LANS SASSy - Summer Argonne Students' Symposium 2017 - PART I

Event Sponsor: 
Mathematics and Computer Science Division - LANS Seminar
Start Date: 
Jul 28 2017 - 1:00pm
Building/Room: 
Building 240/Room 1406 - 1407
Location: 
Argonne National Laboratory
Speaker(s): 
Various Speakers, Summer Students in LANS (MCS Division)

Summer students in LANS give talks on their research.

Session I
Chair: Kibaek Kim

Fethi, Amal 1:00PM
Title: SPUDS: Smart Pipeline for Urban Data Science
Abstract: SPUDS is an automated machine learning pipeline developed to help the City of Chicago to process its restaurants inspections data, and plan the next inspections efficiently. The software performs different preprocessing tasks, but its main role is to optimize the parameters of a machine learning algorithm to fit the best the input data. We investigated two issues: the metric used to evaluate a model’s performance, and the influence of data balancing, and will propose a strategy for each case. I will then show how SPUDS allows us to plan the inspections better than City of Chicago’s team.

Kabre, Julienne 1:15PM
Title: A Physics-Preserving Scheme for the Poisson-Nernst-Planck Equations
Abstract: This project is about the design of a 1 D finite difference energy-preserving discretization for the Poisson-Nernst-Planck Equations. The scheme is of second accurate in both space and time. Comparisons are made between this energy dynamics preserving scheme and a standard finite difference scheme, showing a difference in satisfying the energy law. Numerical results are presented.

Buranosky, Matthew 1:30PM
Title: AWA Optimization Problems
Abstract: In this talk, we discuss a numerical approach to determining a global emittance minimum of an rf photoinjector. We focus on data sets obtained from an OPAL simulation code used to predict the objective values returned from bound-constrained sets of parameters. We visually present the effects on global emittance that result from adjusting the parameters individually across a given domain while holding the others constant around a base set. Using a linear regression model, we take steps to figuring out the range of variable values in which the OPAL simulation code returns values that cannot be trusted.

Villalobos, German 1:45PM
Title: Parameter Optimization for the HFBTHO program
Abstract: The UNEDF project focuses on gaining a better understanding of the interaction of nuclear particles, combining knowledge from various backgrounds including physics, mathematics, and computer science. The HFBTHO program, built in Fortran, was developed to study a various types of nuclei. The goal of our work is to optimize the process of evaluating various residuals so as to find the best possible parameters which will yield minimal deviation from actual experimental results. This presentation discusses the steps taken to find optimal parameters, including making use of the libEnsemble program to perform uniform random sampling.

Lan, Yu-Hsiang 2:00PM
Title: Regularization in Tomographic Reconstruction
Abstract: The Poisson-Nernst-Planck (PNP) equation is a well-known model of ion transport describing many physical and biological phenomena. Due to ionic sizes, steric repulsion may appear in crowded ions of several biological systems, so the classical PNP equations become unreliable since the PNP equations represent ions as point particles of size zero. We consider a modified PNP model that can potentially include the steric repulsion ion-size effect and the selectivity of important types of calcium and sodium ion channels.

I'll present spectral element schemes and implementation for solving the PNP-steric models, and discuss implementation of unsteady and steady-state PNP solvers in NekCEM including the treatments of Dirichlet, Neumann, Robin boundary conditions. Also, I'll demonstrate convergence studies for validating our schemes, provided with some preliminary results of the dynamics of charge concentrations on a real protein structure KcsA.

Break 2:15 - 2:30

Session II
Chair: Prasanna Balaprakash

Huang, Xiao 2:30PM
Title: Optimize the ppa price of NREL simulated solar power plant
Abstract: The aim of this research is to analyze a concentrated solar power plant simulation code produced by NREL, and specifically, to minimize the ppa (power purchase agreement) price of the solar system by tuning 10 operating parameters. The study we did was in three aspects. Firstly, because the functional relationship is unknown to us, we did some exploration over the parametric space to study the smoothness and stochasticity of the ppa price as a function of the parameters. Secondly, due to the expensive functional evaluation, we built an emulator model using gradient boosting that can evaluate the ppa price much faster at the price of a tolerable deviation. This emulator has a by-product of evaluate the importance of each parameter. Lastly, we will try different optimization methods to this problem, including Bayesian global optimization and model-based derivative-free trust region method.

Tripathy, Rohit 2:45PM
Title: Using deep neural networks for wind speed forecasting
Abstract: Forecasting wind speed is important from the perspective of numerous wind engineering applications such as design and placement of wind turbines. In this project, we adopt ideas from deep learning to construct a spatio-temporal model for wind speeds. The wind speed observations are obtained from a Doppler LiDAR sensor in Oklahoma. We also consider strategies to fuse information from numerical weather prediction (NWP) models, so as to negate the effect of noisy field measurements. Specifically, we demonstrate the application of long short term memory (LSTM) RNNs forecast wind speeds in the short / medium term.

Wagner, Riley 3:00PM
Title: ADOL-C Python Test Suite Performance Measurement
Abstract: Algorithmic differentiation (AD) is a technique used to compute the derivative of a function in computer programming at machine precision. ADOL-C is an AD tool for functions written in C++ code. Recently, ADOL-C has been interfaced with Python using SWIG so that functions written in Python may be differentiated using the tool. My research has been centered around a test suite of Python function examples written in the summer of 2015. I have rewritten the test suite into a class system that includes the ability to time the function and record to a CSV file statistics involving the runtime with various lengths for the array input(s). It also includes the ability to trace the function using ADOL-C and compute the gradient, which are timed as well. I will talk about how this new format prevents repetitive code and improves organization, as well as the insights I have made in observing the gathered data graphically.

Patel, Vivak 3:15PM
Title: Structural Identifiability with Stochastic Inputs
Abstract: With applications ranging from pharmacology to power system operations, structural or system identifiability is a common problem when analyzing data related by a dynamical system. Fortunately, when the system input is controllable, structural identifiability can be established by verifying straightforward, computable conditions. However, when the system input is stochastic, the previous checks for structural identifiability do not generalize. In this work, we establish necessary and sufficient conditions for structural identifiability with partial measurements and stochastic input for a specific component of a linear dynamical system, whose selection is motivated by problems in power systems operations.

Miscellaneous Information: 

Refreshments will be served.

Please click below to add this event to your calendar.

[schedule.ics]