TL;DR
This paper introduces a "Data-Enabled Physics-Informed Neural Network" (DEPINN) that leverages a small amount of experimental data to significantly improve the accuracy and efficiency of solving neutron diffusion eigenvalue problems (NDEPs) in nuclear reactor physics. The DEPINN framework, combined with an adaptive optimization strategy, demonstrates promising potential for real-world engineering applications.
What Problem Does This Paper Address?
Accurately predicting neutron behavior within a nuclear reactor core is crucial for its safe and efficient operation. Neutron diffusion eigenvalue problems (NDEPs) are fundamental to this, but solving them for complex, industrial-scale reactor geometries can be computationally intensive using traditional numerical methods. While Physics-Informed Neural Networks (PINNs) offer a mesh-free alternative, their application to complex NDEPs, especially with realistic two-group diffusion equations, has been limited. This paper aims to bridge this gap by developing and rigorously evaluating a DEPINN approach that incorporates experimental data to achieve engineering-acceptable accuracy for these challenging problems.
Methods
The core of the paper is the Data-Enabled Physics-Informed Neural Network (DEPINN). Unlike standard PINNs that rely solely on physical laws encoded in the loss function, DEPINN augments this with a small set of prior data points, typically obtained from experiments or simulations. This prior data helps to constrain the solution and accelerate convergence.
Key methodological contributions include:
- Adaptive Optimization Strategy: The authors propose an adaptive optimization method that intelligently switches between the Adam and LBFGS optimizers. This strategy aims to leverage Adam's initial efficiency and LBFGS's precision, preventing issues like premature convergence or oscillations.
- Loss Function Design: The loss function incorporates three components: residual loss (for PDE accuracy), boundary condition loss, and a prior data loss (SSE - Sum of Squared Errors). The paper investigates the impact of weighting these components.
- Comprehensive Numerical Study: The DEPINN is tested on a range of benchmark problems, from simple mono-energetic diffusion in spherical and cylindrical geometries to a complex two-group diffusion problem based on the 2D IAEA Benchmark Problem (IBP). This allows for a thorough evaluation of its performance across varying complexities.
Datasets
The paper utilizes data from three benchmark problems for training and evaluation:
- Finite Spherical Reactor: Modeled by a one-dimensional mono-energetic diffusion equation.
- Finite Cylindrical Reactor: Modeled by a two-dimensional mono-energetic diffusion equation.
- 2D IAEA Benchmark Problem (IBP): A realistic, two-dimensional two-group diffusion equation problem adapted from a practical nuclear reactor.
For the DEPINN, a "very small amount of prior data" is used. This data is described as being "simulated from experiments" or obtained from a "finite element solver." The exact number of prior points for the 2D IBP is stated as 38x2.
Key Insights and Ideas
- The Power of Prior Data: Even a small amount of prior data can dramatically improve the accuracy and convergence speed of PINNs for complex NDEPs. This is crucial for practical engineering applications where obtaining extensive experimental data might be infeasible.
- Adaptive Optimization is Key: Simply using Adam or LBFGS might not be optimal. The proposed adaptive strategy, which intelligently switches between them, significantly enhances training efficiency and robustness.
- Loss Function Weighting Matters: The relative importance of the residual, boundary, and prior data loss terms needs careful consideration. The paper shows that the prior data loss term (Lp) plays a critical role, especially when data is scarce.
- Generalization Potential: The DEPINN, once trained, shows potential for generalization to slightly different parameter settings, a desirable trait for engineering tools.
- Complexity Trade-offs: While DEPINN excels, the complexity of the problem (e.g., two-group vs. mono-energetic, complex geometry) does impact accuracy, though it still meets engineering acceptance criteria for the 2D IBP.
Key Results
- Engineering Accuracy Achieved: The DEPINN, with its adaptive optimization and prior data, successfully solves NDEPs for complex geometries and two-group diffusion equations, meeting engineering acceptance criteria.
- Improved Accuracy over Standard PINN: For the mono-energetic cases, the DEPINN outperforms previous PINN-based methods reported in the literature in terms of MSE and APE.
- Adaptive Optimizer Performance: Strategy-3 (adaptive Adam-LBFGS) consistently outperforms fixed-epoch Adam or LBFGS strategies, leading to better results with efficient training.
- Sensitivity to Sampling: The placement of prior data points significantly impacts performance. Fixing prior points in regions with less severe coefficient changes and allowing others to adapt proved beneficial.
- 2D IBP Performance: For the 2D IBP, the DEPINN achieves relative errors in flux (u, v) and eigenvalue (keff) that satisfy engineering acceptance criteria (e.g., relative error of keff < 0.005).
Limitations and Future Work
- Data Requirements: While DEPINN reduces the need for extensive data, the paper acknowledges that "different materials, thus more training points or prior data are necessary for further improving the accuracy."
- Computational Cost: The paper notes that simulation time is "almost proportional to the sample rates," suggesting that for very high sampling rates, computational cost can still be a factor.
- Generalization to 3D: The current work focuses on 2D problems. Future work should include testing on "three-dimensional, large-scale reactor physics problems."
- Noise and Uncertainty: The paper suggests exploring "performance on noise observations" and developing "new PINN structures for real complex engineering problems where no observation is available."
- Physical Interoperability: Further investigation is needed to "explain the physical interoperability of the network."
Summary updated