1. The station that had to guess the ocean’s mood

Stanisław Lem’s Solaris returns as the recalled work because the scientists on Prometheus spend much of the novel trying to build a model of the ocean that can predict its behavior. Their models are always slightly wrong, and the wrongness is costly. A LEO desktop faces a smaller version of the same problem: it must maintain a model of its own thermal state, and the model must be good enough to keep components alive without being so complex that the computer cannot run it.

This entry wonders how the desktop estimates and models its own temperatures.

2. Temperature is local, not global

A spacecraft does not have one temperature. It has many: the battery, the processor, the radiator, the solar array backside, the actuator, the payload. Each has its own sensor, its own thermal mass, and its own view of the environment. A single alarm threshold is not enough. The autonomy needs a map: which nodes are hot, which are cold, which are changing fast, and which are stable.

The first problem is sensor placement. A sensor on the outside of a box does not measure the junction temperature of the chip inside. It measures a point on the case. The autonomy must infer internal temperatures from external measurements using a thermal resistance model. This is the same problem as estimating battery state of charge from voltage: the measurement is a proxy, and the relationship is approximate.

3. The live thermal model

A nodal thermal model is not just a design tool; it can be a runtime tool. The desktop can run a simplified version of the model on orbit, updating node temperatures from sensor readings and propagating them forward. The model predicts what will happen if the Sun angle changes, if a payload turns on, if a heater cycles, or if a radiator goes into shadow.

The key choices are:

  • Granularity: how many nodes to track. Too many nodes waste compute; too few nodes miss critical gradients.
  • Time step: how often to update. Thermal changes are slower than electrical changes, so the model can run at seconds or minutes rather than milliseconds.
  • Inputs: which measured temperatures, currents, attitudes, and environment fluxes to feed into the model.
  • Calibration: how to adjust conduction paths and radiative couplings so the model matches reality.

A useful principle is to keep the runtime model simpler than the design model but to re-correlate it against ground data whenever a major attachment changes.

4. Estimation when sensors fail

Sensors can fail or drift. A thermistor can read incorrectly after radiation exposure. A wiring fault can produce an out-of-range value. The autonomy must detect bad sensors and fall back to the model. If the battery temperature sensor fails, the model can estimate battery temperature from the charge current, the heater state, and the temperatures of nearby nodes. The estimate is worse than a real sensor, but it is better than flying blind.

5. What this changes

  • Thermal state estimation is a distributed inference problem across many sensors and nodes.
  • A runtime nodal model can predict temperatures and support decisions.
  • Sensor failures must be detected and handled by model-based fallback.
  • The next entry will wonder about heat trading between cells.