Development of a novel machine learning-based adaptive resampling algorithm for nuclear data processing

Efficient processing of nuclear cross-sections data is critical for advanced reactor physics and safety assessments. Existing workflows of using nuclear data in Hierarchical Data Format, version 5 (HDF5 format) rely on intermediate file formats, such as A Compact ENDF (ACE) files generated via NJOY, which introduce inefficiencies in nuclear data processing. This work presents two novel computational techniques that streamline nuclear data processing and modification. First, a machine learning-based resampling algorithm is presented for nuclear cross-section data stored in HDF5 format, designed to intelligently retain critical threshold points while reducing data redundancy. Second, a direct HDF5 modification framework is introduced, eliminating the need for legacy file conversion steps and enabling direct edits to OpenMC-compatible nuclear data libraries. This methodology employs an adaptive resampling strategy that dynamically adjusts point densities across diverse neutron energy regions, preserving resonance structures and threshold behaviors while achieving significant data compression. Benchmarking against established models-such as K-Nearest Neighbors and Gaussian Processes-indicates that the ML-based approach offers lower errors and enhanced computational efficiency. This integrated framework improves nuclear data accessibility and expedites simulations, reactor core design, uncertainty quantification, and neutronics analysis.