AI Meets Nuclear Physics: Toward More Accurate Photonuclear Cross Sections
A research team from the Shanghai Institute of Applied Physics, Chinese Academy of Sciences, along with collaborators, has applied Bayesian neural networks (BNNs) to improve the fitting of photonuclear (γ,n) cross-sections. By training a neural network with two hidden layers on a consistent experimental dataset, the team achieved reliable predictions while avoiding overfitting. Their analysis of absolute and relative errors confirmed that the model learned the data efficiently.
Compared to the traditional TENDL-2021 database, the BNN approach showed better accuracy in modeling low-energy thresholds, the giant dipole resonance (GDR) peak, and the high-energy tails of cross-sections. Notably, BNNs performed well even when data was sparse or biased, offering trustworthy predictions for cross-sections that are difficult or impossible to measure directly. The study also emphasized the need for consistent training data by examining differences between datasets from various laboratories. This method is expected to support future research at the Shanghai Laser Electron Gamma Source (SLEGS) beamline in fields such as nuclear astrophysics, nuclear material science, and radiation protection.
Enhanced Accuracy and Predictive Power
The BNN model consistently outperformed traditional evaluations like TENDL-2021, achieving lower average absolute errors and stable predictions across multiple nuclides. It also accurately predicted (γ, n) cross-sections for nuclei not included in the training data. This capability is particularly valuable for estimating nuclear reaction data of unstable or experimentally inaccessible isotopes, which are vital for studies of the r-process in astrophysics.
Quantifying Systematic Discrepancies Across Laboratories
By comparing datasets from Lawrence Livermore National Laboratory (LLNL) and Saclay, the BNN method identified systematic differences and estimated potential biases. This analysis provides a useful tool for standardizing nuclear data and improving the reliability of experimental results across different facilities.
Guidance for Future SLEGS Experiments
This research offers valuable support for upcoming precision measurements at the SLEGS beamline in Shanghai. The improved photonuclear cross-section data will enable more focused experiments and help validate theoretical models, enhancing the quality and impact of photonuclear research.
Access the full study via DOI: 10.1007/s41365-024-01611-1.
Your membership also unlocks:
