Understanding algae.
From cultures
to single cells.

The loop, said slowly

The planned inputs are pH, temperature, wall irradiance, optical density at 750 nm, dissolved oxygen and conductivity. We want to use them to estimate biomass, internal nitrogen quota and dissolved inorganic carbon. Droop quota kinetics, Steele light response and Beer-Lambert attenuation are candidate model components; fitting them to real culture data is work still ahead.

The forecast target is seventy-two hours. A local rule-based planner would compare possible changes to carbon supply, nutrients, airflow and light. An optional assistant could explain proposals or accept operator requests, but would not have direct actuator authority. We need to test whether the model remains useful over that forecast horizon.

The proposed safety structure has three layers: fixed limits on mass, duty cycle, pH, temperature and dose frequency; a model check on the proposed action; and independent limits on the ESP32 actuator controller. A watchdog and physical emergency stop are also design requirements. We cannot call these protections reliable until fault testing demonstrates them.

What the TPU is for

Optical density measures light attenuation, not organism identity. A camera could provide a second evidence stream by looking for coils, fragments and unfamiliar shapes. We are considering a quantized vision model on a Coral accelerator, with training and compilation performed separately. A useful classifier needs labeled images and tests against bubbles, debris and organisms it has not seen before.

We have not established inference speed, power draw or contamination-detection accuracy for this proposed setup. Vision should be able to request caution, not authorize a pump. That separation is a design requirement, not an experimental result.

The edge brain

The proposed controller uses a Raspberry Pi 5 for telemetry, the growth model and local policy, with a Coral USB Accelerator for camera inference. A local Qwen assistant is an optional interface idea, not a dependency for control. Resource use, thermal behavior and safe fallback operation all need testing on the assembled hardware.

What the mesh is for

Algaephyte Mesh is a proposed way to share fitted model parameters rather than photos or raw traces. Candidate parameters include μ_max, I_opt, K_s, T_opt and Q_min. MQTT is the planned transport. Whether parameters transfer usefully between vessels and strains must be tested; we are not claiming that a network already exists or speeds up cultivation.

Droop: the quota inside the cell

Droop's cell-quota kinetics decouple uptake from growth: cells store nitrogen, and division depends on that internal quota rather than the concentration in the medium. That single idea explains why a starved culture keeps dividing after you feed it, and why over-dosing nitrate buys you nothing but bacteria.

Steele: light has an optimum, then a cliff

Steele's curve handles light — growth rises to an optimum irradiance and falls again under photoinhibition. Brightest is not best: past the optimum, the outer shell of the culture bleaches, oxygen supersaturates, and photosynthesis starts to poison itself. The twin treats irradiance as a control surface, not a schedule.

Beer-Lambert: a dense culture is a stack of dark jackets

Beer-Lambert gives every radial shell of the vessel its own light climate, so the outer millimetre of a dense culture can be photoinhibited while the axis sits below compensation. A single wall sensor cannot see this. The twin can, because it integrates Steele's curve through the radial shells.

Alkalinity, not pH

Most growers chase pH because pH is the number a ten-dollar probe will give you. In an alkaline Arthrospira medium the bicarbonate/carbonate pool is simultaneously the inorganic carbon supply, the overnight buffer, and the reason almost nothing else can live in the vessel. pH is what that pool looks like from the outside. The twin treats dissolved inorganic carbon as a state variable, not a setpoint.

Dissolved oxygen: when photosynthesis poisons itself

As dissolved oxygen climbs through the afternoon, photosynthesis starts to supersaturate — the culture poisons itself. The first proposal is usually more air. The second is a modest dim. The third, if you have been stubborn about air, is a siesta. Oxygen stress is a state variable in the twin, not a footnote.

A request such as "raise alkalinity to target" leaves out essential information: the mass, rate, current chemistry and equipment limits. Our design must reject incomplete requests rather than guess. We plan to test malformed proposals, stale sensor values, lost communications and stuck outputs before trusting automatic dosing.

The transport design uses MQTT QoS 1, authenticated connections and TLS for remote traffic. Small payloads would suit limited connections. A last-will message can indicate a lost connection, but cannot establish whether a culture is healthy or a pump is working.

The proposed sharing concerns the growth model, not on-device camera training. We need to establish how to handle outliers, incompatible strains and malicious inputs. Local cultivation should not depend on the broker; disconnect tests will need to confirm that.