Mechanical and Civil Engineering Seminar
Machine Learning approaches for predicting Seismic, Acoustic and Atmospheric Data
Abstract: Over the last decade, machine learning (ML) approaches have been increasingly relied upon for predicting responses of interest and quantifying uncertainties. The strengths of ML approaches are their speed and the ever-increasing ability to predict new scenarios where experimental/simulation data is not available. For this purpose, selecting the right approach for a specific application requires previous knowledge of the characteristics of the data. For example, is it a time series? do we have multiple outputs? is the data sparse? At Lawrence Livermore National Laboratory ML approaches have been applied to accelerate predictions for applications where accurate, fast responses are a priority. This presentation will cover the following problems: 1) K-nearest neighbors (k-NN) based approach that predicts the deposition plume after of a contaminant release given weather conditions. The technique is enhanced using imputation, rotation, and translation.2) Gradient boosting, random forests, neural networks, and other algorithms to estimate probability distribution functions of atmospheric sources from sensor measurements for multiple atmospheric models at scales ranging from 100's of meters to 1000's of kilometers. 3) Deep-learning network (DLN) to emulate seismic-phase travel times that are predicted using a global 3-dimensional earth model. The trained DLN can accurately predict seismic-phase travel times between an arbitrary event and station locations. 4) Two branch network multi-fidelity approach that combinesdeep learning and physics based models to identify earthquakes from other events recorded on seismic stations. This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.
Bio: Giselle is a simulation data scientist for Above and Below-Ground Physics in the Atmospheric Science Research and Applications group at the Lawrence Livermore National Laboratory's Physical Life Science directorate. Her current work focuses on machine learning and uncertainty quantification approaches applied to atmospheric flow, transport & hazard assessment. Giselle received her bachelor's degree in nuclear engineering from Balseiro Institute, San Carlos de Bariloche, Rio Negro, Argentina in June 2014. Her thesis was focused on the surveillance program for the Argentinian nuclear reactor CAREM 25. She received her master's degree in mechanical engineering from the University of Florida in December 2016. She received her Ph.D. in aerospace engineering from the University of Florida in December 2018. She was under the supervision of Dr. Raphael T. Haftka and Dr. S. Balachandar. Her thesis topic was on the quantification of particle departure from axisymmetry in multiphase cylindricaldetonations. In 2019, she was a postdoctoral research associate in the Verification and Analysis Group, X Computational Group at Los Alamos National Laboratory working on machine learning applied to fracture mechanics and inertial confinement fusion physics