Model what the physics gets wrong.
Every object in orbit leaves a public behavioural trace: two-line element records since 1957, across 60,000+ catalogued objects. ORBIT-FM pretrains one transformer on the residuals between that record and an SGP4 physics baseline, so it spends all its capacity on what physics cannot explain: manoeuvres, drag mis-modelling, operator behaviour.
Built for Space Domain Awareness users who cannot route data through foreign cloud. Trained on UK-owned hardware, delivered on premises and air-gapped, on the same class of compact NVIDIA machine it was trained on. No operational data, weights or roadmap ever leaves the customer's facility.
Move the threshold. Watch what each detector catches.
Every stem is a real candidate event scored by the frozen Phase-0 prototype (up) and the classical MAD z-score baseline (down). Diamonds on the centre rail are real manoeuvres from CNES/IDS DORIS precise-orbit records. A candidate counts as a hit within ±24 h of an unmatched label, exactly as in the published evaluation.
The prototype beats the classical detector on months it has never seen.
A 2.9M-parameter transformer, trained on 1.1M residual time steps from 335 low-Earth-orbit objects, was evaluated on strictly held-out months against 574 real manoeuvre events for nine DORIS-tracked satellites. Ninety-eight of those events fall in the held-out window.
| Detector | Best F1 | Operating point | Precision | Recall |
|---|---|---|---|---|
| ORBIT-FM Phase-0 prototype (ctx-256) | 0.628 | z = 18 | 0.656 | 0.602 |
| Classical MAD z-score detector | 0.568 | z = 24 | 0.641 | 0.510 |
| ORBIT-FM prototype (ctx-512, second node) | 0.609 | z = 18 | Independent replication; ctx-256 remains primary | |
At matched precision the model returns equal or better recall at every baseline operating point from z = 3 upward, peaking at +9.2 percentage points at the baseline's own F1-optimal point. Among the model-only catches are documented Sentinel-3A, CryoSat-2, Jason-3, HY-2C and HY-2D burns that a fixed statistical threshold does not see at all.
What did not work, reported alongside what did
- Consecutive residuals are near-white: persistence scores 0.911 relative error, so per-step forecasting is hard by construction. Value lies in population-scale breadth, not per-object memorisation.
- Longer context did not help. A 512-step model replicates the result (F1 0.609) but does not beat 256 steps. The scaling path is more objects and longer history, not longer windows.
- Small beats large at this data scale. A 21.7M-parameter model overfit within 400 steps; the 2.9M model is the one that generalises. Data first, parameters second.
- An early pipeline clipped the regression targets at ±20 IQR, flattening the very spikes detection depends on. Found, fixed, and reported.
Don't model the orbit. Model the residual.
SGP4 already describes the deterministic part of an orbit well. What it cannot describe is the interesting part. Each object's history is encoded as residuals against an SGP4 propagation of its own previous element set, in the radial, along-track and cross-track frame, fused with space-weather indices. A single decoder-style transformer is pretrained to predict the next residual step across the whole population.
Because the pretraining task is generic, downstream capabilities need little task-specific engineering: manoeuvre detection is simply the surprise of the model at each step, thresholded and merged into events. The same representation is the starting point for anomaly flagging, behavioural identification and lifetime estimation.
Space Domain Awareness, built once instead of one bespoke system at a time.
Today's SDA stack is a separate hand-built tool for every question, each tuned and maintained on its own. A foundation model over the public orbital record changes the economics: pretrain once, then adapt.
Manoeuvre detection
Flag station-keeping, orbit-raising and collision-avoidance burns from public elements alone. Beats the classical detector on held-out data.
Anomaly and off-nominal flagging
Surprise without a matching burn signature: tumbling, fragmentation precursors, unexpected decay.
Behavioural identification
Characterise operators by their manoeuvre cadence and style. Who is this object behaving like?
Payload-activity characterisation
Distinguish active, dormant and end-of-life behaviour from the trajectory record over months.
Lifetime and re-entry estimation
Decay forecasts that learn the drag error physics leaves behind, especially through solar maximum.
Propagation-error correction
Predict the residual ahead of time and hand operators a corrected state for conjunction screening.
Four product lines from one model.
Manoeuvre and anomaly feed In prototype
Event feed by API or offline batch, from public elements. The Phase-0 detector already runs a weekly unattended pass on live Space-Track data. A first pilot integrates an operating feed, not a promise.
Orbital lifetime and re-entry module
Decay and re-entry estimates for underwriting, regulation and deorbit-compliance reporting, with the drag error learned rather than assumed.
Propagation-correction service
A learned correction to SGP4 propagation, delivered as a corrected state vector for screening and planning.
Sovereign on-premises licence Strategic
The full model, deployed air-gapped on the same class of compact NVIDIA hardware it was trained on. No operational data, weights or capability roadmap leaves the customer's facility.
No foreign cloud in the loop.
The entire programme, from data pipeline to pretraining to the weekly live detector, runs on a two-node NVIDIA DGX Spark cluster that Neuravant owns and operates. Nothing about the result depends on funding arriving or on a provider's terms.
- Nodes
- 2 × NVIDIA DGX Spark, GB10 Grace Blackwell
- Per node
- 20-core ARM64 · 121 GB unified memory
- Interconnect
- Direct 200 GbE ConnectX link
- Precision
- BF16 mixed precision on Blackwell
- Location
- On premises, United Kingdom
- Container
- NGC PyTorch 25.10 · CUDA 13.0
- Framework
- PyTorch 2.12 · NCCL 2.29
- Delivery
- Air-gapped packaging on the same hardware class; TensorRT-LLM and NIM under evaluation
- Data terms
- Space-Track user agreement respected: no basic SSA data redistributed
- Programme
- Neuravant AI · NVIDIA Inception member, August 2026
The detector runs every week against the live catalogue.
Product line 1 is not a roadmap item. The frozen prototype scans the nine labelled satellites each Monday and writes what it found to a continuous record. Over the trailing year it flagged 97 events, matching the labelled base rate.
| Run | Window | Flagged | Events |
|---|---|---|---|
| 2026-09-14 | trailing 8 days | 0 | none above threshold |
| 2026-09-09 | trailing 8 days | 0 | none above threshold |
| 2026-09-01 | trailing 8 days | 1 | 2026-08-26 11:40 UTC · Sentinel-3A (41335) · z = 19 |
| 2026-08-24 | trailing 8 days | 0 | none above threshold |
| 2026-08-05 | trailing 8 days | 0 | none above threshold |
Quiet weeks are the honest norm: the nine reference satellites manoeuvre roughly every few weeks each. The value of the record is that it is continuous and checkable, not that it is busy.
Phase 0 is done and evidenced. Phase 2 is a scale problem.
Phase 0 complete
Representation proven on 335 objects. Held-out F1 0.628 vs 0.568. Evidence frozen at tag phase0-v1.0, technical report published including negative results.
Weekly live detector
Frozen checkpoint runs unattended against current Space-Track data. Continuous log since 5 August.
NVIDIA Inception
Member of the NVIDIA Inception programme for technical resources and go-to-market support.
Public benchmark in preparation
A versioned manoeuvre-detection benchmark: 574 DORIS-derived labels, evaluation harness and reference baseline, with held-out objects and held-out time. Labels and code only; participants fetch elements under their own Space-Track accounts. We define the yardstick, then compete on it.
Pretrain on the full archive
Scale the same representation from 335 objects to the 60,000-object public catalogue, with an astrodynamics consultant engaged for SGP4 baseline verification, label validation and orbit-regime stratification.
Trial the feed. Test the benchmark. Tell us where it breaks.
We are talking to early pilot partners and to researchers who want a shared, versioned yardstick for manoeuvre detection from public elements. We will walk through the evaluation in detail, including its limitations: nine labelled satellites, precision understated by construction, a feed that is a funded-programme deliverable rather than a product on sale today.
team@neuravant.ai