Conference Agenda

The paper outlines ongoing research in a national funded joint project, applying machine learning methods to predict late-phase phenomena observed during severe accidents. The aim is to produce resource-efficient simulations that improve the understanding and predictive capabilities for these late-phase phenomena. Emphasis is on the long-term coolability of debris beds in the vessel, its remelting and possible relocation in the cavity as ex-vessel debris bed. For this purpose, a machine learning model is intended to be integrated to a PSS inhouse version of AC² program package, developed by Gesellschaft für Anlagen- und Reaktorsicherheit (GRS) gGmbH, to make the complex calculation of heat transfer and dryout heat flux more efficient.
The aim is to develop an artificial neural network as a machine learning model. Several different network types are initially developed for this purpose, for example a physically informed neural network (PINN). As work progresses, these models should become more complex and be able to provide a fast and reliable prediction of the debris bed coolability for as many different boundary conditions as possible. According to the current status, the tested network types are already able to produce good preliminary results. They are able to predict the dryout heating power, selected as the output variable, to a good approximation. The training data are generated using the COCOMO-3D detail code from the Institute for Nuclear Energy and Energy Systems (IKE) at the University of Stuttgart.

The recent developments in data-driven methods have paved the way to new methodologies to provide accurate state reconstruction of engineering systems; nuclear reactors represent particularly challenging applications for this task due to the complexity of the strongly coupled physics involved and the extremely harsh and hostile environments, especially for new technologies such as Generation-IV reactors. Data-driven techniques can combine different sources of information, including computational proxy models and local noisy measurements on the system, to robustly estimate the state. This work leverages the novel Shallow Recurrent Decoder architecture to infer the entire state vector (including neutron fluxes, precursors concentrations, temperature, pressure and velocity) of a reactor from three out-of-core time-series neutron flux measurements alone. In particular, this work extends the standard architecture to treat parametric time-series data, ensuring the possibility of investigating different accidental scenarios and showing the capabilities of this approach to provide an accurate state estimation in various operating conditions. This paper considers as a test case the Molten Salt Fast Reactor, a Generation-IV reactor concept, characterised by strong coupling between the neutronics and the thermal hydraulics due to the liquid nature of the fuel. The promising results of this work are further strengthened by the possibility of quantifying the uncertainty associated with the state estimation, due to the considerably low training cost. The accurate reconstruction of every characteristic field in real-time makes this approach suitable for monitoring and control purposes in the framework of a reactor digital twin.

In the event of a severe accident in a nuclear power plant, accident prediction using artificial intelligence (AI) has gained attention as a promising Accident Management Support Tool (AMST). A notable approach is the development of surrogate models for accelerated accident prediction through deep learning-based supervised learning. Such models alleviate the computational complexity of severe accident analysis codes by training on data generated from the codes, significantly reducing the computational costs.

However, surrogate models often present structural challenges, leading to low-resolution predictions and increased uncertainty, hindering effective decision-making for operators. This issue contradicts the essential requirements for AMST reliability. Structural issues arise from low temporal resolution and information loss during data preprocessing for training, limiting the model's accuracy due to cumulative computational errors in time series forecasting. Consequently, using surrogate models to predict thermal-hydraulic variables with refined time resolution during accident progression can yield unreliable results.

To address these challenges, this study aims to develop an auxiliary surrogate model to support accident prediction by identifying time varying patterns in the accident prediction data. This model is designed to predict the onset time and amplitude of physical variations, enhancing the accuracy and reliability of surrogate-based predictions during severe accidents.

This work is performed in the frame of the EU-funded project ASSAS (Artificial intelligence for Simulation of Severe AccidentS), which aims at developing a basic-principles severe accident simulator for a generic PWR-1300MW, by replacing models from ASTEC (severe accident code developed by IRSN) with machine-learning surrogate models. PSI’s tasks address essentially the CESAR module (thermal-hydraulic solver) in the primary and secondary circuits. This paper proposes a surrogate model able to reproduce the thermal-hydraulic behaviour of the reactor cooling system (RCS) during the early phase of a Station Blackout (SBO), i.e., until hydrogen generation, with a significant speed-up factor.

A suitable training dataset must be generated. A base case scenario is considered, involving a SBO without any safety measures until the onset of core oxidation. From this base case, various calculations are performed by depressurizing remotely the primary circuit at different times, followed by the recovery of emergency water injection, also at different times. From these ASTEC calculations only the variables needed for the surrogate model development are extracted. These include state variables for each control volume within the RCS domain, boundary conditions, and additional variables that provide information about the overall evolution of the accident and are useful for Machine Learning. The surrogate model is expected to compute each time-step, like ASTEC does, while also accounting for user decisions interactively during the accident simulation. The Machine Learning methods considered in this work are based on artificial neural networks, and more specifically recurrent neural networks, which are commonly used for time-series.

Small Modular Reactors (SMRs) are gaining increasing attention due to their enhanced safety features, flexibility, and scalability. Ensuring timely and accurate assessments of break sizes during break accidents is crucial for maintaining the safe and reliable operation of SMRs. However, current methods for evaluating break sizes mainly rely on the personal judgment of operators, which often fail to meet the speed and accuracy requirements in high-risk, time-sensitive situations. This limitation can hinder effective decision-making and risk management. Recent advancements in artificial intelligence (AI) have accelerated the development of data-driven methods for break size prediction, demonstrating significant potential for improving operational reliability. Machine learning models, particularly those with interpretability features, can provide real-time, data-driven predictions of break sizes, offering a faster and more accurate alternative to traditional methods. Furthermore, the interpretability of these models can foster greater trust in AI systems, particularly in safety-critical environments such as nuclear reactors. This study investigates Direct Vessel Injection (DVI) break accidents, utilizing explainable artificial intelligence (XAI) and hyperparameter optimization techniques to develop predictive models for break size. The results demonstrate that these models enable rapid and accurate prediction of DVI break sizes based on actual operational parameters. The findings of this research provide valuable insights for developing break size prediction models base on SMRs, contributing to improved safety and operational efficiency.

This research focuses on the safety evaluation of a design extension condition involving multiple steam generator tube rupture (MSGTR) scenarios. A series of operator actions are proposed to mitigate the accident, including depressurization, auxiliary spray operation, and steam generator blowdown. The efficacy of these actions is evaluated under various uncertainties using the best estimate plus uncertainty (BEPU) approach through RELAP5/DAKOTA coupling. The generated ensemble of system responses is used to develop an AI-based prediction model. Tools of explainable artificial intelligence, specifically a combination of attention mechanisms, gradient-based attribution, and parameter interaction analysis, are implemented to examine the model's decision-making process. This framework reveals phase-specific patterns and dynamic shifts in parameter relevance as the accident progresses through different stages–from initial break flow and pressure response, through various operator interventions, to final stabilization. The analysis quantifies the coupling between primary and secondary systems, particularly during critical phases of depressurization and cooldown, while demonstrating the model's adherence to established thermal-hydraulic principles. The result highlights the AI model's general alignment with established thermal-hydraulic principles, suggesting its potential for integration into nuclear safety management, provided its transparency and interpretability continue to be rigorously validated.