A sovereign foundation model of orbital behaviour. Every object in orbit leaves a public behavioural trace — continuous two-line element records since 1957, covering 60,000+ catalogued objects. ORBIT-FM is a research programme to pretrain a single model on that record, fused with space-weather indices, so that multiple Space Domain Awareness capabilities emerge from one model — trained entirely on UK-owned, on-premises compute.
Evaluation precedes scale — every claim below is backed by a frozen, versioned evidence set.
A self-funded Phase-0 prototype encodes orbital histories as residuals against an SGP4 physics baseline — the model learns precisely what physics misses: manoeuvres, drag error, sensor artefacts. Trained on 335 low-Earth-orbit objects, it was evaluated on strictly held-out months against 574 real manoeuvre events derived from CNES/IDS DORIS precise-orbit records.
Result: F1 0.628 versus 0.568 for the standard classical detector, with equal-or-better recall at every matched precision operating point — up to nine percentage points more. The result held across two model configurations trained independently, one per cluster node. Negative results are reported alongside positive ones.
Read the Phase-0 technical report →From a single pretrained model of orbital behaviour, SDA capabilities emerge with minimal task-specific engineering.
Not a policy aspiration — an engineering property of how the model is built and delivered.
We are engaging an astrodynamics consultant (SGP4 baseline verification, manoeuvre-label validation, orbit-regime stratification) and talking to early pilot partners — SSA platforms, insurers and government SDA users. If that's you, we'd like to hear from you.
team@neuravant.ai