Uncertainty Quantification of Classifying Mechanical Breach in the Crash and Burn Problem

Technical Report · Sat Jun 01 04:00:00 EDT 2019

DOI:https://doi.org/10.2172/1762365· OSTI ID:1762365

Frankel, Ari Louis ^[1]

---

1. Sandia National Lab. (SNL-NM), Albuquerque, NM (United States)

The prediction of mechanical breach in the Crash and Burn problem is important for designing the full system to minimize the probability of loss of containment. Due to uncertainties in the impact angle, material properties, and material models, it is not possible to define an exact critical impact speed at which the system breaches. Furthermore, the cost of running a large scale sampling study to determine the empirical probability of breach is prohibitive. In this work, surrogate models from machine learning, namely logistic regression and artificial neural networks, are introduced to predict binary classification of pass versus breach from a limited set of samples. The structure and calibration of these classifiers is discussed, and a set of metrics for describing the performance of the classifiers is introduced. The classifiers are used on the UUR version of the Crash model to attempt prediction of failure probability and perform variable sensitivity analysis. Where a single sample of the computational model can take hundreds of CPU-hours, training and evaluating a classifier can take seconds or less, thus giving high predictive power in a relatively short time.

• Research Organization: Sandia National Laboratories (SNL-CA), Livermore, CA (United States)
• Sponsoring Organization: USDOE National Nuclear Security Administration (NNSA)
• DOE Contract Number: AC04-94AL85000; NA0003525
• OSTI ID: 1762365
• Report Number(s): SAND--2019-4786; 675091
• Country of Publication: United States
• Language: English

Related Subjects

42 ENGINEERING