Space: SSAO and Knowledge Graphs for Orbit
Space is the domain where the numbers are excellent and the meaning is missing. Orbital mechanics is a solved science: given a state vector and an epoch you can propagate a trajectory to useful precision. What nobody solved is the semantics around it. Which object is this, really? Is it the same object it was before the fragmentation event? Who operates it now, under which licensing jurisdiction, and does its disposal plan satisfy the rule that applies to it? Those are ontology questions sitting on top of a numerical catalogue, and they are why space situational awareness keeps rediscovering knowledge graphs.
What the orbital data layer actually is
Three artefacts carry most of the world’s operational space data. The first is the public US catalogue distributed through Space-Track, operated by the 18th Space Defense Squadron of the US Space Force, which publishes two-line element sets and the newer general perturbations format keyed to each object’s catalogue number, along with conjunction warnings, launch data and reentry predictions. The second is the CCSDS Orbit Data Messages family, standard 502.0-B, which defines the Orbit Parameter Message, the Orbit Mean-Elements Message and the Orbit Ephemeris Message; the current recommended standard edition was issued in April 2023. These are the interchange formats that let one agency’s ephemeris be read by another’s software. The third is ESA’s DISCOS, the Database and Information System Characterising Objects in Space, a single-source reference covering launch details, orbit histories, physical properties and mission descriptions for tens of thousands of objects tracked since Sputnik 1, with millions of orbit records behind it and ESA’s annual Space Environment Report as its public output.
Now the limits. A catalogue number identifies a tracked object, not a satellite as an operated asset, and the mapping between them is neither stable nor one to one. International designators encode launch, not mission. When a rocket body fragments, the pieces get new catalogue entries and the relationship to the parent object survives, if at all, in prose. Ownership transfers, mission changes, re-registrations and end-of-life status are largely outside the format. Provenance is absent: a message tells you a state vector, not who computed it, from which observations, with what uncertainty policy. Every one of these gaps is exactly what an ontology and a graph are for.
The ontologies: SSAO and the Space Object Ontology
The Space Situational Awareness Ontology, usually SSAO, is Robert Rovetto’s ontology of the SSA domain: a defined terminology for the objects, observations and activities involved in awareness of the near-Earth and orbital environment. Its stated goals are data sharing and fusion across space object catalogues and SSA databases, automated reasoning over that fused data, and semantic interoperability between systems. It is developed and published openly on GitHub as a modular part of a broader suite of space domain ontologies, and it is realistically a scholarly, individually driven artefact rather than an agency-blessed standard. Treat it as a well-argued starting vocabulary, not as something a mission assurance board has signed off.
The Space Object Ontology is the other main line. Presented by Cox, Nebelecky, Rudnicki, Tagliaferri, Crassidis and Smith at the 19th International Conference on Information Fusion in 2016, it characterises space objects by identifiers, design specifications, components and subsystems, capabilities and vulnerabilities, origins, missions, orbital elements, patterns of life, processes, operational statuses and associated persons, organisations or nations. Architecturally it is orthodox realist practice: BFO at the top, the Common Core Ontologies for mid-level structure, domain classes beneath. That matters commercially, because it is the same stack the US Department of Defense and Intelligence Community adopted as baseline in 2024. If you are building for a US defence or intelligence space customer, a BFO plus CCO grounded space model is the conformant answer.
The honest state of play: unlike defence with IES, or biology with the Gene Ontology, space has no single dominant, governed, production ontology. There are credible academic ontologies, agency databases with proprietary internal models, and a large amount of tabular data. That is an opportunity rather than a defect, and it is why publishable crosswalk work in this sector is disproportionately valuable.
What breaks · the orbital element set treated as the object
A catalogue ingest models each two-line element set as the satellite itself: one node per TLE, carrying inclination, eccentricity and epoch as properties of the spacecraft. It reads naturally, and every conjunction screening built on it is subtly wrong.
An element set is a state estimate valid at an epoch, not the object: the same spacecraft generates thousands of them, and a manoeuvre produces a genuinely different orbit for the same physical body. Under this modelling, either each new element set overwrites the last, and the manoeuvre history that conjunction analysis depends on is destroyed, or each becomes its own object, and the catalogue reports one satellite as hundreds of separate space objects. This is lesson 3’s identity-through-change problem at 7 km/s: the fix is the same one the register used, an enduring object with time-bounded states, so that “where was it on this date” and “what is it now” are both answerable and neither overwrites the other.
Turning debris rules into graph constraints
The rules a space operator must satisfy are the perfect test case for putting policy on a graph, because they are conditional, jurisdictional and time-bound. The Inter-Agency Space Debris Coordination Committee guidelines define protected regions, low Earth orbit up to 2,000 km altitude and the geostationary region, and set the disposal expectations that shaped the well-known twenty-five-year post-mission lifetime limit for LEO. Those guidelines are voluntary and were the model for the UN COPUOS space debris mitigation guidelines of 2007. ISO 24113, Space systems: space debris mitigation requirements, converts the principles into engineering requirements for spacecraft designers and operators; first published in 2011, it was most recently updated in 2023. Regulators then diverge: the US Federal Communications Commission requires deorbit of licensed LEO satellites within five years of mission completion, a sharp tightening of the earlier twenty-five-year expectation. Policy pressure continues through ESA’s Zero Debris Charter, unveiled in 2023 and signed by twelve countries in May 2024, with New Zealand and Mexico the first signatories from outside Europe later that year, committing to debris-neutral activity by 2030.
Encode that in a spreadsheet and you get a compliance matrix nobody trusts. Encode it in a graph and the interesting queries become one hop: which of our missions operate in a protected region, which regulator’s rule applies given the licensing state, which disposal plans satisfy that rule, and which claims rest on an ephemeris older than its validity window. The constraint layer is a SHACL job, not an OWL job: “every mission in the protected LEO region must have a disposal plan with a stated post-mission lifetime and a source” is a closed-world requirement, and OWL will happily stay silent about the missing plan while SHACL fails the build.
| Artefact | What it is and does | Maintainer | Licence and access | Real deployment and limits |
| Space-Track catalogue | Public element sets, general perturbations data, conjunction warnings and reentry predictions keyed to catalogue numbers | 18th Space Defense Squadron, US Space Force | Free with registered account and terms of use | Operational conjunction assessment worldwide; carries no mission, ownership or provenance semantics |
| CCSDS Orbit Data Messages, 502.0-B | Standard interchange formats: Orbit Parameter, Orbit Mean-Elements and Orbit Ephemeris Messages | Consultative Committee for Space Data Systems | Recommended standards published free by CCSDS; current edition April 2023 | Agency to agency ephemeris exchange, and the basis of civil space traffic coordination message specifications |
| ESA DISCOS | Reference database of launch details, orbit histories, physical properties and mission descriptions for tracked objects since 1957 | European Space Agency Space Debris Office | Access via ESA services; annual Space Environment Report published openly | Underpins ESA debris environment modelling and public reporting |
| SSAO and Space Object Ontology | Formal vocabularies for space objects, observations, missions, capabilities and operational status | SSAO by Robert Rovetto; Space Object Ontology from the Buffalo and CUBRC community | Open repositories and published papers; no agency mandate | Space Object Ontology follows the BFO plus CCO stack adopted as US defence and intelligence baseline in 2024 |
| IADC guidelines and ISO 24113 | Protected region definitions and disposal expectations; the engineering requirements standard for debris mitigation | IADC member agencies; ISO for 24113 | IADC guidelines published openly and voluntary; ISO 24113 text paid, updated 2023 | Basis of national licensing rules, including the FCC five-year deorbit requirement for US-licensed LEO satellites |
Check yourself
A conjunction analysis pipeline must answer “what was this object’s orbit at the time of the 12 March close approach?” Which modelling supports it?
- The object is an enduring individual; each element set is a time-bounded state estimate with an epoch and a provenance link to its source.
- The object node carries current orbital elements, overwritten whenever a new element set is published.
- Each published element set is its own space object, identified by its epoch.
- Orbital elements are stored as free text in a description property for analyst reading.
Show the answer, and why each wrong option is wrong
A is correct. The question names a time, so the model must carry time explicitly: states with epochs, attached to one persistent object, plus provenance because element sets from different sources disagree and the analyst needs to know which one the answer used.
B answers only “now”, and 12 March is not now; the historical orbit was destroyed by the overwrite, which is the register’s lesson 3 failure in a domain where the consequence is a missed conjunction. C preserves history and destroys identity: the catalogue can no longer say how many objects are in orbit, and every cross-source match has to reconstruct which thousands of “objects” are one spacecraft. D makes the data unqueryable by construction; text is not a model, and no screening algorithm can filter on a paragraph.
Neurosymbolic checking of orbital claims
Space is an unusually good showcase for the verifier loop that runs through this course. An analyst asks a natural language question about an object, a manoeuvre or a compliance status. A language model is excellent at reading the question, drafting the query and phrasing the answer, and unreliable at every fact in between: it will confabulate a catalogue number, assign an object to the wrong orbital regime, or quote a rule from the wrong jurisdiction. The fix is not a better prompt, it is a symbolic gate.
Three gates apply well here. Constrain generation to identifiers that exist: the graph-constrained reasoning approach of arXiv 2410.13080 compiles valid paths into a trie so the model can only emit entities and relations present in the graph, which kills invented catalogue numbers outright. Ground retrieval in a curated ontology rather than free-form extraction: OG-RAG, arXiv 2412.15235, reported a 55 per cent improvement in fact recall and 40 per cent in correctness by anchoring retrieval in domain ontology structure. Then check the proposed assertions against the model: the loop in arXiv 2504.07640 has an OWL reasoner consistency-check what the language model asserted and feeds explanations of the inconsistency back for repair. In orbital terms the reasoner can catch that an object cannot be both in the geostationary protected region and have a stated mean motion of fifteen revolutions per day, or that a disposal claim cites a rule that does not apply to its licensing state.
None of this requires new science. It requires a graph with real identities, an ontology that states what cannot be true, and a pipeline that refuses to pass unverified claims through to a human. That combination is the subject of our neurosymbolic AI in space course, which builds an orbital knowledge graph and the checking layer over it. For a sector where a wrong answer is a manoeuvre nobody needed or a collision nobody avoided, that is the right architecture.
