Bayesian neural network evaluation method on the neutron-induced fission product yields of Thorium-232

Application of Bayesian Neural Networks for Thorium-232 Fission Yield Prediction
A research team led by Chun-Wang Ma has developed a Bayesian neural network framework to predict fission product yields of thorium-232 (²³²Th) under neutron irradiation. The approach incorporates nuclear physics principles—such as the odd-even effect and isospin symmetry —to enhance prediction reliability across neutron energy ranges where experimental data is sparse.

Addressing Nuclear Data Gaps
Current nuclear data libraries (e.g., JENDL, ENDF, CENDL, and JEFF) provide fission yield measurements for ²³²Th at only a few discrete neutron energies: thermal neutrons (0.0253 eV), 0.5 MeV, and 14 MeV. The scarcity of experimental data between these energies presents challenges for reactor design and safety analysis. The Bayesian neural network method offers a systematic solution to interpolate and extrapolate fission yields across the neutron energy spectrum while rigorously quantifying prediction uncertainties.

Physics-Informed Machine Learning Approach
The research team implemented a two-hidden-layer Bayesian neural network architecture with physical constraints deeply integrated into the modeling process. The framework incorporates nuclear physics principles, including the odd-even effect in proton numbers and isospin symmetry, as input features. This integration of physical constraints mitigates unphysical predictions and significantly improves the model's ability to reproduce fine structures in the mass yield distribution. Furthermore, the approach generates probabilistic predictions with quantified uncertainties, providing valuable data references for nuclear safety analysis.

Applications in Nuclear Technology
The predicted fission yields have direct practical applications for thorium-based reactor systems, notably China's Thorium Molten Salt Reactor program. Accurate fission yield data is crucial for supporting burn-up credit analysis, reactivity feedback assessment, and spent fuel characterization. Additionally, the methodology provides essential data for medical isotope production and reactor antineutrino spectrum calculations. Validations for isotopes such as ⁹⁵Zr, ⁹⁹Mo, ¹³²Te, and ¹³¹I demonstrate strong agreement with available experimental measurements across neutron energies from thermal to 14 MeV.

Future Research Directions
Looking ahead, the research team plans to extend the Bayesian framework to predict fission yields for other actinide nuclei and to develope methods for incorporating additional physical constraints. This approach represents an evolution in nuclear data evaluation methodology by effectively integrating machine learning capabilities with established physical principles. The work demonstrates that machine learning methods can powerfully complement traditional nuclear physics approaches when properly constrained by physical principles.

The Bayesian neural network framework provides a systematic method for nuclear data evaluation that quantifies both prediction values and their associated uncertainties. By addressing specific gaps in existing nuclear databases while maintaining consistency with established physical principles, this methodology offers a highly promising approach for future nuclear data assessment efforts.