Enhancing Nuclear Safety: How AI Powers Reliable Decision Support in Critical Systems

      The reliability and safety of complex engineered systems are paramount, especially in high-risk industries like nuclear energy, aviation, maritime transportation, and chemical processing. In these sectors, the ability to quickly and accurately synthesize information from vast archives of regulatory documents, operational procedures, and technical guidance is not just beneficial—it's critical for preventing accidents, mitigating risks, and ensuring operational continuity. Incomplete or delayed information retrieval can have catastrophic consequences, directly compromising risk analysis, fault diagnosis, and overall operational decision-making.

      Recent advancements in artificial intelligence (AI) and machine learning (ML), particularly in generative AI (GenAI), have introduced powerful new capabilities for natural language understanding and automated reasoning. These technologies offer promising avenues for decision support in safety-critical environments. However, applying general-purpose AI tools to highly specialized fields like nuclear engineering presents unique challenges, primarily stemming from fragmented documentation, the risk of AI "hallucinations," and stringent data security and privacy requirements. Addressing these specific needs requires a tailored approach, one that goes beyond generic AI and embeds domain-specific intelligence directly into the decision-making process.

Bridging the Knowledge Gap: The Challenge for AI in Nuclear Engineering

      The nuclear science and engineering sector faces a persistent bottleneck in knowledge management. Critical technical data, regulatory guidance, and modeling and simulation (M&S) benchmarks are often scattered across disparate international and national silos, such as the International Atomic Energy Agency (IAEA) and various national regulatory bodies. The sheer volume and heterogeneity of these authoritative resources can overwhelm professionals, making efficient and accurate information retrieval exceedingly difficult. This challenge is particularly acute for nuclear Safety, Security, and Safeguards (commonly known as nuclear-3S) applications.

      General-purpose Large Language Models (LLMs), while adept at summarization and complex question-answering for broad topics, fall short in specialized nuclear domains due to several fundamental issues. Firstly, their domain reliability is limited; trained on vast public datasets, they possess only superficial fluency in intricate nuclear concepts. When pressed for specifics—such as thermal-hydraulic safety margins, plant-specific standard operating procedures, or detailed security recommendations—they tend to "hallucinate," generating plausible but incorrect or fabricated outputs. Secondly, data security and privacy are non-negotiable in nuclear engineering. Uploading sensitive or proprietary documents to commercial or public AI platforms poses unacceptable risks of data leakage and intellectual property theft. Given the sensitive nature of design details and security information, strict control over data flow is imperative. Finally, while organizational LLM platforms might offer some cost advantages, they often don't fully resolve domain reliability or provide the stringent data control and security guardrails demanded by sensitive nuclear workflows.

Introducing RADIANT-LLM: An Agentic RAG Framework for Nuclear Applications

      To address these complex challenges, researchers have developed RADIANT-LLM (Retrieval-Augmented, Domain-Intelligent Agent for Nuclear Technologies using LLM). This multi-modal Retrieval-Augmented Generation (RAG) framework is specifically designed to provide reliable decision support in nuclear safety, security, and safeguards applications. RAG fundamentally enhances an LLM's capabilities by providing it with access to a relevant, verified knowledge base before it generates a response. This ensures that the AI's output is grounded in factual, domain-specific information, significantly reducing the risk of hallucinations.

      RADIANT-LLM employs a local-first, model-agnostic architecture. This means the system is designed to operate on local infrastructure, ensuring data sovereignty and removing dependencies on external cloud services. This local deployment is crucial for industries dealing with highly sensitive data, enabling full control over information flows and security. The framework also integrates a multi-modal document ingestion pipeline, meaning it can process and understand information not just from text, but also from figures, diagrams, and other visual elements within technical documents. This capability allows for page- and figure-level retrieval, offering a comprehensive understanding of complex blueprints and schematics. ARSA Technology implements similar local processing capabilities with its AI Box Series, providing pre-configured edge AI systems for rapid, on-site deployment in environments where data privacy and low latency are critical.

Ensuring Accuracy and Trust: How RADIANT-LLM Mitigates AI Hallucinations

      A core innovation of RADIANT-LLM lies in its agentic layer. This layer acts like an intelligent assistant, coordinating various domain-specific tools and processes to ensure high-quality, trustworthy outputs. It enforces citation-backed responses, complete with provenance tracking, meaning every piece of information generated can be traced back to its original source within the structured, metadata-rich knowledge base. This feature is vital for transparency and auditability in safety-critical environments. The framework also supports human-in-the-loop validation, allowing domain experts to review and validate AI-generated responses, further reducing the risks of hallucination and ensuring the output aligns with expert judgment.

      The specialized knowledge base, enriched with metadata, categorizes and cross-references information at a granular level, including specific pages and figures. This detailed organization allows the RAG system to retrieve the most precise and relevant context for any given query. For enterprises requiring bespoke AI functionalities tailored to their unique operational complexities and data structures, solutions like ARSA Technology's Custom AI Solutions can implement such intricate knowledge management and agentic systems, ensuring that AI delivers measurable financial outcomes and operational efficiency. The integration of such frameworks within existing infrastructure is key to transforming passive data into predictive intelligence, enabling stakeholders to make informed decisions faster and more reliably.

Validating Trustworthiness: Performance in Real-World Scenarios

      To rigorously evaluate the RADIANT-LLM framework, the researchers developed a suite of domain-aware metrics. These included Context Precision (CoP), which measures how relevant the retrieved information is to the query; Hallucination Rate (HR), quantifying how often the AI generates incorrect or fabricated information; and Visual Recall (ViR), assessing the system's ability to accurately retrieve visual data. These metrics were applied to expert-curated benchmarks derived from real-world Used Nuclear Fuel Storage Facility design guidance.

      The results of this evaluation were highly encouraging. Across varying knowledge base sizes, Context Precision (CoP) and Visual Recall (ViR) consistently remained within an impressive 85–98% band. Crucially, the hallucination rates observed with RADIANT-LLM were substantially lower than those seen in general-purpose LLM deployments. When the same nuclear-specific queries were posed to commercial LLM platforms without the RAG layer, both hallucinations and citation errors increased markedly. This rigorous validation underscores that a locally controlled, multi-modal RAG framework with domain-specific retrieval and robust provenance enforcement is not merely advantageous but necessary to achieve the factual accuracy, transparency, and auditability that nuclear engineering workflows demand. This principle is vital in critical sectors where ARSA provides AI Video Analytics, ensuring systems can identify PPE compliance, traffic violations, crowd density, and restricted area intrusions with 99.7% accuracy, providing reliable operational intelligence.

The Future of Secure and Auditable AI in Critical Infrastructure

      The development of frameworks like RADIANT-LLM marks a significant step forward in the application of AI to safety-critical industries. By focusing on domain-specific knowledge, local processing, multi-modal data interpretation, and rigorous traceability, these systems overcome the inherent limitations of generic AI models. This approach ensures that AI can serve as a truly reliable decision support tool, enhancing not just efficiency but also the fundamental safety and security of complex operations.

      For enterprises and governments operating in sensitive environments, the ability to deploy AI that guarantees data ownership, minimizes latency, and adheres to stringent compliance requirements is paramount. The principles demonstrated by RADIANT-LLM – such as robust provenance, human oversight, and a deep understanding of domain nuances – are critical for building trustworthy AI solutions across various industries. As technology providers, ARSA Technology is committed to delivering production-ready AI and IoT solutions engineered for precision, scalability, privacy, and operational reliability in diverse sectors, including defense, smart cities, and industrial operations.

      For organizations seeking to implement cutting-edge AI and IoT solutions that deliver measurable impact and enhance security in mission-critical operations, it is essential to partner with experts who understand both the technological frontier and the practical realities of deployment.

      **Source:** Ndum Ndum, Z., Tao, J., Ford, J., Yim, M., & Liu, Y. (2026). RADIANT-LLM: an Agentic Retrieval Augmented Generation Framework for Reliable Decision Support in Safety-Critical Nuclear Engineering. arXiv preprint arXiv:2604.22755. Available at: https://arxiv.org/abs/2604.22755

      To explore how ARSA Technology’s AI and IoT solutions can enhance reliability, security, and operational efficiency in your critical infrastructure, please contact ARSA for a free consultation.