Conference Agenda

During a severe nuclear accident, a prolonged loss of coolant causes fuel elements to overheat and melt due to their own decay heat. As this molten fuel interacts with surrounding structural materials, it forms a debris bed - a complex, porous, heat-generating accumulation of fragmented particles. In extreme scenarios, the molten core can migrate to the lower plenum of the reactor pressure vessel (RPV). In the presence of residual cooling water, the material may fragment and form a debris bed composed of irregularly shaped and sized particles. [1][2]

Effective quenching of this debris is vital to prevent remelting; failure to cool the bed can lead to structural failure of the RPV and the catastrophic release of core material, posing significant safety risks. To mitigate these risks, experimental studies are necessary to optimize coolability by analysing how factors such as porosity, particle geometry, and flooding strategies influence the quenching process. [3] The FLOAT facility at the Institute of Nuclear Technology and Energy Systems (IKE), University of Stuttgart is specifically designed to study the quenching phenomena. It utilizes a cylindrical debris bed with 200 mm diameter and 300 mm height that is heated to specific temperatures using an inductive heating system. Quenching experiments are performed under a range of conditions, systematically varying parameters. While 45 embedded thermocouples provide extensive temperature data, the facility also addresses the need for visual insights into two-phase flow near the vessel wall. To achieve this, 12 video cameras are positioned around the transparent glass outer wall of the test section. These cameras record the quenching process, allowing for the detailed visualization of quench front progression and two-phase flow behaviour within the bed.

Previous research presented at NENE 2025 successfully developed a CNN to segment experimental visual data from FLOAT into wet, dry and background regions [4]. While this established a robust optical classification framework, it remained limited to segmentation without any predictive capability. The present study advances this work by introducing a spatiotemporal predictive framework that completely forecasts the quench front progression of the two-phase flow visually. This approach aims to forecast the complete spatial and temporal progression of the two-phase flow visually. The proposed model utilizes the previously segmented chronological image frames as ground-truth training data. To effectively capture both the spatial features of the debris bed and the transient dynamics of the quench front, a Convolutional Long Short-Term Memory (ConvLSTM) architecture is employed. Unlike standard LSTMs, the ConvLSTM architecture replaces fully connected matrix multiplications with convolution operations at its internal gates. This critical modification enables the network to simultaneously retain the 2D spatial topology of the debris bed images while learning the chronological dependencies of the advancing fluid flow.

The predictive model is explicitly conditioned on a comprehensive set of experimental input parameters, such as initial bed temperature, porosity, flooding configuration, and cooling water temperature. These physical parameters are integrated into the network to guide the hidden states of the ConvLSTM, enabling the model to output a chronologically ordered series of segmented images that provide the complete visual progression of the quenching process over time. The generated image sequences are rigorously evaluated against actual experimental segmentations using performance metrics such as the Structural Similarity Index Measure (SSIM), Dice Coefficient and pixel-wise classification accuracy to ensure reliable performance. Furthermore, this framework enables the validation of thermal-hydraulic system codes against complex, multi-dimensional experimental observations. The model will be evaluated against experimental segmentations using these metrics, with success criteria defined by its ability to correctly reproduce quench front geometry and progression speed across varying porosity and flooding configurations. This framework is intended to reduce experimental overhead by enabling rapid screening of parameter combinations prior to physical testing, directly supporting debris bed coolability assessments in severe accident management.

References:

[1] Ahmed, Z. et al. (2022) ‘Experimental investigation on the coolability of nuclear reactor debris beds using seawater’, International Journal of Heat and Mass Transfer, 184, p. 122347. doi:10.1016/j.ijheatmasstransfer.2021.122347.

[2] Leininger, S., Kulenovic, R., and Laurien, E. (2015) ‘Experimental Investigations on the Coolability of Debris Beds under Variation of Inflow Conditions’, Proceedings of the 16th International Topical Meeting on Nuclear Reactor Thermal Hydraulics (NURETH-16, Chicago, IL, August 30–September 4, 2015). Available at: https://glc.ans.org/nureth-16/data/papers/12903.pdf.

[3] Li, Y. et al. (2025) ‘Debris Bed coolability in hypothetical core disruptive accidents: Theory, experiment, and Numerical Simulation Review’, Progress in Nuclear Energy, 184, p. 105700. doi:10.1016/j.pnucene.2025.105700.

[4] Ananya, N. R., Kulenovic, R., and Starflinger, J. (2025) ‘Development of a Machine Learning Model for Optical Two-Phase Flow Detection in Debris Bed’, Proceedings of the 34th International Conference Nuclear Energy for New Europe (NENE 2025, Bled, Slovenia, September 8–11, 2025). Available at: https://www.djs.si/upload/nene/2025/NENE_2025_Conference_Proceedings.pdf.

Large language models (LLMs) are a class of artificial intelligence tools based on deep
neural networks that can understand and generate text at scale. They are increasingly useful
as interfaces to technical knowledge, but their direct use in nuclear engineering remains lim-
ited by traceability, confidentiality, hallucination risk, and poor access to site- or institution-
specific documentation. This work presents an in-development local LLM deployment for
deep searching through heterogeneous nuclear and technical document collections using agen-
tic retrieval-augmented generation (RAG) on GPUs (graphical-processing units) available on
non-dedicated local servers and workstations. The system supports internal knowledge re-
trieval across archival reports, technical documentation, logs, forms, schematics, NPP-related
material, textbooks, journals, photographs, news sources, and other supporting documents.
The corpus is heterogeneous in both format and language, including Slovenian, English, and
mixed-language technical material. All processing is performed locally, without sending doc-
uments or queries to external cloud services. This is essential for maintaining confidentiality
and traceability in the nuclear field.

The system runs on a local cluster of non-dedicated servers and workstations equipped
with a heterogeneous set of cluster, professional, and workstation-grade GPUs. Requests are
managed by a router that schedules work according to available resources and can automati-
cally fall back to smaller local models under resource pressure. This keeps the system usable
during peak demand or when some machines are occupied by other primary tasks. Documents
are brought into the system through a processing pipeline that reads scanned pages, interprets
figures and tables, and stores extracted content in a searchable database. Rather than per-
forming a single database lookup, the system uses agentic RAG: when a user asks a question,
it tries multiple search strategies, reformulates the query, tests alternative terminology, and
combines meaning-based and keyword search before passing the most relevant material to the
main model for answering. Additional tools can look up acronyms, fetch web content, retrieve
data from the local TRIGA meteorological station and check reactor schedules.

The system is not intended to replace expert judgement, but to reduce the time needed to
search through large and poorly organised document archives. Answers are grounded in re-
trieved source material and link back to the original documents, which is especially important
in nuclear engineering, where fluent but unsupported responses can be misleading. The user
interface reflects this requirement by providing a live view of backend activity, model deci-
sions, tool calls, retrieved sources, and intermediate reasoning output. Retrieved sources can
be opened and inspected while the answer is still being generated. Unlike most commercial AI
tools, the system is not limited to short interactions and can support longer, in-depth research
queries.

Initial use cases show the practical value of this approach. The system has been used
to locate original shipment documents and specifications for TRIGA reactor graphite from
thousands of old scanned documents, identify when specific reactor components were replaced
by analysing logs and articles, find relevant nuclear references during technical writing, and
support reactor physics exercises for students. These examples demonstrate that the main
value of the system is not only text generation, but faster access to otherwise difficult-to-find
institutional knowledge.

The system is under active development, and document ingestion is still ongoing. Known
challenges include OCR errors, inconsistent quality of scanned documents, duplicated or re-
vised documents, source weighting, access-control requirements, and non-deterministic re-
trieval performance. In some cases, the system finds the relevant material quickly; in others,
similar questions may require additional search attempts or expert guidance. Despite these
limitations, the work demonstrates that locally hosted, GPU-accelerated AI infrastructure for
technical document retrieval is practical in demanding nuclear research environments