Bayesian neural net sharpens thorium 232 fission yield data

by Robert Schreiber
Berlin, Germany (SPX) Jan 13, 2026

A research team led by Chun-Wang Ma has developed a Bayesian neural network framework to predict independent and cumulative fission product yields for thorium 232 under neutron irradiation across a broad energy range. The work targets longstanding gaps and inconsistencies in evaluated nuclear data libraries, which currently provide thorium 232 fission yields only at a few discrete neutron energies.

Current evaluations in major nuclear data libraries such as JENDL, ENDF, CENDL and JEFF list fission yields for thorium 232 primarily at thermal neutron energy near 0.0253 electronvolts, at 0.5 megaelectronvolts and at 14 megaelectronvolts. Between these energies, experimental measurements are sparse, which complicates nuclear reactor design, safety assessment and fuel cycle analysis for thorium based systems.

The new Bayesian neural network approach offers a systematic method to interpolate and extrapolate fission yields between and beyond the measured points while rigorously quantifying prediction uncertainties. The model produces probabilistic yield distributions that supply both central values and uncertainty bands, which are essential for advanced reactor calculations and safety margins.

The team implemented a two hidden layer Bayesian neural network architecture and embedded nuclear physics concepts directly into the model inputs and structure. In particular, the framework incorporates effects related to odd and even proton numbers and isospin symmetry, which strongly influence the structure of fission product distributions. These physics informed features help suppress unphysical behavior and improve the reproduction of fine structures in the mass yield curves.

By constraining the neural network with established physical principles, the researchers report a significant enhancement in the model's ability to match evaluated and experimental data where it exists. The Bayesian framework not only reproduces the main features of the fission yield distributions but also refines the description of detailed mass dependent patterns across different neutron energies.

The predicted fission yields have direct relevance for the design and operation of thorium fueled reactors, including programs such as China's thorium molten salt reactor effort. More accurate and complete fission yield data support burn up credit analysis, reactivity feedback evaluation and the characterization of spent fuel inventories over the reactor lifetime.

Beyond reactor physics, the authors note that improved thorium 232 fission yields can benefit medical isotope production and calculations of reactor antineutrino spectra. The Bayesian neural network results for specific isotopes, including zirconium 95, molybdenum 99, tellurium 132 and iodine 131, show strong agreement with available experimental data from thermal up to 14 megaelectronvolt neutron energies.

The study demonstrates that Bayesian neural networks can serve as a powerful complement to traditional nuclear data evaluation techniques when they are tightly linked to underlying physics. By combining machine learning flexibility with nuclear structure and reaction constraints, the method yields predictions that are both quantitatively precise and physically consistent.

Looking ahead, the research team plans to extend the Bayesian framework to other actinide nuclides beyond thorium 232. They also intend to incorporate additional physical constraints and observables to further refine the predictive capability and robustness of the approach.

The authors argue that this physics constrained machine learning methodology marks an evolution in nuclear data evaluation practices. It offers a structured way to address specific gaps in existing databases while preserving compatibility with well established nuclear models and measurements.

By providing high precision fission yield predictions together with quantified uncertainties, the Bayesian neural network framework delivers foundational data for future fourth generation nuclear energy systems. It also offers tools for nuclear waste transmutation studies and advanced fuel cycle optimization involving thorium based materials.

The work appears in the journal Nuclear Science and Techniques under the title "Bayesian neural network evaluation method on the neutron induced fission product yields of 232Th." The published article presents detailed comparisons between the Bayesian predictions, evaluated data libraries and experimental results across multiple neutron energies.

Research Report:Bayesian neural network evaluation method on the neutron-induced fission product yields of 232Th