Use of innovative technologies to support RWMO knowledge management, the GRS approach

The safe disposal of radioactive waste represents one of the most demanding and long-term challenges faced by modern society, that requires an in-depth understanding of complex technical, geological and safety-related processes. There is a significant risk of knowledge loss over long time periods, especially when institutions dissolve, are restructured or when documentation is not adequately maintained or managed. Ensuring the continuity of knowledge transfer across generations is crucial to maintaining safety and security by considering past experiences and historical data. Artificial intelligence (AI) and its applications, which are currently at the forefront of technological discussions and solutions, could offer innovative tools to overcome these Knowledge Management (KM) challenges.

The application of AI in RWMO knowledge management requires a structured approach to meet the special requirements of nuclear safety: all knowledge must meet the highest standards of accuracy, reliability and validation. AI systems must be equipped with mechanisms that check data against authorized sources and prevent the generation of false or falsified information. Transparency and traceability are crucial to ensure that users can understand how results are generated and that decisions are based on validated and trustworthy information. RWMO knowledge management systems need to be dynamic and able to integrate new information as it becomes available.

In this regard, AI applications like Large Language Models (LLMs) can provide valuable support to RWMO knowledge management, but face significant challenges such as hallucinations, outdated training data, probabilistic outputs, and high resource requirements for training and operation. Therefore, pure LLMs like Generative Pre-trained Transformers (GPT) alone cannot meet the necessary requirements. Corresponding AI solutions should be based on a reliable and secure knowledge pool, which is built upon a solid foundation of high-quality data that is well structured, comprehensible and accessible to realize the full potential. This includes the use of ontologies to represent relationships between concepts, decision trees, and structured documents such as “lessons learned”, “set of essential records” or “regulations”.

GRS research in corresponding projects like KISS (FKZ 15S9448B, funded by BMBF) shows that the implementation of the knowledge pool in a so-called Graph RAG scenario seems to be appropriate for such a system. Here a LLM-Chatbot serves mainly as a communicating interface, while the data is provided by a separate knowledge pool.

The knowledge pool is based on a multi-layered architecture designed to provide contextual information efficiently. It consists of four layers: Content layer, metadata layer, semantic layer and access layer. The Content Layer stores raw data, including documents, and multimedia, as the foundation. The Metadata Layer enriches this data with attributes like tags, categories, and versioning for better organization and retrieval. The Semantic Layer transforms content into meaningful structures using ontologies, knowledge graphs, and embeddings.

This system architecture enhances LLM performance and ensures scalability and flexibility. Since the system understands the context and the relationships, more meaningful results can be obtained.

This approach is also applicable to other applications when knowledge, information and data are strongly interconnected. E.g. competence databases, systems for data management or systems to trace regulatory decisions.