Link to the code: Zae-Project / arkspace-core

How LLMs Are Managing the Parts That Keep Satellites Flying

Before a capacitor goes into a satellite, it needs to be qualified. That means demonstrating it will survive radiation, thermal cycling, vacuum, vibration, and whatever specific combination of conditions the mission profile demands. Qualification is expensive. A single component test campaign can run from tens of thousands to hundreds of thousands of euros, depending on the specification.

What the aerospace industry has not solved cleanly is the knowledge management problem underneath qualification. Spacecraft designers at large manufacturers like Thales Alenia Space work across dozens of programs simultaneously. A component qualified for one program under one specification may be directly usable on another, or it may require additional testing, or the required qualification may already exist in a database that nobody thought to check. Figuring out which situation applies requires querying fragmented internal data systems, cross-referencing part numbers, and comparing qualification records against current requirements. None of this is automated. Most of it is done manually by engineers.

A 2026 paper from Politecnico di Milano, Quantia Consulting, Motus ml, and Thales Alenia Space describes a production system that changes this. The authors — Antonio De Santis, Marco Balduini, Matteo Belcao, Andrea Proia, Marco Brambilla, and Emanuele Della Valle — combined large language models with knowledge graphs to build a semantic data integration layer over Thales’s component qualification databases. The system has been running in production at Thales Alenia Space with over 50 authenticated users, and ESA is in active discussions about an industry-wide deployment.

The Data Problem in Aerospace Manufacturing

Two databases hold the relevant information at Thales. The Product Lifecycle Management database (PLM-DB) contains component specifications: part numbers, manufacturer names, technical parameters. The Quality Control database tracks qualification status: which components have been tested, to which standards, with what results.

The problem is that these databases were not built to talk to each other in a semantically consistent way. Part numbers appear in different formats across records. Manufacturer names are inconsistent. Querying whether a specific resistor has a valid qualification for a new program requires cross-referencing two systems with mismatched schemas, unresolved ambiguities in part identification, and no automated pathway for identifying alternative components with equivalent qualifications.

The manual process takes, on average, 40 person-minutes per component. For a spacecraft with hundreds to thousands of components, each requiring qualification verification, the cumulative effort is substantial. The cost is not just time. It is the risk of missed reuse opportunities: qualifications that exist but are not found, leading to redundant testing campaigns that could have been avoided.

The Solution: LLMs Plus Knowledge Graphs

The system the paper describes operates in three phases.

The first phase is data cleaning. LLMs extract and normalize part numbers from unstructured text fields across both databases. Manufacturer names — which appear in dozens of variant forms across records — are standardized. Human experts review a sample of the LLM’s outputs and provide corrections, which feed back into improving the model’s performance. This phase creates a consistent foundation for the integration that follows.

The second phase builds the semantic layer. An ontology — a formal vocabulary defining what “component,” “qualification,” “part number,” and their relationships mean — is mapped across both databases using a system called Ontop, which translates relational database content into a virtual knowledge graph. The graph does not store a copy of the data. It provides a semantic view over the existing systems that allows them to be queried as if they were unified. SPARQL, a query language for knowledge graphs, is used to retrieve information across both databases simultaneously.

The third phase is the query interface. Given a new program’s component requirements, the system searches for direct qualifications (the exact part number has been qualified to the required standard) and qualification-by-similarity (a sufficiently similar component has been qualified, and its qualification may be transferable). Similarity search uses vector embeddings, finding components with comparable technical parameters even when the part numbers are different.

The performance results are specific. Direct qualification identification achieves 92% F1-score. Qualification by similarity achieves 94% F1-score. The time per component drops from 40 person-minutes to 5. At scale — the paper models this at 10,000 components — the system achieves over 80% effort savings compared to the manual process, and outperforms RAG (retrieval-augmented generation) approaches that work better only at smaller scales.

Why This Matters for Satellite Manufacturing

The qualification data problem the paper describes is not specific to Thales. It is structural to the aerospace industry. Any manufacturer operating across multiple programs with heterogeneous internal systems faces the same fragmentation. The knowledge that qualification X exists for component Y, and could substitute for component Z that nobody thought to check, sits in databases that engineers cannot easily cross-query.

The consequences are real and measurable. Redundant qualification testing is direct cost. Missed reuse opportunities extend development timelines. Qualification errors — components used without verified qualification — are mission risk.

For satellite programs, where components must be qualified not just to commercial standards but to radiation, vacuum, and thermal specifications relevant to their orbital environment, the qualification database problem is compounded. The space-specific qualification landscape is smaller than the commercial electronics landscape. Radiation-hardened components are designed and tested by a limited set of manufacturers. The qualification records for space-grade components are distributed across manufacturers, national space agencies, and research institutions, with no centralized queryable database.

The paper’s approach — semantic integration via virtual knowledge graphs rather than physical data migration — is relevant here precisely because it does not require rebuilding the existing database infrastructure. It provides a query layer over systems that continue to operate independently, removing the integration barrier without the organizational cost of a unified database migration project.

The ESA Connection

The paper notes that discussions are underway between Thales Alenia Space, ESA, and other European space industry stakeholders about extending the system to a shared qualification service across the sector. The concept is a federated knowledge graph that allows multiple organizations to query each other’s qualification records — subject to access controls — without centralizing their data.

If implemented, this would address a significant inefficiency in the European space industry. Currently, qualification campaigns conducted by one prime contractor are largely invisible to others. A radiation-hardened chip qualified by Airbus Defence and Space for one ESA program may be directly relevant to a Thales program, but finding that out requires informal communication, contractual relationships, or industry databases that are incomplete and manually maintained.

The TRL for the Thales production deployment is 4-5: demonstrated in a relevant industrial environment with real operational users. The European-wide feasibility study is in progress. The timeline to broader deployment depends on organizational agreements and data governance frameworks, not technical feasibility.

Connection to Orbital Computing Development

The challenge of qualifying radiation-hardened neuromorphic chips for space is a direct instance of the problem this paper addresses.

Carnegie Mellon’s 22nm FinFET radiation-hardened chip designs, heading toward a 2026 CubeSat test, represent a small number of devices with specific radiation performance data. Intel’s Loihi 2 neuromorphic processor has commercial qualification data but limited space-specific radiation testing. Qualifying these systems for operational use in a constellation like the Exocortex Constellation requires demonstrating that each component meets the mission’s total ionizing dose, single-event effect, and thermal cycling requirements.

A semantic knowledge graph over existing space-grade component qualification data — including radiation tolerance LEO standards — would allow designers to identify which commercial neuromorphic components have partial qualifications that could be extended, which rad-hard designs have relevant heritage, and what additional testing is required to close the gaps. This is exactly the efficiency gain the paper demonstrates at Thales.

The deeper point is architectural. The paper’s system does not replace engineering judgment. It removes the information access barrier that prevents engineers from applying their judgment effectively. The same logic applies across the orbital computing development pipeline: the bottleneck is not technical understanding, it is organized access to the existing technical record.

Official Sources