NE Seminar: “Combining Predictive Modeling and Machine Learning to Search for Forensic Signatures in Nuclear Fallout”

Date(s) - 10/13/2022
1:55 pm - 2:55 pm

Rhines Hall 125


John Mattingly, Ph.D.

Professor, Nuclear Engineering
NC State University

Dr. John Mattingly is a Professor of Nuclear Engineering (NE) at North Carolina State University (NCSU), where he has worked since 2011. At NCSU, John directs a team of graduate students conducting research on applications of neutron and gamma radiation detection, imaging, inverse analysis, and machine learning to nuclear nonproliferation, emergency response, and forensics. He served as the Chief Scientist and Principal Investigator of the National Nuclear Security Administration’s (NNSA’s) Consortium for Nonproliferation Enabling Capabilities (CNEC), where he directed and coordinated the research of professors and students at ten universities and scientists at four national laboratories, who worked to develop new technologies and policies to support the next generation of proliferation detection and deterrence capabilities. Prior to joining the NCSU faculty, John worked at Sandia National Laboratories from 2003 to 2011, where he served as an emergency response and forensic analyst, and at Oak Ridge National Laboratory from 1997 to 2003, where he developed active neutron measurements for nonproliferation and arms control. He earned his Ph.D. in NE from University of Tennessee in 1998.


One objective of forensic analysis of post-detonation nuclear debris, a.k.a. fallout, is to assess the nuclear explosive’s degree of sophistication. Classification of the explosive type as either a primitive fission device or a more advanced device employing fusion boosting has forensic value: it can narrow the field of the explosive’s potential origins. However, classifying explosive type based on fallout composition is challenging for several reasons. Fission products vaporized in the explosion form compounds (principally oxides) that exhibit different condensation and diffusion rates, so they are fractionated as they condense and diffuse into fallout particles as they coalesce, and fractionation significantly alters fallout composition.

Few measurements of actual fallout composition have ever been conducted, and most of those measurements were done on fallout from US atmospheric nuclear tests in remote areas in the South Pacific and Nevada in the 1950s and 1960s. Furthermore, it is challenging to reproduce the fractionation process in controlled, laboratory experiments. Consequently, forensic analysts rely in part on predictive models of fallout fractionation, but those models are approximate, they are computationally expensive, and their parameters exhibit substantial uncertainty.

NC State has developed methods to efficiently propagate uncertainty in fallout fractionation models onto predictions of fallout composition using techniques borrowed from machine learning (ML): reduced-order modeling using active subspace decomposition (ASD) and surrogate modeling using Gaussian-process regression (GPR). These developments can enable forensic analysts to identify combinations of fission products that classify explosive type, even when the fractionation model parameters exhibit large uncertainties.

John Mattingly will describe the physical processes governing fallout formation, predictive models of fallout fractionation and composition, and ML methods to classify explosive type from fallout composition.