Using Machine Learning Methods to Predict Bias in Nuclear Criticality Safety

Journal Article · 28 March 2019 · Journal of Computational and Theoretical Transport

DOI: https://doi.org/10.1080/23324309.2019.1585877 · OSTI ID:1544713

Grechanuk, Pavel Aleksandrovi ^[1]; Rising, Michael Evan  ^[2]; Palmer, Todd ^[1]

---

1. Oregon State Univ., Corvallis, OR (United States)
2. Los Alamos National Lab. (LANL), Los Alamos, NM (United States)

The manuscript provides a description of the application of machine-learning (ML) tools to the prediction of bias in criticality safety analysis. In particular, a set of over 1000 experiments included in the Whisper package were fed into a variety of ML algorithms (notably Random Forest and AdaBoost) implemented in SciKit-Learn using k-eigenvalue sensitivities (with and without energy dependence) for individual nuclides, and optionally, the simulated k_eff as the training features. Ultimately, the ML model was used to predict the bias (k_sim - k_exp). The results indicate that use of energy-integrated sensitivity profiles with k_sim as training features led to the best predictions as quantified by root-mean square and mean absolute errors. In particular, the best-case estimates came from AdaBoost, with a mean absolute error of 0.00174, which is less than the mean experimental uncertainty of 0.00328 for the experiments included.

• Research Organization: Los Alamos National Lab. (LANL), Los Alamos, NM (United States)
• Sponsoring Organization: USDOE National Nuclear Security Administration (NNSA)
• Grant/Contract Number: 89233218CNA000001
• OSTI ID: 1544713
• Report Number(s): LA-UR--18-24800
• Journal Information: Journal of Computational and Theoretical Transport, Journal Name: Journal of Computational and Theoretical Transport Journal Issue: 4-6 Vol. 47; ISSN 2332-4309
• Publisher: Taylor and FrancisCopyright Statement
• Country of Publication: United States
• Language: English