In the field of nuclear reactor physics, the demand for high-fidelity core analysis has intensified, necessitating numerical methods that can resolve complex 3D power distributions with pin-wise precision. While traditional deterministic solvers—suc...
In the field of nuclear reactor physics, the demand for high-fidelity core analysis has intensified, necessitating numerical methods that can resolve complex 3D power distributions with pin-wise precision. While traditional deterministic solvers—such as Finite Difference Method (FDM) and Nodal Methods—have successfully served as the industry standard, they are fundamentally governed by iterative numerical schemes. This iterative nature imposes a cumulative computational burden, particularly in multi-query applications like core design optimization, where the solver requires re-convergence from scratch for each configuration.
To overcome this computational bottleneck, data-driven deep learning initially emerged as a compelling surrogate modeling technique. These approaches offer rapid, non-iterative inference speeds suitable for real-time applications. However, purely data-driven models face fundamental limitations in the nuclear domain: they rely heavily on massive labeled datasets—which are computationally expensive to generate using high-fidelity codes—and lack rigorous physical guarantees, raising safety concerns regarding their "black-box" predictions.
Addressing these limitations, Physics-Informed Neural Networks (PINNs) have emerged as a promising alternative. By embedding the governing equation directly into the loss function, PINNs ensure physical consistency without relying on labeled data. Yet, a critical misalignment remains between standard PINN architectures and reactor physics. Most existing PINNs rely on Fully Connected Networks (FNNs), which treat spatial domains as flattened vectors. This approach discards the inherent lattice topology of reactor cores, leading to severe scalability challenges in high-dimensional 3D problems. Consequently, there is a need for a specialized architecture that aligns the geometric inductive bias of deep learning with the structural regularity of nuclear reactors.
To realize this vision, this dissertation proposes a Physics-Informed Convolutional Neural Network (PICNN) framework for the analysis of reactor physics problems. Unlike standard Fully Connected Networks, this architecture treats the reactor core as a 3D tensor, leveraging the geometric alignment between the reactor's lattice structure and the grid-based topology of Convolutional Neural Networks (CNNs).
Within this convolutional framework, the primary objective was to accurately solve the neutron diffusion equation while maintaining the spatial integrity of the reactor data. To achieve this, a standard U-Net architecture was adopted as the backbone, utilizing its hierarchical feature extraction capabilities to capture the multi-scale physics of neutron transport. In the developmental phase, an architectural decision was required regarding the estimation of the effective multiplication factor (keff), a system parameter that depends on the flux distribution itself.
Two distinct strategies were rigorously investigated using a 2D Pressurized Water Reactor (PWR) benchmark problem: a 'Direct Prediction' model, which treats the eigenvalue as a learnable parameter inferred implicitly via an auxiliary fully connected layer, and an 'Explicit Calculation' model, which mathematically derives the eigenvalue from the network’s predicted flux utilizing the physical definition of the Rayleigh Quotient. Comparative analysis demonstrated the clear superiority of the Explicit Calculation strategy; while the Direct model exhibited instability, the Explicit model achieved a mean flux error of approximately 0.3% and a keff difference within 10 pcm, ensuring strict adherence to the neutron balance principle. Consequently, this explicit formulation was established as the core methodology, trained in a self-supervised manner by minimizing the residuals of the discretized diffusion equation without requiring labeled data.
With the architectural methodology finalized, the scalability and fidelity of the proposed framework were systematically validated across multi-dimensional regimes. Building upon the fundamental feasibility established in 2D PWR benchmark, the framework was seamlessly extended to a 3D Small Modular Reactor (SMR) core. Despite the increased dimensionality, the PICNN maintained high fidelity, achieving a flux Root Mean Square (RMS) error of 1.56% and a keff error of 17.5 pcm. Most significantly, the solver demonstrated a transformative advantage in computational efficiency. By replacing the iterative procedures of traditional fine-mesh solvers with a non-iterative neural inference, the proposed model reduced the computation time for a full 3D core state to approximately 0.20 seconds. This sub-second latency validates the framework's potential as a real-time surrogate engine for extensive design optimization and on-the-fly safety analysis.
Beyond forward prediction, the versatility of the framework was demonstrated by extending it to the inverse problem of power reconstruction. Recognizing that real-world operations involve input uncertainties and limited instrumentation, the architecture was adapted to assimilate sparse signals from in-core detectors. The resulting unified model successfully functioned as a virtual sensor, reconstructing the full 3D power distribution from sparse measurements with a flux RMS error of less than 1.3%. This performance was facilitated by the synergistic integration of measurement and physics: the assimilation of sparse observations provides critical constraints to mitigate local predictive biases, while the strict enforcement of the diffusion equation serves as a rigorous regularizer that effectively suppresses the propagation of stochastic sensor noise.
In conclusion, by demonstrating that deep learning can achieve both computational efficiency and physical consistency, this dissertation establishes a seminal methodological framework, paving the way for the next generation of scientific machine learning research in nuclear reactor analysis.