FECSG-ML: Feature Engineering for Nuclear Reaction Cross Sections Generation Using Machine Learning.

Area of Science:

• Nuclear science and engineering.
• Computational physics and data science.

Background:

• Accurate nuclear data, including cross sections, are vital for nuclear reactor design and safety.
• Existing nuclear databases like EXFOR have limitations due to scarce, discrepant, or erroneous experimental data requiring manual evaluation.
• Manual evaluation of nuclear data is prone to biases and uncertainties.

Purpose of the Study:

• To develop a machine learning framework (FECSG-ML) for generating nuclear reaction cross sections data.
• To provide a data-driven alternative to manual evaluation of nuclear databases.
• To improve the accuracy and reliability of nuclear cross sections data for scientific applications.

Main Methods:

• Utilized transfer learning by pre-training on the ENDF/B-VIII.0 dataset and fine-tuning with the EXFOR database.
• Employed machine learning to convert discrete cross sections data into a continuous format for various isotopes.
• Integrated ensemble learning for optimizing the prediction of multiple cross sections data sets.

Main Results:

• The FECSG-ML model demonstrated high accuracy, with generated regression curves closely matching EXFOR data points.
• The model outperformed the ENDF/B-VIII.0 evaluation database in terms of accuracy.
• Generated nuclear cross sections data were successfully applied in OpenMC simulations for pin fuel assemblies and CANDU reactors.

Conclusions:

• The FECSG-ML framework offers a robust method for generating reliable nuclear cross sections data.
• This approach effectively substitutes traditional, potentially biased, nuclear database evaluation processes.
• The study highlights the potential of machine learning to advance nuclear science and engineering through improved data generation.