ESA have developed a method to automatically extract critical information from space mission data to populate a Knowledge Graph.

This tool aims to improve knowledge reuse when designing new missions.

In the business of designing and launching space missions, knowledge management can encourage innovation and boost efficiency.

Recent studies show that the demand for satellites will grow more than fourfold during the years 2021 to 2030. To keep up with this boom, ESA is finding new ways to structure and reuse knowledge.

However, key information about ESA missions - ranging from their requirements, objectives and payloads to their propulsion systems and target environments – are scattered across various documents, existing in different formats.

ESA wanted to provide a unified view of the space ecosystem, to enable mission experts to find this unstructured information more efficiently.

Leveraging recent technology developments in a field of artificial intelligence known as Natural Language Processing (NLP), researchers at ESA evaluated how Language Models (LMs) can support the restructuring of key, publicly available mission information.

They developed an initial proof-of-concept relying on a graph database, known as a Knowledge Graph.

“The goal with the Knowledge Graph is to unify and centralise the disseminated data we have accumulated on space missions, providing engineers and users with a homogeneous overview of the space ecosystem,” says ESA Research Fellow, Audrey Berquand.

“We wanted to go from stand-alone, textual, mission descriptions found on the eoPortal Directory, to a structured and unified database. This way users would not have to read through thousands of webpages - instead, they could query all knowledge from a single interface.”

The Knowledge Graph is essentially a database that structures data in nodes and edges - relations linking these nodes. It has an additional semantic layer, which allows reasoning, so that it can infer implicit information encoded into the existing structure.

Automatic population of the Knowledge Graph is achieved by using LMs to parse data from unstructured datasets. The information is partially verified through comparison with a human validated dataset.

The team chose to use a common pre-trained LM, the Generative Pre-trained Transformer 3 (GPT-3), which uses deep learning to generate human-like text.

“Our method is innovative as we use one of the most powerful NLP models available, GPT-3 from OpenAI, to deal with unstructured datasets,” explains Berquand.

Several case studies have been developed to illustrate the potential applications of the Knowledge Graph in facilitating knowledge management and reuse during space mission design.

The team presented their methodology recently at the 10th International Systems and Concurrent Engineering for Space Applications Conference (SECESA 2022).

Future tasks include exploring the replacement of GPT-3 with open-source models and aligning the Knowledge Graph with the ESA Space System Ontology.