Conference Agenda

The cold trap of a liquid sodium purification system maintains concentration of impurities below an acceptable level to prevent deterioration of sodium fast reactor (SFR) components. A cold trap is typically monitored with multiple thermal hydraulic sensors. Timely detection of incipient anomalies in cold trap operation is important for efficient SFR operation and maintenance. Previous work developed a deep learning long short-term memory (LSTM) autoencoder for loss-of-coolant type anomaly detection in cold trap of the liquid sodium purification system at the Mechanisms Engineering Test Loop (METL) thermal hydraulic facility at Argonne National Laboratory. We found that relative delays in response time for multi-modal sensor monitoring systems affect anomaly detection time and certainty. We have developed a novel machine learning (ML) method to estimate sensor response delays in detection of signals related to anomaly events, and to use this information to augment the data to improve detection time. The anomaly signal is detected by establishing a threshold using the density distribution of the loss for the training data. Relative sensor delays were determined during testing by finding the times when the loss of each sensor rises above their respective threshold values. The time delays were then used for synchronization of the data. The augmented data was fed back to the LSTM autoencoder to detect the anomaly using sensor-averaged loss. A parametric study was conducted, in which the anomaly was gradually reduced until the signal-to-noise ratios (SNRs) approached unity. Results indicate that synchronization improves anomaly detection, especially for lower SNR anomalies.

The High Temperature Test Facility is an integral test facility located at Oregon State University, modeled after the Modular High Temperature Gas Reactor. It is designed to provide benchmark data for phenomena such as lower plenum mixing, depressurized conduction cooldown, pressurized conduction cooldown, and normal operational conditions. Numerous sensors are installed throughout the facility to measure variables like temperature, pressure, and mass flowrate, with data recorded at 2 Hz frequency. Several methods are under study for field reconstruction in online monitoring applications. One method that is promising with sequential data, but not well-studied in nuclear engineering is Recurrent Neural Networks (RNN). This study focuses on developing data-driven RNN models,specifically gated recurrent units (GRU) and long short-term memory (LSTM), to predict sensor outputs at various locations within HTTF. The models are trained on data from one subset of sensors and applied to predict the outputs of similar sensors - this was done 400 times, with 200 permutations of LSTM and GRU models each. Mean absolute error (MAE) was used as a performance metric to evaluate the predictions. It was found that 71 of the sensors can be used to train LSTM and GRU models, which can then predict the data of the other 71 sensors very well. The MAE of the predictions ranged from 0.28°C to 4.41°C for all models and permutations. Generally, the LSTM models have a higher accuracy relative to the GRU models with overall (average MAE value of 0.721°C for LSTM as opposed to 0.788°C for GRU).

The Shallow Recurrent Decoder networks are a novel paradigm recently introduced for state estimation, combining sparse observations with high-dimensional model data. This architecture features important advantages compared to standard data-driven methods including: the ability to use only three sensors (even randomly selected) for reconstructing the entire dynamics of a physical system; the ability to train on compressed data spanned by a reduced basis; the ability to measure a single field variable (easy to measure) and reconstruct coupled spatio-temporal fields that are not observable and minimal hyper-parameter tuning. This approach has been verified on different test cases within different fields including nuclear reactors, even though an application to a real experimental facility, adopting the employment of in-situ observed quantities, is missing. This work aims to fill this gap by applying the Shallow Recurrent Decoder architecture to the DYNASTY facility, built at Politecnico di Milano, which studies the natural circulation established by internally heated fluids for Generation IV applications, especially in the case of Circulating Fuel reactors. The RELAP5 code is used to generate the high-fidelity data, and temperature measurements extracted by the facility are used as input for the state estimation. The results of this work will provide a validation of the Shallow Recurrent Decoder architecture to engineering systems, showing the capabilities of this approach to provide and accurate state estimation.

Advanced reactor and fuel designs could be crucial in decarbonizing our energy portfolio. However, their development and implementation come with specific challenges, often related to the novelty of such designs, that must be addressed to ensure that such systems operate safely, reliably, and economically viable. Strategies like the preventive maintenance of such systems can support achieving these goals by potentially reducing the maintenance and operation costs while preserving the safety and reliability of such systems.

However, the preventive maintenance of nuclear reactors may rely on real-time monitoring of some physical properties of such systems, which can be challenging. Many probe designs cannot withstand the reactor's extreme conditions (e.g., temperature, radiation). In this context, physics-informed Convolutional Neural Networks (CNNs) offer a promising non-intrusive alternative for reconstructing physical fields, such as temperature and stress distributions, using minimal sensor data. This work presents the integration of machine-learning-aided tools with coupled thermomechanical and thermal-hydraulic simulations to assess the behavior of fuel rods during both steady-state and accident scenarios.

To train our CNN, we leveraged the capabilities of the MOOSE framework to build computational models representing the fuel rod thermomechanical behavior during steady-state operation and its response during a transient situation, such as an accident condition. These models were used to build the necessary datasets to train and test the prediction performed by the CNN architecture. These efforts provide a foundation for real-time monitoring and enhanced safety assessments of advanced reactor designs, addressing challenges in operational efficiency and accident management.