7 days (expected ~+6 °C daytime)
• Moisture per zone → water per protocol, log volumes
• Sump tubes: extract any free water, measure EC, record mL, dispose in bin bag
• Check pad lead, glands, no water on the XPS
• Zone D pH check (every 2 weeks), re-dose if drift > 0.5
• Bench: rotate cups, count/measure per schedule

Manual logs (shared spreadsheet, one tab each): sowing & counts; heights & leaves; EC/pH extracts (incl. calibration line); infiltration & penetrometer; watering & leachate; interventions (what, grams, cost); incidents.

End of term (by 18 Dec): harvest cress, dry biomass; radish measured (leave in); disable heater (heat_enable off, HA automation paused); LED 12 h; water all zones to 60 %, cover front with cling film (with 2 cm gap); teacher check once a week in the holiday (HA moisture alert at < 30 %); export data snapshot CSV.

Disposal: salty and vinegar-treated soil goes into a sealed bag in the general waste — never onto school garden beds or lawn (it would really salinise them). Control-zone soil can go to compost.

Safety & risks

Educational outcomes & assessment

Sub-BOM

Cost-cut 2026-10-20: the three optional biology lines (Salicornia, inoculant, worms) are removed; add in spring if the restoration round needs them.

Price basis: Romanian retail estimates incl. VAT, Oct 2026 — electronics optimusdigital.ro / cleste.ro / ardushop.ro; tank & aquaristics emag.ro / pet shops; building and garden items dedeman.ro / hornbach.ro; heat pad & thermal parts AliExpress fallback. Shared items (PHPEN, TDSPEN, CALSOL, Pi hub, router) are in the common BOM and shown here at 0.

8-week timeline (26 Oct – 18 Dec 2026)

Phase 2 (Jan–May 2027): apply recipes to zones A–D, re-sow clover/barley/orache, monthly EC/pH/infiltration, spring repeat of S1 on restored soil, earthworm avoidance tests, final report.

4–5. The Future Is Not What It Used to Be (2026 vs 2056)

Concept (refined)

Two identical 45 L glass tanks get the same substrate batch, seed lot, pot layout, LED strip and lid. They differ only in a climate treatment package written in software:

• Tank A, mc-04-future2026: current Bucharest growing-season conditions, adapted to the classroom.
• Tank B, mc-05-future2056: a 2056 scenario. Air is +3.5 °C warmer. Rain over each 14-day cycle has the same total volume, but it comes as one extreme event, one small shower and then a 10-day drought. CO2 is held at 700–800 ppm while the lights are on.

Both tanks carry the full hardware set: heat pad, CO2 inlet, servo vent flap, pump and rain bars. The two ESPHome YAMLs differ only in substitutions:. Because of this, the treatments can be swapped between tanks in spring (crossover design) to separate the climate effect from any "tank effect".

This is a scenario experiment: heat, rainfall pattern and CO2 change together. Run 1 cannot attribute an effect to any single factor, and students must say so in their reports. A single-factor decomposition run is planned for spring (Run 3, CO2 dosing off).

Climate scenario definition

Honest limits of the CO2 treatment (state this in student reports).

1. Classroom air is not "2026 outdoor air". Occupied classrooms regularly reach 1 000–2 000 ppm (Fisk 2017, Indoor Air 27:1039, review of school ventilation; EN 16798-1 Cat. II wants < 800 ppm above outdoor). During lessons the 2026 tank can therefore sit at 1 200 ppm, above the 2056 target. Mitigations:
   • Put the pair in the least-occupied room (lab, not homeroom).
   • Ventilate windows at every break.
   • Report a treatment-fidelity metric: the share of light-period hours with CO2_2056 − CO2_2026 ≥ 200 ppm.
   • Expected fidelity is ≈ 75–80 %, because about 130 of the 168 hours in a week are unoccupied.
2. The 2056 tank is semi-sealed. Plants deplete CO2 in light, so it is a dosing controller, not a "set and forget". The yeast generator overproduces by ≈ 10–40×, so the servo pinch valve meters it.
3. A single SCD41 per tank samples one point. Mixing relies on heat-pad and LED convection. Check homogeneity during commissioning by moving the sensor to 3 heights.
4. The SCD41 automatic self-calibration (ASC) assumes the sensor sees ~400 ppm weekly. The 2056 tank never does, so ASC is disabled on both sensors. Calibration is manual (see Instrumentation).

Science track

Driving question. If Bucharest's climate in 2056 is hotter, has its rain in fewer and heavier events, and has more CO2, how will common crop plants germinate, grow and use water compared with today?

Hypotheses (directional, testable by mid-December or in spring):

Variables.

• Independent variable: climate package (2026 vs 2056).
• Dependent variables: emergence, height, leaf count, fresh/dry mass, hypocotyl diameter, root:shoot ratio, pot mass loss, drainage, stress score, SOIL %, canopy green area from photos.
• Controlled: tank model, lid, substrate batch and fill mass, seed lot, sowing depth and count, pot positions (same randomised map), light (lux matched), photoperiod, ventilation schedule, total water.
• Logged covariates: room CO2 (proxy = mc-04 right after a breath), room temperature, RH.

Replication.

• Per tank: 15 pots (9.5 × 9.5 × 9 cm) in a 5 × 3 grid.
  • Science: 3 species × 3 replicate pots = 9 pots.
  • Engineering: 6 radish pots (2 control, 2 mulch, 2 hydrogel). The engineering controls are biologically identical to the science radish pots, so radish control n = 5 per tank.
• Pot positions come from one random draw (dice/spreadsheet). The same map is used in both tanks.
• Every Monday all pots move one position along a fixed serpentine rotation, identically in both tanks. This averages out edge, heat-pad and fan gradients.
• Within-tank pots are pseudo-replicates of the treatment: there is only one tank per treatment. The spring crossover (Run 2: mc-04 becomes 2056, mc-05 becomes 2026) gives two independent tank × treatment combinations and shows whether a "tank effect" exists.
• Cress is re-sown every ~10 days (3 cycles before Christmas), which gives temporal replication.

Sampling schedule.

Analysis.

• Treatment fidelity first. Daily mean ΔT (target 3.5 ± 0.5 °C), the light-period CO2 contrast, and lux ratio. Data from days that failed fidelity is flagged, not deleted.
• Paired comparison by position. Pot i in A vs pot i in B (same species, same position): difference ± SD. With n = 3 per species, report effect size and all individual points (dot plot). Welch t-test or Mann–Whitney as an exercise only, with a power caveat.
• Growth curves. Height vs day per pot. RGR = (ln H2 − ln H1)/(t2 − t1). Biomass RGR from day 7 thinnings to day 31 harvest.
• Germination curves. Cumulative % vs time. T50 by linear interpolation.
• Water. Daily mass loss (g/day) per pot. Drainage fraction = drainage ÷ delivered. Water-use efficiency = dry mass ÷ water transpired.
• Sensor curves. Grafana overlays of SOIL A/B vs time with rain events annotated. Drying rate (slope % per day) after each event.
• Crossover (spring). Effect = mean over both runs of (2056 − 2026). Tank effect = mean of (Tank B − Tank A) across runs.

What shows by mid-December vs spring.

• By ~Dec 18:
  • H1 complete for all 3 species (sown Nov 16).
  • H3 and H4 over 2 drought cycles (Nov 23 – Dec 21).
  • 3 cress cycles and 1 wheat clip.
  • Radish science harvest (day 31, Dec 17). Roots will be small (≈ 1–2 cm), which is enough for the dry-mass comparison.
  • First engineering results (mulch/hydrogel moisture curves).
• Spring (Feb–May):
  • Run 2 crossover (Feb 1 – Mar 26): same protocol, tanks swapped.
  • Run 3 decomposition (Apr 12 – May 28): 2056 heat + rain with CO2 dosing off vs full 2056, to estimate the CO2 contribution alone.
  • Engineering iteration 2 (clover living mulch, drip stakes).

Engineering track

Problem statement. "The 2056 climate stresses plants through heat, a cloudburst that drains away, and a 10-day drought. Using only low-cost, low-maintenance interventions a Bucharest gardener could copy, make radish plants in the 2056 tank produce at least 80 % of the 2026 radish dry mass, using no extra water."

Requirements.

1. Water budget unchanged: 200 mL/pot per 14 days.
2. No extra electricity per pot.
3. Cost ≤ 2 RON per pot.
4. Materials non-toxic and available from Dedeman/garden shops.
5. Must not change light at canopy by > 10 % (shade cloth tested separately).
6. Measurable effect on at least one of: pot moisture at end of drought, drainage fraction, dry mass, stress days.