Abstract
Since it is difficult to directly measure neutron energy, the neutron spectrum is estimated from the specific responses of the neutron reaction. Although several algorithms have been proposed to reconstruct the neutron spectrum, no attempt has been made to apply a multi-objective optimization technique. This study is based on the idea that reconstructing the spectrum by taking into consideration various prior information simultaneously enables to obtain more reasonable results. The genetic multi-objective optimization scheme was applied to derive the Pareto front of spectrum from activation foil responses. The two objectives of maximizing the Shannon information entropy of the neutron spectrum and minimizing the relative error of the responses were considered. By applying the algorithm, we were able to successfully reduce the solution candidates to Pareto front and improve the validity of the unfolded spectrum.
Introduction
In nuclear engineering and radiation safety, determining the energy spectrum of neutrons is a fundamental challenge. Because neutron energy cannot be measured directly, researchers rely on "activation foils"—materials with well-characterized reaction cross-sections. By exposing these foils to a neutron field and measuring the resulting radioactivity, scientists attempt to reconstruct the original energy distribution. This process, known as "neutron spectrum unfolding," is a classic ill-posed inverse problem where multiple theoretical spectra can satisfy the same experimental data. Traditional unfolding methods often focus on a single objective, which can lead to unphysical oscillations or overfitting. This paper explores a Multi-Objective Genetic Algorithm (MOGA) approach to find a more robust set of solutions.
Methodology
The researchers implemented a MOGA to evolve a population of candidate spectra. Rather than evolving the spectrum directly, the algorithm optimizes a "shifting function"—a third-order polynomial—that modifies an initial guess spectrum. This approach naturally enforces smoothness, preventing the jagged results common in other methods.
Optimization Objectives
The algorithm evaluates candidates based on two competing criteria:
- Minimizing Relative Error: Ensuring the calculated response of the foils matches the simulated measurements.
- Maximizing Shannon Entropy: Based on the Maximum Entropy Principle, this ensures the spectrum is the most unbiased distribution consistent with the data.
Experimental Setup
The study utilized a numerical simulation to validate the MOGA approach. Using MCNPX 2.7 and the ENDF/B-VII nuclear data library, the team simulated a parallel neutron beam entering water. To test the algorithm's robustness, they intentionally introduced a 20% error in the initial guess spectrum by artificially increasing the simulated water density. Seven specific nuclear reactions were used as the "measured" data points for the unfolding process.
Key Results
After 300 generations with a population of 500, the algorithm successfully converged on a stable Pareto front. The Pareto front represents the set of optimal solutions where one objective cannot be improved without degrading the other.
Despite the significantly flawed initial guess, the MOGA successfully reconstructed a spectrum that closely matched the "true" simulated distribution. Interestingly, the solutions on the Pareto front exhibited very similar entropy values, suggesting that for this specific configuration, minimizing relative error was the primary driver for improving spectral validity. The use of the polynomial shifting function effectively filtered the infinite search space into a manageable set of high-quality candidates.
Conclusion and Future Work
This study demonstrates that multi-objective optimization is a viable and powerful tool for neutron spectrum unfolding. By balancing measurement accuracy with information entropy, the MOGA provides a physically sound way to navigate the ambiguities of inverse problems. Future work will focus on applying this method to real-world experimental data, where measurement noise and unpredictable uncertainties will further test the algorithm's resilience.
Summary updated