Hi all,
Quick honest one this week.
After Week 6 I had a loose end nagging me. I'd built a 5G-NTN terminal that holds a timing + Doppler correction and refreshes it on a fixed schedule — but a satellite races past the horizon and then crawls overhead, so the correction goes stale much faster at some points than others. Refreshing at a constant rate felt wasteful. The obvious fix: adapt the rate through the pass.
So this week I built the adaptive version properly and raced it against the fixed one, on a real STARLINK-1008 pass over my Aberdeen ground station.
It lost.
The whole Week 7 story in one figure — a real Starlink pass over Aberdeen, showing why "simple and fixed" already wins.
The adaptive schedulers actually used more refreshes than a sensibly-chosen fixed rate. Two reasons, both physical: the satellite never gets high enough here for the geometry to genuinely calm down (the pass is timing-limited 91% of the way through), and in the one spot adaptation could have helped — right overhead — a second constraint (the frequency budget) quietly takes over and clamps it anyway.
I'll be honest, I spent a bit of the week disappointed before I realised the negative result is the deliverable. It tells you plainly: don't bother adding a within-pass adaptive loop for this deployment. Week 6's simpler design was already the right call. That's a genuinely useful thing to be able to say.
The bit that did work: the rate you should pick changes from pass to pass, and it turns out you can predict it from a single thing the terminal already measures — Doppler — via a one-line law. I then trained a small model on 40 passes to see what it would learn, and it rediscovered that exact law on its own (R² = 0.995). I love when that happens — the AI and the hand-derived physics agreeing is worth more than either alone.
Two threads from earlier weeks (the pre-comp work and the ML work) are starting to converge, which is nice to see with the flagship in sight.
Full write-up on Medium: https://medium.com/p/3ed91c41f980
Code, tests and plots on GitHub: github.com/shaifiqbal/ntn-adaptive-precomp
More next week.
Shaif