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

Event Sponsor: 
Mathematics and Computer Science Division - LANS Seminar
Start Date: 
Aug 18 2017 - 1:00pm
Building/Room: 
Building 240/Room 1404 - 1405
Location: 
Argonne National Laboratory
Speaker(s): 
Various Speakers, Summer Students in LANS (MCS Division)

Session I
Chair: Julie Bessac

Gady, Sarah 1:00PM
Title: Stochastic Optimal Power Flow Applied to University of Chicago Campus
Abstract: Interest in cogeneration has risen considerably over the previous years due to both the possible financial and environmental benefits. Cogeneration technology produces heat and power simultaneously near the point of consumption, allowing for greater overall system efficiency and financial savings. While there are considerable benefits within this technology, one of the primary problems is how to design a system for variable loads and how to operate the technology optimally over a long time horizon. This work extends a previous model developed at Argonne, where five technologies are considered for the system: solid oxide fuel cells, combined heat and power solid oxide fuel cells, water storage tanks, water boilers, as well as lithium ion batteries. This model is extended to incorporate individual building electric demand profiles gathered from the University of Chicago. The profiles are clustered using k-medoids clustering, partitioning the buildings into 10 distinc t clusters. Electric demand profiles are generated for each building type in addition to the corresponding heating demand. Optimal cogeneration systems and operational profiles are presented based on the data clustering. In addition, the model is currently being extended to incorporate uncertainty within the demand profiles through multiple demand scenarios based on the different clusters.

Ghai, Aditi 1:15PM
Title: Multigrid Method for Tomographic Reconstruction

Abstract: Tomographic reconstruction is a non-invasive imaging technique allowing for understanding the internal structure of a 3D sample under consideration. Arising from inversion of a sequence of projections at multiple angles, the problem is set up via least squares optimization formulation and is usually very ill- conditioned and rank deficient. Our focus is to develop multigrid methods which can efficiently solve the sparse linear systems by experimenting with different coarsening, smoothening and regularization techniques. We present results related to algebraic multigrid with classical coarsening and other semi-coarsening techniques with different smoothers.

Kim, Youngseok 1:30PM
Title:
Abstract:

Moen, Connor 1:45PM
Title: Precipitation and Prediction

Abstract: As rainfall events in Chicago become increasingly intense and unpredictable, the damage caused by flooding increases as well; this project focuses on tackling the issue of predicting said flooding. My research involved the collection and analysis of precipitation data from available sensors and soil moisture data from the Thoreau Sensor Network. The main difficulty of the project was uncovering a reliable method to use precipitation as a predictor. With this method tested and verified, the next step is to create a framework for machine learning models that can be used to predict flooding once additional data can be made available.

Morgan, Hannah 2:00PM
Title: A stochastic performance model for pipelined Krylov methods

Abstract: In this work, we investigate the performance of Krylov and pipelined Krylov methods in a probabilistic setting under different workload distributions. We view runtimes as stochastic to account for the operating system noise present in computer simulations and compute the expected runtimes of Krylov and pipelined Krylov methods to measure their performance. We introduce the further complication of uneven workload distribution, a common occurrence in evolutionary problems.

Break 2:15PM - 2:30PM

Session II
Chair: Charlotte Haley

Kent, Carson 2:30PM
Title: Nonlinear robust optimization

Abstract: The parameters defining an optimization problem can often not be known with sufficient accuracy and, hence, are subject to a non-trivial amount of uncertainty. Further, in many practical settings, it has been shown that this uncertainty can result in a nominal solution which is highly infeasible with respect to some possible value of the uncertain parameters. Such a difficultly has given rise to the notion of a "robust" optimization, where feasibility of a solution is required with respect to all possible values of the uncertain parameters. In this work, we demonstrate an extension of robust optimization to the setting of nonlinear robust constraints in a manner that removes traditional regularity/convexity assumptions.

Sood, Kanika 2:45PM
Title: Analyzing HPC software Github repositories to identify and compute software productivity metrics

Abstract: Scientific software is rapidly growing with a focus on enhancing capabilities, accuracy and performance. However, software productivity has received insufficient attention during the development and maintenance of scientific software projects. In this work, we propose new time-dependent metrics that can quantify software productivity by mining HPC software GitHub repositories for issue tracking data and developer information. Using this data, we analyze the correlations between project issues and characteristics of the projects. The metrics can be used to better understand the trends of software development workflows and provide objective measurements of productivity. We demonstrate our approach on several HPC projects which have been in existence for more than three years.

Gupta, Shweta 3:00PM
Title: Analysis of Software Mailing Lists to Understand Productivity Issues

Abstract: Understanding various aspects of scientific software productivity can help developers to assess the impact of code changes over time and also to improve workflow. Email is a widely used communication platform among software developers and users because requirements and bugs can be expressed in any form. An important source of information is the analysis of mailing lists, requiring text mining due to the unstructured nature of email. I have used a statistical topic modeling algorithm, Latent Semantic Algorithm (LDA), to categorize and analyze over 20,000 emails for PETSc, with an objective of extracting metrics such as the time complexity of working on a new feature or bug fix, the total number of members addressing an issue, and other topics.

Shi, Yunong 3:15PM
Title: Using formal verification to facilitate the migration of scientific software

Abstract: We try to verify various numerical simulation codes, including Nekbone, burgers equation. These codes heavily rely on tensor calculus and floating-point arithmetic, which leads to large search space and non-associative property. To overcome these difficulties, we use state-of-the-art formal verification tools, CIVL and CodeThorn, to check the equivalence between codes optimized for two different architectures.

Miscellaneous Information: 

Refreshments will be served.

Please click here to add this event to your calendar. [schedule.ics]