The (γ, n) cross-section is important in nuclear engineering transport calculations. The measurements of the (γ, n) reaction for some isotopes show significant discrepancies from different laboratories. Since experimental data analysis is the first tasks in nuclear data evaluation, identifying outlier data in measurements is crucial for improving the quality of nuclear data. Therefore, this work employs Variational AutoEncoder (VAE) approach to analyze experimental measurements of (γ, n) cross sections for nuclear mass from 29 to 207, aiming to provide more reliable experimental information for nuclear data evaluation.
Based on the proton Z and nuclear mass A, we constructed a Variational AutoEncoder network designed for outliers identification in measurement of (γ, n). The silhouette coeffcient method and K-Means algorithm were used to perform clustering the latent variables of VAE. Subsequently, the experimental data with and without the outliers were compared with the IAEA-2019-PD to assess VAE in application of photoneutron measurements evaluation.
The results demonstrate that VAE can effectively identify outliers in the measurements of (γ, n). After excluding outliers, the (γ, n) cross-section for 54Fe, 63Cu, 181Ta, 206Pb and 207Pb showed higher consistency with the IAEA-2019-PD evaluation results. However, 29Si and 141Pr deviated from the IAEA- 2019-PD evaluation results yet, which requires more analysis to the measurements itself in future.
The Variational AutoEncoder method effectively identifies outliers and mines the latent structures in experimental data of (γ, n) reaction. It provides more reliable experimental information for nuclear data evaluation and validating the potential application of this method in nuclear data research. However, generalization capability of Variational AutoEncoder still needs further developed especially the issues with uneven energy distribution for various measurements.