The Research Pivot — Temperature Optimisation

2026-03-04

I ran a structured literature review prompt through Claude — not a casual question, but a systematic search across thermoperiodism, tissue culture temperature management, and optimal control theory in plant science. Claude synthesised findings from 60+ papers in a few hours. A human doing this manually would need weeks in a university library. The result was more interesting than I expected.

The headline finding: no published study has applied optimal control theory to TC temperature management. Every lab in the world defaults to static 25°C +/- 2. The scattered studies that tested dynamic temperature regimes found dramatic improvements, but nobody connected the dots.

The evidence

The literature is thin but the results that exist are striking:

• Bennett et al. (1991) — potato microtuberisation showed a 25% yield increase simply by cycling between day and night temperatures instead of holding static 25°C.
• Abiri et al. (2025) — cannabis tissue culture reduced hyperhydricity (the most common cause of culture loss) from 85% to just 10% by applying a 2°C bottom-cooling differential. Hyperhydricity went from culture-killing to nearly eliminated with a temperature tweak.
• Kvaalen & Johnsen (2008) — Norway spruce showed epigenetic memory lasting 20+ years from the temperature experienced during a single developmental window in embryogenesis. The temperature during one phase of development permanently altered the tree’s cold hardiness decades later.
• The DIF framework — the day/night temperature differential (DIF) is well-studied in greenhouse horticulture for controlling stem elongation. But it has NEVER been tested in vitro, despite the molecular mechanisms (auxin transport, cell expansion) being cell-autonomous and operating at the single-cell level.

For mint specifically, nobody has tested anything beyond static 20 vs 25°C. The gap is wide open.

The mathematical frameworks

The tools for optimal temperature control already exist — they’ve just never been pointed at tissue culture:

• Seginer’s Pontryagin approach — optimal control theory applied to greenhouse climate, proven to outperform static setpoints
• Stochastic MDP — Markov decision processes for sequential environmental decisions under uncertainty
• Reinforcement learning — model-free discovery of optimal trajectories through trial and error

All validated in greenhouse systems. None applied to TC. The leap from greenhouse to growth chamber is technically trivial — smaller space, fewer variables, faster feedback loops. The research gap isn’t a lack of methods; it’s a lack of anyone connecting the control theory literature to the tissue culture literature.

The bigger picture: a citrus paper and a reality check

The same day, I uploaded a paper on gamma irradiation of citrus tissue (Agisimanto et al., 2016 — Sains Malaysiana). The researchers irradiated Citrus reticulata nucellus tissue to induce mutations, then screened regenerated plants with molecular markers. I said to Claude: “I would love to do something like this. This is the kind of research unit we want to build. A lab that Claude controls to perform cutting-edge research that can be scaled.”

That triggered a reframe of the whole project. The mint organogenesis isn’t the product — it’s the training wheels for the system. Every step in the citrus paper’s workflow (design treatment → apply treatment → culture → observe → subculture → screen → analyse) maps to something Claude can direct. The actual goal is an AI-directed plant breeding and research platform.

Then came the honest critique. Claude pushed back hard:

• “You can’t do TC yet.” I’ve never sterilised a single explant. 80-90% contamination is normal for beginners.
• “Claude isn’t actually an agent.” At this point it’s a stateless conversation, not a persistent process watching the lab.
• “The camera doesn’t make me a scientist.” Interpreting callus morphology is subtle. Trained technicians get it wrong.
• “EMS (chemical mutagenesis) is dangerous.” Needs a fume hood, proper PPE, waste disposal. Not a home lab.
• “Molecular screening is a wall.” PCR, gel electrophoresis, primers — £2-5k minimum, or outsource every run.
• “The timeline is brutal.” A novel patentable variety is 3-5 years minimum.
• “The environment isn’t solved.” 22.9% combined suitability. The radiator problem hasn’t been fixed yet.

My response: “Brilliant.”

The decision was to stop pitching the platform and start doing the science — but build the platform infrastructure in parallel so it’s ready when the biology catches up. Two parallel tracks: biology (get mint alive, learn sterile technique) and platform (sensor → AI → actuator → camera, species-agnostic). The critique didn’t weaken the project. It made it credible. A project that knows exactly what it can’t do yet is more trustworthy than one that claims it can do everything.

The actuator: a Shelly Plug S

Simplifying to the first controllable variable — temperature — led to a discussion about hardware. I initially suggested wiring a transistor to the ESP32 for mains control. My project partner Zi had a better idea: “Something that works out of the box, like a switch we can control with an API using the Pi.”

The answer: a Shelly Plug S (UK version, ~£20-25). Why specifically this:

• Local HTTP REST API out of the box — curl http://<shelly-ip>/relay/0?turn=on and curl http://<shelly-ip>/relay/0?turn=off
• No cloud dependency, no hub, no firmware flashing
• Works on local WiFi — Claude on the Pi curls it directly
• The radiator plugs into the Shelly, the Shelly plugs into the wall — Claude IS the thermostat
• Binary on/off is enough for bang-bang temperature control: if temp < 20°C turn on, if temp > 24°C turn off

When I sent Zi the updated research plan PDF, his reaction: “Wow. Wow. Literally wow. Great pdf. Shelly Plug S. YES M9. Has it arrived?”

He also sent a message about ensemble agents: “I wonder if it’s good as a researcher too. Do you know how to ensemble agents? Have them talk to each other. Collaborate. E.g. codex 5.3 + opus working on same problemo. Critiquing, observing, each other.” This maps directly to the two-loop architecture we’d build three days later — inner agent (technician) on the Pi, outer agent (director) on the laptop.

The experiment

First variable we control: temperature. The Shelly Plug S on the radiator gives Claude a complete observe-decide-act loop: read temperature from the DHT22, decide whether to heat, curl the Shelly on or off, observe the result. No cloud dependency, no app, no human in the loop.

We run mint cultures at multiple temperature profiles in parallel:

• Static 25°C — control
• Diurnal cycling 25/18°C — day/night
• Decreasing ramp — across the culture cycle
• Whatever Claude discovers — via reinforcement learning

The Pi camera (next purchase) photographs cultures daily. Claude compares growth across treatments and adjusts the next round.

The question isn’t “can mint grow in TC” — that’s known. The question is “what temperature function optimises mint callus induction and shoot multiplication, discovered by an AI running closed-loop control?” That’s a publishable result.

What this actually is

The mint is the test subject, not the product. We’re building an AI-directed research platform that can discover optimal environmental trajectories for any species in tissue culture. Temperature is the first variable because we can control it with a £20 plug. Light, humidity, and CO2 follow. The infrastructure we’re building (sensor → AI → actuator → camera → AI) is species-agnostic. Once it works for mint temperature, the same system runs lavender, aroids, or anything else.

The principle

We add automation and complexity only when human intervention becomes a limiting factor. The research comes first. The infrastructure serves the science, not the other way around.

If Phase 0 fails because of contamination in the still air box, the answer is better aseptic technique — not a robot arm. If Phase 1 fails because the PGR concentrations are wrong, the answer is another round of experiments designed by Claude — not a bioreactor. We solve problems as they appear, with the simplest tool that works.

What success looks like

Near term (month 1-2): Callus growing on plates. Claude monitoring environment 24/7. Me checking the lab dashboard on my phone and doing physical work when Claude says it’s time. A working rhythm established.

Medium term (month 2-4): Somatic embryos forming. Claude has run 2-3 rounds of protocol optimisation, each informed by the last. An optimised PGR combination for our conditions and species. Embryos converting to plantlets.

Long term (month 4+): Experiments running passively. Liquid cultures in flasks, medium changes every 1-2 weeks. Claude sends a summary to your phone: everything’s fine, embryos at cotyledonary stage, next subculture Wednesday. The flat lab is producing plants. And when we’re ready to scale, the entire system — hardware, software, protocols, Claude’s knowledge base — lifts and shifts to a real space.

Immediate next steps

1. Order Shelly Plug S + Pi camera module
2. Wire relay to ESP32 for heat mat backup
3. Fix room environment — turn radiator TRV down, verify with 48h dataset
4. Pour first MS + BAP media batch
5. Establish mint explants — spring growth just emerging, ideal timing
6. Deploy camera pipeline on Pi so Claude has eyes

Biology and platform build in parallel. The system is ready to run the first real experiment as soon as cultures are alive.