Physics-Informed Multi-Stage Deep Learning Framework Development for Digital Twin-Centred State-Based Reactor Power Prediction

James Daniell, Kazuma Kobayashi, Susmita Naskar, Dinesh Kumar, Souvik Chakraborty, Ayodeji Alajo, Ethan Taber, Joseph Graham, Syed B. Alam

Abstract

Computationally efficient and trustworthy machine learning algorithms are essential for the development of Digital Twin (DT) frameworks in the nuclear industry. This study focuses on the "ML-driven prediction algorithms" component, developing a multi-stage predictive model consisting of two feedforward Deep Neural Networks (DNNs) to determine the final steady-state power of a reactor transient. The architecture converts probabilistic classification into continuous output variables to improve reliability. Tested with inputs from the Missouri S&T Reactor (MSTR), the combined model yields 96% classification accuracy for the first stage and 92% absolute prediction accuracy for the second stage. This paper discusses the development procedure and the role of such models in real-time DT operations.

Introduction: The Need for Speed in Nuclear Safety

In nuclear reactor operations, safety and efficiency hinge on the ability to predict reactor behavior during "transients"—periods of change in state, such as adjusting control rods to modify power levels. Traditionally, engineers rely on high-fidelity physics codes like SCALE or TRACE. While these tools provide high accuracy, they are "computationally thirsty," often requiring significant time to run. This latency makes them unsuitable for real-time decision support during rapidly unfolding events.

Digital Twins (DTs) offer a solution by providing a virtual mirror of the physical asset. However, a functional DT requires prediction algorithms that are both fast and physically grounded. Standard Deep Neural Networks (DNNs) are fast but often act as "black boxes" that lack physical intuition and are prone to overfitting. This research bridges the gap by creating a physics-informed, multi-stage deep learning framework designed for real-time reactor power prediction.

Methodology: A Multi-Stage Architecture

The proposed framework breaks the complex prediction problem into two manageable stages to enhance reliability and ease of analysis:

  1. Classification Stage: A front-end DNN categorizes the expected power output into one of five discrete "bins" (e.g., <90W, <900W, etc.). This stage acts as a probabilistic filter, narrowing down the search space for the subsequent model.
  2. Regression Stage: A back-end DNN takes the classification output alongside the original input data to predict a precise, continuous value for the final steady-state power level.

The researchers explored different input configurations, including an "All-In-One" (AIO) approach and a "Split-Input" approach that preserves the physical structure of the data by separating initial states from final target states. Furthermore, they compared models using raw control rod heights against those using calculated "reactivity," a parameter with more direct physical significance in nuclear kinetics.

Data Strategy: Bridging Real-World Logs and Physics Simulations

A significant challenge in applying deep learning to nuclear reactors is the scarcity of high-quality transient data. This study utilized a hybrid dataset approach:

  • Real-World Data: The team collected 542 observations from historical transient logs at the Missouri University of Science and Technology Reactor (MSTR) spanning 2013 to 2014.
A sample hourly operating log from the Missouri S&T Reactor (MSTR) used to build the real-world dataset.
  • Physics-Informed Data Augmentation: To provide the thousands of data points required for robust DNN training, the authors used the power defect equation—a fundamental principle of nuclear physics—to generate synthetic data.
  • Noise Enhancement: Random noise was intentionally added to the simulated data. This "noise enhancement" prevents the model from simply memorizing perfect theoretical curves, forcing it to learn the underlying patterns and generalize better to real-world sensor data.
Comparison of power distributions between the original MSTR dataset and the physics-informed generated dataset, demonstrating the efficacy of the augmentation strategy.

Key Insights and Results

The framework demonstrated that data-driven models can achieve high precision when guided by physical principles:

  • Domain Knowledge Matters: Models utilizing reactivity as an input significantly outperformed those using raw rod positions. This confirms that feeding models parameters with physical meaning is more effective than expecting the network to derive physics from scratch.
  • Architecture and Accuracy: The "Split-Input" models, which respect the temporal relationship of the transient, outperformed AIO models. The final framework achieved a 96% classification accuracy and a Mean Absolute Percentage Error (MAPE) of 2.3% for power prediction.
  • Operational Viability: Unlike traditional simulations that take minutes or hours, these DNN-based predictions are nearly instantaneous. This speed allows for real-time monitoring and "what-if" scenario testing by reactor operators.

Conclusion and Future Directions

This study successfully demonstrates a multi-stage, physics-informed deep learning framework capable of serving as a predictive engine for a nuclear reactor Digital Twin. While the current model focuses on predicting the final steady-state power, future work will involve incorporating Long Short-Term Memory (LSTM) or Convolutional Neural Networks (CNNs) to predict the entire time-series trajectory of a transient. Additionally, integrating formal Uncertainty Quantification (UQ) will be vital for providing operators with confidence intervals, a prerequisite for safety-critical deployment in the nuclear industry.

Summary updated

Abstract

Computationally efficient and trustworthy machine learning algorithms are necessary for Digital Twin (DT) framework development. Generally speaking, DT-enabling technologies consist of five major components: (i) Machine learning (ML)-driven prediction algorithm, (ii) Temporal synchronization between physics and digital assets utilizing advanced sensors/instrumentation, (iii) uncertainty propagation, and (iv) DT operational framework. Unfortunately, there is still a significant gap in developing those components for nuclear plant operation. In order to address this gap, this study specifically focuses on the "ML-driven prediction algorithms" as a viable component for the nuclear reactor operation while assessing the reliability and efficacy of the proposed model. Therefore, as a DT prediction component, this study develops a multi-stage predictive model consisting of two feedforward Deep Learning using Neural Networks (DNNs) to determine the final steady-state power of a reactor transient for a nuclear reactor/plant. The goal of the multi-stage model architecture is to convert probabilistic classification to continuous output variables to improve reliability and ease of analysis. Four regression models are developed and tested with input from the first stage model to predict a single value representing the reactor power output. The combined model yields 96% classification accuracy for the first stage and 92% absolute prediction accuracy for the second stage. The development procedure is discussed so that the method can be applied generally to similar systems. An analysis of the role similar models would fill in DTs is performed.
BibTeX
@misc{daniell2022physics,
  author = {James Daniell and Kazuma Kobayashi and Susmita Naskar and Dinesh Kumar and Souvik Chakraborty and Ayodeji Alajo and Ethan Taber and Joseph Graham and Syed B. Alam},
  title = {{Physics-Informed Multi-Stage Deep Learning Framework Development for Digital Twin-Centred State-Based Reactor Power Prediction}},
  year = {2022},
  eprint = {2211.13157},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2211.13157},
}
Figure