Mentors

Brooke Belisle & Alan Calder

Art, Computational Astrophysics

Progress in understanding astrophysical phenomena commonly involves modeling processes that can not be directly observed, and visualization is a critical part of obtaining insight from the data. But how does our understanding of reality rest on the particular tools and techniques that we use to visualize it? For this summer's research experience, astrophysicist Alan Calder and media aesthetics scholar Brooke Belisle will collaboratively mentor students who are interested in both the technical and aesthetic dimensions of scientific visualization. This could involve creative approaches to visualizing data sets from Calder's research; using public sources (star catalogs, NASA archives) to build a compelling visualization; comparing historical methods or alternative techniques for astronomical visualization; or creating a cinematic or immersive media project (as used to depict space in movies, museum installations, and artworks). Students will have the chance to learn about techniques and problems of astronomical simulation and to experiment with its visual possibilities.

Yong Chen

School of Marine & Atmospheric Sciences

Hudson River Estuary (HRE) provides critical spawning and nursery habitats for many fish species of significantly ecological and economical importance. The Hudson River Biological Monitoring Program (HRBMP) is a comprehensive fish survey program conducted on the HRE from the Troy Dam in Albany to Battery Park since 1974. It includes a suite of four river wide surveys targeting early life history stages, from eggs to juveniles, of all finfish species and river strata along with water quality data. The historical HRBMP dataset, managed by Dr. Chen’s Lab in the School of Marine and Atmospheric Sciences, provides a great opportunity to improve our understanding of the dynamics of fish populations and communities in the HRE. Several invasive species have been identified in the HRBMP and some of them have had significant impacts on the HRE ecosystem dynamics. For this project, we plan to use this rich historical dataset to evaluate and model the spatio-temporal dynamics of invasive species in the HRE and identify key environmental drives that might regulate the dynamics. We will quantify the impacts of climate-induced changes on the dynamics of invasive species in the HRE ecosystem and develop recommendations for the HRE management and conservation.

Rezaul Chowdhury

Computer Science

FFT-based Fast Stencil Computations on Ookami.  A stencil is a pattern used to compute the value of a cell in a spatial grid at some time step from the values of nearby cells at prior time steps. A stencil computation applies a given stencil to the cells in a spatial grid for some given number of timesteps. Stencils are used in a wide variety of fields across industry and scientific computing, including the simulation of physical systems, traffic flows, meteorology, stochastic and fractional differential equations, chemistry, modeling of erosion, fluid dynamics, quantitative finance, and even cellular automata. Stencils often originate from discretization of partial differential equations or PDEs, which lie at the center of many modern computational science and engineering fields. All currently available stencil algorithms that can accept arbitrary linear stencils perform work linear in both N and T, where N is the number of cells in the spatial grid and T is the number of timesteps. Very recently we have designed improved algorithms for linear stencils based on Fast Fourier Transforms (FFT) which perform work sublinear in either N or T while keeping the dependence on the other parameter unchanged. Implementations of these algorithms outperform existing fastest stencil implementations on Intel KNL (Knights Landing) and Skylake processors. The goal of the current project is to explore how the key features of Ookami, such as SVE and HBM, can be used to further improve the performance of our stencil algorithms. We also plan to benchmark GPU implementations of our algorithms on the Ookami GPU node.

Yuefan Deng

Applied Mathematics & Statistics

A group of graduate students, postdocs, and junior faculty in Prof. Yuefan Deng's group focus their research on developing algorithms of machine learning, multi-scale modeling, parallel computing for the latest supercomputer architectures including the IBM Power9 based Summit to simulate the SARS-CoV-2 as well as human platelets.  We hope to broaden the applications of the algorithms and software that we call IPDyna to other proteins and other cells.  Welcome REU researchers to our group.

Marivi Fernandez-Serra

Physics & Astronomy

Machine Learning Density Functional. The student will evaluate the importance of data sampling in optimization of exchange and correlation functionals in density functional theory. For this, the student will learn how to use the code psych to generate accurate data and then will use our XC-Diff code to optimize the exchange and correlation functional. The goal of the project will be to assess the differences between imposing constraints in the transformation of the data before entering the neural network versus imposing constraints directly on the data. The project will also evaluate how models trained in molecular systems generalize for solids and look for optimal generalization strategies.

Jason Jones

Computational Social Science

People are sometimes asked to describe themselves with a short text.  When creating a profile on Twitter, for example, the user is asked to "Describe yourself in 160 characters or less."  Dr. Jones' lab uses such short text to study personally expressed identity - i.e. which words individuals choose to describe themselves.  DCD students will analyze millions of observations of public social media profiles to study changes in personally expressed identity over time.  Dr. Jones' lab has observed that the use of political words (e.g. conservative, progressive) is increasing while other words such as religious words (e.g. Christian, atheist) remain at a steady prevalence.  Students will compute and visualize networks of words and time series of word prevalence to examine the temporal dynamics of self-conception and self-presentation.

Benjamin Levine

Chemistry

The design of materials for use in energy conversion and storage applications is essential to a sustainable future. Design efforts are most efficient when guided by knowledge of the relationship between the atomistic structure of a material and its function. Computational electronic structure theory enables researchers to discover such relationships without ever setting foot in a traditional chemistry laboratory. The Levine group works to develop and apply advanced computational techniques for determining the properties of materials for solar energy conversion and other applications. An undergraduate student in the Levine group will utilize computational electronic structure methods to search “chemical space” for molecules with desirable properties for application in energy conversion or storage. Toward this end, the student will benchmark the accuracy of ab initio methods for this purpose, in the process gaining familiarity with the approximate nature of quantum chemical calculations. Then they will choose a series of molecules to study in order to identify structure- function relationships. In the process, they will learn about the application of high performance computers to chemical problems and the microscopic physical processes that underly important
applications in energy applications.

Heather Lynch

Ecology & Evolution

This project aims to understand the biophysical processes that drive prey availability for penguins, and as a result could impact penguin biogeography on the West Antarctic Peninsula. Utilizing 3-D physical ocean model simulations in conjunction with penguin colony presence/absences, satellite-based diet estimates, colony reproductive success estimates, and colony counts, our goal is to determine if penguin utilize areas of high accumulation and/or retention of prey-like species near their colonies along the West Antarctic Peninsula. From these relationships, we hope to develop a habitat suitability model to predict future penguin colony locations and persistence. Potential DCD student projects include, but are not limited to, comparison of simulated and observed prey distributions; examining prey pathways around the peninsula, designing and analyzing results from model experiments based on observed prey fields; calculating residence times based on bathymetric features, penguin colony location, or foraging behavior; and assisting in habitat suitability model development. Students will work primarily with model output so experience working with NETCDF files and/or large (10GB +) datasets and automating analyses would be beneficial but is not required. Students should have experience in at least one programming language (R preferred but others are welcome).

Georgios Moutsanidis

Civil Engineering

Civil infrastructure is often susceptible to extreme actions, such as blast loads, hypervelocity impacts, or hydrodynamic events like tsunamis and coastal floods. Accurate and robust simulation of such actions is crucial for the safe design and construction of structural systems. However, numerical modeling of extreme events is a demanding task since it requires methods that can handle extreme inelastic deformations, damage, fracture, fragmentation, and free- surface turbulent flows. Participating students will have the opportunity to work with George Moutsanidis and his students on developing a numerical framework for the simulation of the aforementioned events, and on applying them to several challenging real-world scenarios.

Klaus Mueller

Computational Science

1. Do data analysis like a fruit fly!

Flies use an algorithmic neuronal strategy to sense and categorize odors. Dasgupta et al. applied insights from the fly system to come up with a solution to a computer science problem. On the basis of the algorithm that flies use to tag an odor and categorize similar ones, the authors generated a new solution to the nearest-neighbor search problem that underlies tasks such as searching for similar images on the web. Your task is to implement this concept and gauge its performance. The paper is here.

2. Teach a robot to make entertaining captions for data visualizations

Play around with new intelligent text generation algorithms like GPT-3 to generate interesting captions  for data visualization. Current techniques mostly produce narration that sound boring and nerdy. These new AI tools can read in lots of text, then connect this with the information about the data, and generate much more interesting stories.

Nav Nidhi Rajput

Materials Science & Chemical Engineering

Work in our lab is focused on using materials informatics to design optimal materials for next-generation energy storage, energy-conversion and carbon dioxide removal technologies. We use a multi-modal approach that combines high-throughput multi-scale simulations with surrogate machine learning models. Our group has developed two state-of-the-art open-source automated computational infrastructures (1) MISPR (Materials Informatics for Structure- Property Relationships; https://www.molmd.org/mispr/html/index.html) and (2) MDPropTools (https://github.com/molmd/mdproptools). MISPR seamlessly integrates density functional theory calculations with classical molecular dynamics simulations, allowing 100-1000s of calculations to be performed in parallel. MDPropTools is a Python-based post-processing routine to perform statistical analysis of properties of molecular dynamics simulations. An undergraduate student in Rajput Lab will be using MISPR and MDPropTools to develop large-scale databases of properties of electrolyte solutions computed using high-throughput multi-scale simulations for their applications in battery technologies. The generated databases will then be used to develop physics-aware machine learning models. In this process, students will apply the fundamental principles of computational chemistry and materials science while also learning advanced computational techniques such as coding in Python, managing databases, and developing machine learning models.

Christopher Pitt Wolfe

School of Marine & Atmospheric Science

Explore simpler Earths. Realistic climate models have become essential tools for predicting climate change, but they are often too complex to use to answer basic questions about the climate system. Starting with a production-class climate model and building a simpler Earth by removing most or all of the land allows us to ask basic questions about the climate. Like, how does the ocean control tropical climate? Or, how does El Niño change if you make the Pacific much wider? How do separate ocean basins influence each other through the atmosphere? This project would involve analyzing and possibly running one of the aquaplanet Earth configurations being developed at the School of Marine and Atmospheric Science. This project will teach students the foundations of climate dynamics and climate modeling.