### Lawrence Berkeley National Laboratory

An acceleration framework for parameter estimation using implicit sampling and adaptive

Robert Baraldi, University of Washington

**Practicum Year:**2018

**Practicum Supervisor:**Matthew Zahr, Luis W. Alvarez Postdoctoral Fellow, Computational Research Division, Mathematics Grou, Lawrence Berkeley National Laboratory

Parameter estimation for partial differential equations is a computationally demanding endeavor, particularly when adopting a Bayesian setting since it amounts to sampling from a posterior probability distribution involving the solution of the governing equations. For efficiency, we use implicit sampling to focus samples in high probability regions of the posterior distribution. However, locating regions of high probability amounts to solving a PDE-constrained optimization, which can be an expensive endeavor due to repeated queries to the primal and adjoint PDE. To remedy this cost, we replace expensive PDE solves efficient, projection-based reduced-order models, embedded in a globally convergent trust region framework. One we have located the maximum a posteriori point, the random maps sampling procedure is re-cast as a one-dimensional PDE-constrained optimization problems, which is also solved using the reduced-order model trust region method. The proposed method based on implicit sampling and reduced-order models is shown to substantially reduce the cost of parameter estimation for a subsurface flow model problem.

Ultrafast Electron Detector Enabled by High Performance Computing

Gabriela Correa, Cornell University

**Practicum Year:**2018

**Practicum Supervisor:**Peter Alexander Ercius, Staff Scientist, National Center for Electron Microscopy, Lawrence Berkeley National Laboratory

Electron microscopy is exuding exponentially more data as detector technologies advance. In particular, direct electron detectorsÃƒÂ¢Ã¢â€šÂ¬Ã¢â‚¬Âwhich provide a plethora of information in materials systemsÃƒÂ¢Ã¢â€šÂ¬Ã¢â‚¬Âare approaching 200 TB/hour through this project. This is made possible by use of a massive embedded system, connecting a direct electron detector to a supercomputer.
Direct electron detector data collected on the electron microscope is pre-processed on FPGA boards then sent across a 400 Gbps fiber optic network to NERSCÃƒÂ¢Ã¢â€šÂ¬Ã¢â€žÂ¢s software defined network, landing on Cori compute nodes. To prioritize bandwidth and data integrity, data on compute nodes are immediately sent to CoriÃƒÂ¢Ã¢â€šÂ¬Ã¢â€žÂ¢s Burst Buffer.
After processing, a portion of the data is immediately returned to the user to guide their experiment in real-time. This simple step would take several months, as opposed to seconds, without the use of HPC. The raw data remaining on Cori is subsequently used for advanced computational imaging. Without HPC the detector may approach a frame rate of 400 fps; with HPC the detector is projected to attain 100,000 fps.

Application of AMReX to an Eulerian method to simulate fluid-structure interactions

Yuexia Lin, Harvard University

**Practicum Year:**2018

**Practicum Supervisor:**Ann Almgren, Group Leader of CCSE, CCSE, Lawrence Berkeley National Laboratory

Adaptive mesh refinement (AMR) is a technique that allows solution of partial differential equations on a hierarchical structured grid, and adapts the level of refinement according to the accuracy of the solution. By providing higher spatial and temporal resolutions only in regions of interest as needed, AMR concentrates computational resources where they are most needed. AMR has been widely applied to many multi-scale and multi-physics problems, including magnetohydrodynamics and combustion.
The summer project has three parts. To learned about AMR strategies, I first implemented a 1D PDE solvers for hyperbolic and elliptic equations with grid refinement to learn operations such as refluxing at the fine-coarse boundary and sub-cycling in time.
The other two parts of the project are carried out concurrently, focusing on different aspects of AMReX applications. In one part, I started by learning the basics of creating an application in AMReX, then implemented simplified version of a fully Eulerian method for solving fluid-structure interaction in 3D, namely the reference map technique. To exploit the grid hierarchy, finer grids are used to discretize the solid body, and therefore are created and destroyed as the solid moves in the fluid. Coarse grids suffice in solving the fluid in simple test cases I developed, however, fine grids can be created to handle more complex flows.
The other part of the project is to investigate the effect of box sizes on GPU computation efficiency by adapting two existing CPU applications in AMReX to run on Tesla P100 NVIDIA GPUs on Summitdev at Oak Ridge National Laboratory. One application solves the heat equation using data held on meshes, and the other solves the Maxwell equations using both mesh and particle data. In this study, I also compare three common GPU programming frameworks for offloading kernels: CUDA, OpenACC, and OpenMP, both in terms of performance, and in terms of ease of programming on the current Summitdev development environment.

Genome Constellation

Kayla McCue, Massachusetts Institute of Technology

**Practicum Year:**2018

**Practicum Supervisor:**Zhong Wang, Group Leader, DOE Joint Genome Institute, Lawrence Berkeley National Laboratory

Genome Constellation is an ongoing project at JGI that explores the evolutionary relationships between microbial organisms using genome-based classification visualization

Climate Science, Deep Learning, and Pattern Discovery: The Madden-Julian Oscillation as a Test Case

Benjamin Toms, Colorado State University

**Practicum Year:**2018

**Practicum Supervisor:**Prabhat, Mr., National Energy Research Scientific Computing, Lawrence Berkeley National Laboratory

I applied deep learning to pattern recognition and discovery within atmospheric science, using a specific type of atmospheric wave called the Madden-Julian Oscillation (MJO) as a test case. The first test was to ensure CNNs could recognize the MJO within spatial fields, while the second test expanded upon this simple premise to see if the CNNs could learn the spatial structure of the MJO for themselves. Most atmospheric scientists do not see how deep learning can benefit their science, so I designed this project to try to identify the avenues that deep learning can be used to increase the efficiency of scientific discovery.

Computational discovery of novel two-dimensional CO2 reduction photocatalysts

Steven Torrisi, Harvard University

**Practicum Year:**2018

**Practicum Supervisor:**Kristin Persson, Professor & Staff Scientist, Chemical Sciences Division, Lawrence Berkeley National Laboratory

Using a database of bulk CO2 reduction photocatalysts and two-dimensional structures, I gauged the feasibility of bulk photocatalysts to be adopted into a two-dimensional form, and the suitability of those two-dimensional structures to be used for photocatalysis of CO2 into useful feedstocks.

Charged Pion Scattering with Massive QED

Casey Berger, University of North Carolina, Chapel Hill

**Practicum Year:**2017

**Practicum Supervisor:**Andre Walker-Loud, Staff Scientist, Nuclear Science, Lawrence Berkeley National Laboratory

This project involved adding QED into the QCD calculations of charged pion scattering. The electromagnetic force on the lattice was regulated using a massive photon, which will be extrapolated to the zero-mass (physical) limit.

Dynamical Models of Biological Subtraction in Genetic Circuits

Nicholas Boffi, Harvard University

**Practicum Year:**2017

**Practicum Supervisor:**Adam Arkin, Professor, Dr., Environmental Genomics and Systems Biology, Lawrence Berkeley National Laboratory

A self-optimizing genetic controller is a biological implementation (in the form of a genetic circuit) of a dynamical system capable of sensing and processing signals in the environment to optimize a relevant objective function. One possible application is in antibiotic drug delivery: one can imagine utilizing a strain of bacteria with a modified genome containing such a genetic circuit as a drug delivery system, where the bacteria attempt to optimize an objective function that measures their energy expenditure and their effectiveness at eliminating the pathogen. The goal would be to simultaneously minimize energy expenditure (so the drug delivery system does not die out, and does not release too much antibiotic) and maximize the death of the pathogen (to cure the host of the infection).
The Arkin lab has designed a biologically realizable dynamical systems model that, in theory, implements such a genetic controller, and the lab is currently in the process of assembling the model in a strain of bacteria. The full genetic circuit can be broken down into submodules which can be built individually, tested, and then combined in much the same way one would imagine building a typical engineering system. One fundamental issue in building the controller is a biological implementation of subtraction: it is often the case in control theory that terms such as "u - v" will appear on the right hand side of some equations, where u and v represent signals present in the system. For example, u might represent the current concentration of a given chemical in the system, and v the desired concentration. Then u - v represents an error term which the system can use to bring itself to the desired state. However, in a biological system, all signals are chemical concentrations and hence are nonnegative. The question then becomes: how does one implement subtraction in a biological system, when the negation operation does not exist?

Exploration of the Electronic Landscape

Sukin Sim, Harvard University

**Practicum Year:**2017

**Practicum Supervisor:**Jarrod McClean, Luis W. Alvarez fellow, Computing Sciences, Lawrence Berkeley National Laboratory

In variational quantum chemistry, the choice of ansatz ("trial wavefunction") is crucial for an accurate and efficient description of a chemical system. Despite the rapid progress in development of methods in this area, there is no systematic way to analyze and evaluate an ansatz. In this work, we develop tools to construct and characterize landscapes of various methods. To investigate difficult landscapes in electronic surfaces, we considered the Hartree-Fock and unitary coupled cluster ansatzes. A deeper understanding of characteristics that make certain methods better suited for studying particular molecules may provide new insights and strategies to approach difficult problems in chemistry.

Exploring methods to compress deep neural networks using union of intersections

Alex Kell, Massachusetts Institute of Technology

**Practicum Year:**2017

**Practicum Supervisor:**Kristofer Bouchard, Research Scientist, Life Sciences and Computational Research Divisions, Lawrence Berkeley National Laboratory

Deep learning has surged in popularity in recent years -- leading to engineering feats in computer vision, automated speech recognition, and artificial intelligence. While these models yield quite impressive performance on a variety of tasks, they can be quite difficult to gain intuitions about. In this work, we explored means of compressing these networks into smaller models, with the hope that this compression may reduce redundancy in expansive models and potentially lead to improvements in intuition and understanding about these networks.

Dynamic Load Balancing for Storm Surge Applications

Maximilian Bremer, University of Texas

**Practicum Year:**2016

**Practicum Supervisor:**Cy Chan, Computer Research Scientist, Computational Research Division, Lawrence Berkeley National Laboratory

Asynchronous computing provides a novel way to deal with the transition to next generation HPC systems. One nice feature of asynchronous runtimes is the relative ease of performing dynamic load balancing. The goal of this practicum was to explore load balancing schemes to improve the efficiency of a hurricane load balancing application. To allow for rapid development, we wrote a simulator which reproduces the load profile of the application and then used this simulator to implement various load balancing strategies.

Spark Based Metagenome Assembler

Jordan Hoffmann, Harvard University

**Practicum Year:**2016

**Practicum Supervisor:**Zhong Wang, Staff Scientist, Group Lead, JGI/Genomics, Lawrence Berkeley National Laboratory

I went to the Joint Genome Institute (JGI) and worked with Dr. Zhong Wang. I built on an existing Spark framework for meta genome assembly. I worked on scaling the code to handle large datasets as well as using tools form machine learning to explore a very high dimensional parameter space.

Computational Optimization of Biosynthetic Gene Clusters

Adam Riesselman, Harvard University

**Practicum Year:**2016

**Practicum Supervisor:**Sam Deutsch, Synthetic Biology Group Lead, Joint Genome Institute, Lawrence Berkeley National Laboratory

Organisms produce a wide array of small molecules that are useful drugs, biofuels, antibiotics, and other chemicals. The genomic instructions to make these molecules are encoded in the genomes of the organisms that produce them. The DNA instructions to these enzymes can be essentially printed out and tested in a host organism that can be grown easily and cheaply in a high-throughput manner.
The landscape of small molecule production can subsequently be sampled with the above technique, but production is usually suboptimal. I used computational and statistical approaches to analyze this data and propose the next experimental step to optimize small molecule production.

Numerical Methods for the Drag-Coupled Particles in Viscous Flow

Jay Stotsky, University of Colorado

**Practicum Year:**2016

**Practicum Supervisor:**Dan Martin and Phil Colella, Mathematician Staff Scientist/Engineer, Computational Research Division, Lawrence Berkeley National Laboratory

With my advisors at the Lawrence Berkeley National Laboratory, I worked on numerical algorithms for simulating fluid flows containing large numbers of embedded particles. Specialized algorithms are needed for the simulation of flows containing particles
because each particle induces a singular force on the fluid which cannot be accurately represented by standard fluid dynamics algorithms.
The two algorithms studied were the Particle-Particle/Particle-Mesh (P3M) algorithm and the Method of Local Corrections (MLC). The P3M algorithm is based on a splitting of the force into short and long range terms. For the long range forces,
grid-based methods of solving for the fluid velocity can be applied. For the short range forces, a special analytical solution for the velocity is used. This algorithm provides a balance between the accuracy of analytical solutions,
and the computational efficiency of grid-based approximations.
The MLC algorithms is based upon the use of a special class of finite-difference stencils known as Mehrstellen stencils. This class of stencils has the special property that when applied to harmonic functions, they are highly accurate, decaying rapidly to zero. Thus, by locally applying the Merhstellen stencil to a regularized version of the singular force, and then solving a global problem on a grid, globally accurate solutions may be obtained efficiently.
Both algorithms, are efficient compared to direct computation of the particle velocities, and have similar computational costs. We found however, that the MLC can obtain more rapid convergence then the P3M algorithm.

Network design for quantifying urban CO2 emissions: Assessing tradeoffs between precision and network density

Alexander Turner, Harvard University

**Practicum Year:**2016

**Practicum Supervisor:**Ronald C. Cohen, Faculty Scientist, Energy and Environment Technologies, Lawrence Berkeley National Laboratory

We submitted a manuscript for publication a couple weeks after I left. Here's the abstract:
The majority of anthropogenic CO2 emissions are attributable to urban areas. While the emissions from urban electricity generation often occur in locations remote from consumption, many of the other emissions occur within the city limits. Evaluating the effectiveness of strategies for controlling these emissions depends on our ability to observe urban CO2 emissions and attribute them to specific activities. Cost effective strategies for doing so have yet to be described. Here we characterize the ability of a prototype measurement network, modeled after the BEACO2N network, in combination with an inverse model based on WRF-STILT to improve our understanding of urban emissions. The measurement network includes 34 measurement sites at roughly 2 km spacing covering an area of roughly 400 km^2. The model uses an hourly 1x1 km^2 emission inventory and 1x1 km^2 meteorological calculations. We perform an ensemble of Bayesian atmospheric inversions to sample the combined effects of uncertainties of the measurements and the model. We vary the estimates of the combined uncertainty of the observations and model over a range of 20 ppm to 0.005 ppm and vary the number of sites from 1 to 34. We use these inversions to develop statistical models that estimate the efficacy of the combined model-observing system at reducing uncertainty in CO2 emissions. We examine uncertainty in estimated CO2 fluxes at the urban scale, as well as for sources embedded within the city such as a line source (e.g., a highway) or a point source (e.g., emissions from the stacks of small industrial facilities). We find that a dense network with moderate precision is the preferred setup for estimating area, line, and point sources from a combined uncertainty and cost perspective. The dense network considered here could estimate weekly CO2 emissions from an urban region with less than 5% error, given our characterization of the combined observation and model uncertainty.

Analysis of high throughput screens using CRISPRi and adapting CRISPRi for probing phage infections

Joy Yang, Massachusetts Institute of Technology

**Practicum Year:**2016

**Practicum Supervisor:**Adam Arkin, Division Director, Physical Biosciences Division, Lawrence Berkeley National Laboratory

There were two main aspects of my practicum:
1. In the Arkin lab, Harneet has completed a high-throughput CRISPR screen of E. coli MG1655, with spacers designed against genes, promoters and transcription factor binding sites. One aspect of my practicum was to use this dataset to explore the features that impact CRISPRi efficacy and to study the organization of regulatory systems.
2. The second was to knock-down the expression of phage genes during infection using CRISPRi. The Polz lab, where I’m doing my thesis work, has many uncharacterized phage, and this could potentially be a very powerful technique for probing phage genes.

High-dimensional variable selection and shrinkage in pseudolikelihood estimators of the inverse covariance matrix

Alnur Ali, Carnegie Mellon University

**Practicum Year:**2015

**Practicum Supervisor:**Sang-Yun Oh, Postdoctoral Researcher, Computational Research Division, Lawrence Berkeley National Laboratory

Here's the abstract from the paper we're planning on submitting on this work:
We introduce "covshrink", a new pseudolikelihood estimator for the inverse covariance matrix that is the solution of a (strongly) convex optimization problem featuring sparsity-inducing and shrinkage penalties. Theoretically, we show that covshrink (a) does not "saturate", i.e., it may have more than np nonzero entries when n > p, where n is the number of observations and p is the number of predictors, generalizing a similar result for the lasso to (undirected) graphical models (b) converges to the unique global solution at a geometric (i.e., linear) rate (for fixed n,p) (c) converges in probability (n,p --> infty) (d) produces positive definite iterates (e) admits simple "screening rules", i.e., it is possible to omit predictors from the optimization problem a priori, yielding significant computational savings without sacrificing predictive accuracy; these properties distinguish covshrink from other pseudolikelihood estimators that we study. Empirically, we show that these theoretical benefits translate into quantitative improvements in variable selection, predictive, and estimation accuracy on synthetic, finance, and neuroscience examples.
Here's a simplified description of the problem setting in case it helps:
Suppose you take some measurements of various parts of the brain; can you construct a graph from these measurements, where vertices correspond to parts of the brain and edges are placed between vertices where there is deemed to be some kind of "meaningful" interaction? (This is hard when there are very few measurements, and the measurements are (say) non-Gaussian.) This graph can aid scientific discovery, and this problem is known as sparse inverse covariance estimation.

"Automating Image Processing for Planet and TNO Searches in PTF" (unofficial)

Ryan McKinnon, Massachusetts Institute of Technology

**Practicum Year:**2015

**Practicum Supervisor:**Peter Nugent, Division Deputy for Scientific Engagement, Computational Research Division, Lawrence Berkeley National Laboratory

This project evolved into developing astrophysical image processing algorithms to apply to archival Palomar Transient Factory data in order to look for fast-moving objects. This requires efficiently combining large images taken over a period of time in varying conditions and distinguishing potential signals on these images from background noise and other artefacts. Computing for this project was done at the National Energy Research Scientific Computing Center.

A Hartree-Fock Application using UPC++ and the New DArray Library

David Ozog, University of Oregon

**Practicum Year:**2015

**Practicum Supervisor:**Kathy Yelick, Associate Lab Director, Computing Sciences, Lawrence Berkeley National Laboratory

The Hartree-Fock (HF) method is the fundamental first step for incorporating quantum mechanics into many-electron simulations of atoms and molecules, and is an important component of many computational chemistry toolkits. To enable innovations in algorithms and exploit next generation exascale systems, it is crucial to support quantum chemistry codes using expressive and convenient programming models and runtime systems that are also efficient and scalable. My project involved implementing an HF application using UPC++, a partitioned global address space model that includes flexible communication, asynchronous remote computation and a powerful multidimensional array library. UPC++ offers runtime features that are useful for HF such as active messages, a rich calculus for array operations, hardware supported fetch-and-add, and functions for ensuring asynchronous runtime progress. I also designed a new distributed array abstraction, DArray, that is convenient for the kinds of random access array updates and linear algebra operations on block distributed arrays with irregular data ownership. I analyzed the performance of atomic fetch-and-add operations (relevant for load balancing) and runtime attentiveness, then compared various techniques and optimizations for each. The final optimized implementation of HF using UPC++ and the DArrays library shows up to 20% improvement over the original code at scales up to 24,000 cores.

Robust Estimation of Auditory Cortex Tuning and Connectivity

Danielle Rager, Carnegie Mellon University

**Practicum Year:**2015

**Practicum Supervisor:**Kristofer Bouchard, Dr., Life Sciences (primary), Computational Research, Lawrence Berkeley National Laboratory

The activity of any given neuron in a network that encodes sensory information is due to both the neuron's response to the sensory stimulus and the input that the neuron receives from other neurons in the network. In auditory cortex, neurons exhibit preferential spiking responses to tones of a certain frequency. We call this a neuron's "frequency tuning". When an animal is presented a tone, a neuron will exhibit a spiking response that is dependent both on its frequency and on the activity of the neurons that connect to it, each of which will have their own frequency tuning. At Lawrence Berkeley National Laboratory, I helped develop regularized, hierarchical regression models that captured the dependency structure between frequency tuning and network connectivity of auditory neurons. Finding stable solutions to these regularized, hierarchical models is computationally expensive and required parallel implementations of the optimization algorithm that were run on NERSC high performance computing resources.

Parallel algorithms for maximum-cardinality high-weight graph matching

Adam Sealfon, Massachusetts Institute of Technology

**Practicum Year:**2015

**Practicum Supervisor:**Aydın Buluç, Research Scientist, Computational Research Division, Lawrence Berkeley National Laboratory

A matching in a graph is a subset of the graph edges such that each vertex is incident to at most one edge in the matching. A matching is of maximum cardinality if no other matching has more edges. A maximum cardinality matching is a perfect matching if every vertex in the graph is matched. For weighted graph, the concept of a maximum weight matching can be defined.
Maximum cardinality or perfect matching is needed in several applications. One application in scientific computing is to permute a matrix to its block triangular form via the Dulmage-Mendelsohn decomposition of bipartite graphs. Once the block triangular form is obtained, in circuit simulations, sparse linear systems of equations can be solved faster, and data structures for sparse orthogonal factors for least-squares problems can be correctly predicted.
Several other important applications in scientific computing require high-weight matchings (for weighted graphs) in addition to the maximum cardinality constraint. One popular scheme, static pivoting, is used to find a permutation that maximizes the diagonal entries of a sparse matrix as a preprocessing step, in order to avoid expensive dynamic pivoting in Gaussian Elimination. Static pivoting is typically performed by finding a maximum weighted matching in the bipartite graph of the sparse matrix. This matching has to be “perfect” or “maximum cardinality” in order to avoid expensive rewrite of legacy code that performs sparse Gaussian elimination. Since static pivoting is essentially a heuristic, the maximum weight requirement can be relaxed to high weight in practice.
Scientists routinely solve very large sparse linear systems that do not fit into the memory of a single node computer. Consequently, distributed-memory parallel solvers, such as SuperLU or MUMPS, became widely used. However, scalable parallel algorithms for graph matching are challenging, especially in distributed memory.
The focus of the project was to explore parallel algorithms for maximum-cardinality high-weight matching, both in bipartite and general graphs. The latter kinds of graphs arise in the matching problems required for solving sparse symmetric indefinite systems. A symmetric maximum matching on a bipartite graph with 2N vertices (N on each part) is equivalent to a maximum matching on a general graph with N vertices. This equivalence holds for weighted matching as well as well for cardinality matching.

Reactive force field modelling of silicate sorption on the hematite r-plane

Brenhin Keller, Princeton University

**Practicum Year:**2014

**Practicum Supervisor:**Glenn Waychunas, Senior Scientist, Earth Sciences Division, Geochemistry Department, Lawrence Berkeley National Laboratory

Iron oxides and oxyhydroxides are abundant both in natural and man-made environments - formed by oxidative weathering of reduced Fe-bearing materials. Due to the high surface area of particles produced by such weathering processes, iron oxide surface chemistry is relevant for a number of potential reaction, catalysis, and transport processes, such as diffusion of heavy metals in acid mine drainage or nuclear wast storage environments. Passivation of Fe-oxide sur- faces by silicate can dramatically alter surface chemistry - reducing sorption of other ions, includ- ing heavy metals. Molecular simulation would provide substantial mechanistic insight into silicate passivation of iron oxide surfaces, but is challenging due to the rarity of sorption events on times- cales accessible by molecular dynamics. In order to approach this problem, we considered three main classes of molecular dynamics (MD) simulation: ab-initio MD, classical MD, and reactive force field MD to assess their applicability to silicate sorption to the hematite r-plane - a naturally-occuring iron oxide surface with particularly well-known surface geometry.

FASTERp: A Feature Array Search Tool for Estimating Resemblance of protein sequences

Derek Macklin, Stanford University

**Practicum Year:**2014

**Practicum Supervisor:**Zhong Wang, Staff Scientist, Group Lead, Genomics (at the Joint Genome Institute), Lawrence Berkeley National Laboratory

Metagenome sequencing efforts have provided a large pool of billions of genes for identifying enzymes with desirable biochemical traits. However, homology search with billions of genes in a rapidly growing database has become increasingly computationally impractical. To address this problem, we have developed a novel alignment-free algorithm for homology search. Specifically, we represent individual proteins as feature vectors that denote the presence or absence of short kmers in the protein sequence. Similarity between feature vectors is then computed using the Tanimoto score, a distance metric that can be rapidly computed on bit string representations of feature vectors. Preliminary results indicate good correlation with optimal alignment algorithms (Spearman r of 0.87, ~1,000,000 proteins from Pfam), as well as with heuristic algorithms such as BLAST (Spearman r of 0.86, ~1,000,000 proteins). Furthermore, a prototype of FASTERp implemented in Python runs approximately four times faster than BLAST on a small scale dataset (~1000 proteins). We are optimizing and scaling to improve FASTERp to enable rapid homology searches against billion-protein databases, thereby enabling more comprehensive gene annotation efforts.

Bidi: scalable bit-distance calculations using MPI + OpenCL

Sarah Middleton, University of Pennsylvania

**Practicum Year:**2014

**Practicum Supervisor:**Zhong Wang, Dr., Joint Genome Institute, Lawrence Berkeley National Laboratory

I developed a package for doing bit-distance calculations at scale. It is a flexible framework that can accommodate many different kinds of bitwise comparison measures (we used bitwise AND/OR operations, but there are others) and facilitates the distribution of these calculations across multiple GPUs or CPUs. I used a combination of OpenCL and MPI, which makes the package very portable--the same code can be run on both GPU and CPU clusters without any extra effort from the user. The specific application of this work that we had in mind during development was large scale protein-vs-protein similarity measuring, where proteins are represented as bit vectors using a transformation developed by another CSGF student, Derek Macklin. For example, this package will enable the comparison of putative proteins from newly sequenced organisms against known proteins in the IMG database, which currently contains over 2 billion proteins.

A convolutional free-space propagator for Maxwell's equations in free-space with application to PIC methods

Victor Minden, Stanford University

**Practicum Year:**2014

**Practicum Supervisor:**Phillip Colella, Group Lead / Senior Scientist, Computational Research / Numerical Algorithms, Lawrence Berkeley National Laboratory

Two broad classes of schemes for solving Maxwell's equations in free-space are finite-difference time-domain schemes, which suffer from stability issues unless the time-step is small and pseudo-spectral schemes, which are difficult to parallelize effectively. In this project, we construct a convolutional propagator in the space domain that is essentially unconditionally stable and simple to parallelize.

Metagenomic Assembly by K-mer Histogram Peelback

Daniel Strouse, Princeton University

**Practicum Year:**2014

**Practicum Supervisor:**Daniel Rokhsar, Chief Informatics Officer/Eukaryote Spr Prgrm Head, Joint Genome Institute, Lawrence Berkeley National Laboratory

One of the central problems of genomics is to piece together the “genetic code” for different species. Unfortunately, these codes can’t simply be read off an organism like a book. Instead, researchers only get access to small snippets of DNA at once. Taking many of these snippets at once from the same genome, researchers then try to piece together the full genome, like you might put together a puzzle. This problem is known as “genomic assembly.”
Now often, researchers aren’t able to isolate snippets of DNA from a single organism at once, but instead get snippets of DNA from across an entire community of organisms. Examples might be all the microbes in a sample of pond water or all the microorganisms in the human gut that help with digestion. Here the assembly problem becomes more complicated; instead of putting together a single puzzle, researchers have to piece together several diffierent puzzles that have been mixed together. This problem is known as “metagenomic assembly.”
This summer, I worked on implementing a new approach to metagenomic assembly and testing it on data taken from groundwater samples.

Estimating urban carbon dioxide fluxes at high resolution from in situ observations

Alexander Turner, Harvard University

**Practicum Year:**2014

**Practicum Supervisor:**Ronald C. Cohen, Faculty Scientist, Energy and Environment Technologies, Lawrence Berkeley National Laboratory

Carbon dioxide (CO2) is an atmospheric trace gas and the largest anthropogenic radiative forcer. CO2 levels have increased from 280 ppm in pre-industrial times to greater than 400 ppm in the present, largely due to changes in fossil fuel emissions, and can be measured via ground stations, aircraft, and satellites. The paradigm in ground-based trace gas measurements has been to employ a sparse network of high-precision instruments that can be used to measure atmospheric concentrations (e.g., the Total Carbon Column Observing Network has a network of 19 Fourier transform spectrometers worldwide). These concentrations are then used to estimate emission fluxes, validate numerical models, and quantify changes in physical processes. However, the BeACON project (http://beacon.berkeley.edu/Overview.aspx) aims to provide a better understanding of the emissions and physical processes governing CO2 by deploying a high density of moderate-precision instruments. We propose to (1) determine the optimal CO2 measurement network for California's Bay Area, (2) employ the BeACON observations to estimate hourly atmospheric fluxes of CO2 at an unprecedented 1 km spatial resolution, and (3) use these hourly fluxes to improve our understanding of the underlying physical processes such as traffic congestion.
The first objective was addressed by performing a series of Observing System Simulation Experiments (OSSEs), an OSSE is where a hypothetical observing system (in this case a network of ground-based instruments) makes measurements from a simulated atmosphere. One can then vary the observing system to determine what the observing system is sensitive to and quantify the benefits of assimilating measurements from different observing systems into a numerical model. This allowed us to objectively define the optimal observing system for California's Bay Area and will inform the design of future CO2 measurement campaigns. The latter objectives were addressed by solving the inverse problem relating the atmospheric observations to the underlying emissions from California's Bay Area.

Hydrodynamic Simulations of high multiplicity proton proton collisions at a center of mass energy of 13 TeV

Dragos Velicanu, Massachusetts Institute of Technology

**Practicum Year:**2014

**Practicum Supervisor:**Matthew Luzum, Dr, Nuclear Science, Lawrence Berkeley National Laboratory

The project is to use the hydrodynamics code, MUSIC, to simulate the hydrodynamic expansion of single proton proton collisions which produce a very high number of particles in the final state. This project aims to reproduce the already measured 7 TeV proton proton measured correlations at the LHC and to predict some measurable quantities from the 13 TeV proton proton run that will occur in 2015. The theoretical predictions from this work can be directly tested in the upcoming higher energy collisions and the results of this comparison will tell us whether hydrodynamics a likely theory for explaining the observed long range correlations in high multiplicity collisions.

Adaptive mesh refinement (AMR) and embedded boundaries (EB) in Chombo

Curtis Lee, Duke University

**Practicum Year:**2013

**Practicum Supervisor:**Hans Johansen, Researcher, Applied Numerical Algorithms Group, Computational Research Division, LBNL, Lawrence Berkeley National Laboratory

This practicum experience involved working with Chombo, an advanced, highly-parallelized software application for the adaptive solution of partial differential equations in a finite-volume, AMR framework. The first half of the practicum was spent learning the C++ programming language, the intricacies of Chombo, various debugging applications, and performance profiling tools. After this was accomplished, I used various components of Chombo to develop and test a 2D time-dependent heat equation solver with embedded boundaries. I next began working towards the development of a heat solver with a moving embedded boundary; this involves the marriage of the embedded boundary technology used in Chombo and other moving interface techniques. I spent the remainder of my practicum working towards this goal, and I will continue to collaborate with my mentor on this project.

High-speed visualization of/interaction with ALS data for previewing algorithm parameter tweaks

Justin Lee, Massachusetts Institute of Technology

**Practicum Year:**2013

**Practicum Supervisor:**Dula Parkinson, Beamline Scientist, Advanced Light Source, Lawrence Berkeley National Laboratory

I worked on developing a web tool for ALS users to visually interact with their large X-ray/CT datasets. The tool was interfaced with NERSC so that users could rapidly preview desired adjustments to their image processing algorithms using my interface, then click a button to re-run their data processing algorithms on NERSC with their selected parameters.

Monte Carlo Radiation Transport in Compact Binary Merger Disks

Sherwood Richers, California Institute of Technology

**Practicum Year:**2013

**Practicum Supervisor:**Daniel Kasen, Professor, Physics, Astronomy, Lawrence Berkeley National Laboratory

We combine the neutrino opacity library NuLib and the Monte Carlo radiation transport code SEDONA to simulate neutrino transport through disks surrounding binary compact object (neutron star-neutron star or neutron star-black hole) merger disks. Monte Carlo radiation transport is far more accurate than the approximate transport schemes currently employed in dynamical simulations, and our transport simulations give a more precise description of the energetics and composition of the neutrino-driven wind. This could help understanding of the observable characteristics of these systems and also of the poorly-understood origin of certain heavy elements like gold and platinum.

Inferring the ecological niches of thousands of microbial species using synthetic ecosystems

Chris Smillie, Massachusetts Institute of Technology

**Practicum Year:**2013

**Practicum Supervisor:**Adam Arkin, Dr, Physical Biosciences Division, Lawrence Berkeley National Laboratory

At MIT, I work on a large DOE project called ENIGMA. One part of ENIGMA focuses on characterizing the groundwater geochemistry and microbial community in a field site at Oak Ridge National Lab. Using field data, I first calculated statistical models relating groundwater geochemistry to the distributions of bacteria that inhabit the Oak Ridge field site. For example, by knowing the temperature, pH, and nitrate levels in the groundwater, can we predict the relative abundances of the specific bacterial species that inhabit the groundwater? This was the first problem I worked on solving during my practicum in the Arkin lab. Having calculated these statistical models, the second part of my practicum was to experimentally test these models by creating synthetic ecosystems in the lab, inoculating with groundwater bacteria, and determining which bacterial species grow. For example, by experimentally manipulating pH, temperature, and nitrate, can we predict the bacteria that will grow in our synthetic ecosystem, therefore testing the predictions of the statistical models we calculated using the field data? This is the second question I worked on during my practicum.

The Shape of Climate: Topological Data Analysis on Climate Data

Melissa Yeung, California Institute of Technology

**Practicum Year:**2013

**Practicum Supervisor:**Dmitriy Morozov, Postdoctoral Fellow, Visualization Group, Lawrence Berkeley National Laboratory

Under the guidance of Dmitriy Morozov and Prabhat at LBL, I used techniques from topological data analysis to study the dynamics of high-dimensional climate data. In particular, we were interested in detecting climate oscillations over a range of spacial and temporal resolutions.

Solving 1,000 Puzzles with 8 Billion Small Pieces: Assembling Draft Genomes from the Cow Rumen Metagenome

Aleah Caulin, University of Pennsylvania

**Practicum Year:**2012

**Practicum Supervisor:**Zhong Wang, Staff Scientist, Genomics (at the Joint Genome Institute), Lawrence Berkeley National Laboratory

Next generation sequencing technologies, applied to metagenomics and single cell genomics, have provided two culture-independent ways of investigating the genetic and metabolic potentials of complex microbial communities. At the DOE’s Joint Genome Institute, over one tera-base of sequence data has been generated from a microbial community adherent to switchgrass in the cow rumen. Assembling individual genomes from this data, with 8 billion short reads from a mixture of over 1,000 species has proven difficult. One of the biggest challenges is that there are no suitable reference genomes that exist for assembly quality assessment. Here we propose to integrate single cell genomics and metagenomics to de novo assemble a set of high quality reference genomes. We developed a three-step strategy: assemble each single amplified genome (SAG) into contigs, bin cells representing the same species, and use these bins to recruit assembled metagenome contigs. This helps to create a more complete draft genome for each species represented by the isolated cells. A hierarchical clustering algorithm based on sequence features, such as tetranucleotide frequency, is then applied to bin the remaining metagenome contigs. Generating these draft genomes is the first step towards the assembly of many uncultured novel organisms from complex communities, which will serve as a foundation to the comprehensive study of community composition, structure and dynamics.

Accuracy checks of Chombo's mapped multiblock framework

Seth Davidovits, Princeton University

**Practicum Year:**2012

**Practicum Supervisor:**Phil Colella, Director Applied Numerical Algorithms Group, Applied Numerical Algorithms Group, Lawrence Berkeley National Laboratory

The Chombo package, developed by the Applied Numerical Algorithms Group at Lawrence Berkeley National Laboratory and their collaborators, provides advanced tools for the solution of partial differential equations with massively parallel computing. Among its features are adaptive mesh refinement, high order accurate finite volume methods, mapped grids, and embedded boundaries. My work focused on the mapped multiblock grid framework. The mapped multiblock grid framework is a technique designed for flexibly handling a variety of complicated geometries, such as a spherical shell geometry relevant for atmospheric simulations, or a x-point geometry associated with simulations of tokamak fusion devices. I conducted accuracy tests of the framework.

Parallelizing three-dimensional viscoelastic models

Phoebe DeVries, Harvard University

**Practicum Year:**2012

**Practicum Supervisor:**Roland Burgmann, Dr., Earth Sciences Division, Lawrence Berkeley National Laboratory

Last week, I wrote a description of my practicum project in the research part of my renewal application -- I hope it is all right if I use the same description here.
The practicum goal was to parallelize one or more of the existing three-dimensional viscoelastic relaxation codes and try to benchmark them against one another. At the Southern California Earthquake Conference, which took place for two days in September right at the beginning of my practicum, and there was a simple benchmarking effort going on at the conference. The extent to which the existing serial semi-analytic and finite-element codes disagreed with one another was striking. Because of the disagreement between these existing codes, we spent more time working on understanding the math in detail to make sure that the model we were going to parallelize was correct. The rate-limiting step is the inverse Fourier transform in cylindrical coordinates which requires thousands of Bessel function calculations. So far we have found that we get a factor of 40 speed up when we put these Bessel function calculations on the GPU. The exciting part of this practicum research is that until now, geodetic slip rate estimates have been based largely on two-dimensional models which assume faults are infinitely long. But, with a three-dimensional code that runs on a GPU, we can build models that are more geometrically complex and tectonically realistic than previously possible.

Shirley Compact Basis

Maxwell Hutchinson, University of Chicago

**Practicum Year:**2012

**Practicum Supervisor:**David Prendergast, Staff Scientist, Molecular Foundry, Lawrence Berkeley National Laboratory

Popular electronic codes, such as VASP, Quantum Espresso, and ABINIT, use plane-wave bases to represent the Kohn-Sham Hamiltonian. In general, these bases are quite large, often upwards of 1 million dimensions. The Shirley compact basis (SCB) is an alternative basis that is generally appropriate for electronic structure calculations and is quite small, usually on the order of 100s of dimensions. I worked on writing the operators used in a variety of physical applications in the SCB. This work is built on the Quantum Espresso framework, and will be released as the QE-Forge project "Shirley".

High resolution global constraints on atmospheric carbon dioxide with a novel statistical framework

Scot Miller, Harvard University

**Practicum Year:**2012

**Practicum Supervisor:**Inez Fung, , Earth Sciences Division, Lawrence Berkeley National Laboratory

My practicum research leverages new statistical tools to improve our understanding of carbon dioxide in the atmosphere as well as carbon dioxide sources. I utilized a global atmospheric model and a wide variety of meteorological and carbon dioxide observations in order to obtain high resolution constraints on CO2 concentrations in the atmosphere as well as the interaction among meteorological variables and CO2. This work provides an improved understanding of CO2 sources, of the uncertainties in these sources and of the effect of meteorological uncertainties on our knowledge of CO2.

Time Integration and Nonlinear Solvers for the Speedup of the Discontinuous Galerkin Finite Element Method

Matthew Zahr, Stanford University

**Practicum Year:**2012

**Practicum Supervisor:**Per-Olof Persson, Assistant Professor, Mathematics Department, Lawrence Berkeley National Laboratory

I spent my practicum assignment working in the mathematics group at LBL under the supervision of Per-Olof Persson on addressing the computational cost of high-order DG methods. The purpose of the project is find computational techniques to speed up DG-FEM, which is notorious for having large memory requirements and high computational cost. The project focuses is finding faster time integration schemes and nonlinear solvers, better preconditioners for linear solvers, and suitable prediction techniques for nonlinear solvers.

Model-based detection of interactions from time-series data

Edward Baskerville, University of Michigan

**Practicum Year:**2011

**Practicum Supervisor:**Adam Arkin, Faculty Scientist, Physical Biosciences Division, Lawrence Berkeley National Laboratory

The purpose of the project was to investigate methods for identifying interactions from time-series data in the context of Bayesian process models, for example autoregressive models, and to investigate applying these methods to microbial data sets, including data from a soil grassland wet-up experiment.

Simulation of Porous Media Flow in a Nuclear Waste Repository

Leslie Dewan, Massachusetts Institute of Technology

**Practicum Year:**2011

**Practicum Supervisor:**John Bell, CCSE Director, Center for Computational Sciences and Engineering, Lawrence Berkeley National Laboratory

During this project, I worked with the Center for Computational Sciences and Engineering at the Lawrence Berkeley National Laboratory to develop accurate simulations of reactive subsurface geochemical flow. Specifically, I investigated the application of parallelized, high-resolution adaptive mesh algorithms to multiphase flow through porous geological media.
There are several factors that make porous media flow a difficult problem to solve. The primary issue is that the reactions occur on a wide range of time and length scales. Porous media flow often contains regions of steep concentration gradients, which require very fine meshes to resolve accurately. Furthermore, there can be multiple phases, each with a complex chemical composition, interacting with the minerals in the rock matrix. It is therefore necessary to develop a detailed representation of the geochemistry, including a large number of reactive species, to accurately model the flow.
My particular simulations focused on modeling the flow of groundwater through the porous rock media surrounding a nuclear waste repository.

Fluctuating Hydrodynamics

Christopher Eldred, Colorado State University

**Practicum Year:**2011

**Practicum Supervisor:**John Bell, , Center for Computational Sciences and Engineering, Lawrence Berkeley National Laboratory

The project was entitled Fluctuating Hydrodynamics. Specifically, we developed a hybrid particle-continuum code for the simulation of fluid flows where the Knudsen number is at or near the continuum limit. Such flows include microfluidics and highly rarified flows. There were two components to the existing code- a Direct Simulation Monte Carlo (DSMC) particle solver and a Landau-Lifshitz Navier Stokes (LLNS) continuum solver. These could be coupled to handle transitional flow regimes. The LLNS solver was also capable of adaptive mesh refinement (AMR). Both the DSMC and LLNS codes had been well tested and published by the time I joined the project. I worked on extending the existing DSMC and LLNS codes to include particles with internal energy (the earlier code only handled particles with translational energy) as a prelude to eventually adding chemical reactions.

Low Mach Number Fluctuating Hydrodynamics

Thomas Fai, New York University

**Practicum Year:**2011

**Practicum Supervisor:**John Bell, Dr., Center for Computational Sciences and Engineering, Lawrence Berkeley National Laboratory

I worked on the development of new methods for simulating the mixing of two fluids due to thermal fluctuations.

Grouping molecules of similar orientation in single-molecule diffraction experiments

Charles Frogner, Massachusetts Institute of Technology

**Practicum Year:**2011

**Practicum Supervisor:**Chao Yang, Staff, Computational Research Division, Lawrence Berkeley National Laboratory

The goal of this project was to develop computational techniques and tools that can be used process diffraction imaging data, to elucidate 3D structures of nanomaterials and biological molecules. I focused on data from non-crystalline X-ray diffraction imaging (Miao, 1999), which can reveal molecular or cellular structures that are difficult to obtain through other existing techniques such as X-ray crystallography, X-ray tomography, and electron microscopy. Since non-crystalline imaging does not rely on building high-quality lenses or growing crystals, it is suitable for structure studies using a wide variety of light sources, including the Linac Coherent Light Source (LCLS) at Stanford and the Next Generation Light Source that is under preparation at LBNL.
In a single molecule diffractive imaging experiment, small droplets, containing on average one molecule per droplet, are placed in the path of an X-ray beam, and 2D diffraction images of these identical molecules are recorded on a detector. There are three major barriers to reconstructing the 3D structure of the molecule from these diffraction images: (i) the signal-to-noise ratio of each diffraction image is low (ii) the molecules are randomly oriented within the droplets, and we don't know their orientations; (iii) after reconstructing a 3D diffraction pattern from the 2D images, we need to retrieve the 3D phase image to obtain a 3D structure. I focused on the first barrier.
The need to boost the signal-to-noise ratio is motivated by current approaches to solving (ii), the orientation determination problem. One such technique was proposed by Singer et al. [1] in the context of cryo-EM data. In this approach, one collects the 2D diffraction images corresponding to many different orientations of the molecule, and identifies correspondences between each pair. Each 2D diffraction image can be imagined as lying on a randomly-oriented plane in the full 3D diffraction image space, so between each pair one will find a line of intersection. The problem then reduces to finding the orientations of the 2D image planes that best match these lines of intersection between pairs of them, and Singer et al. [1] formulates a relaxed version of this problem that centers on computing the eigenvectors associated with the largest eigenvalues of a symmetric matrix.
The approach of Singer et al., and an extension of this idea developed by my proposed supervisor at LBL, Dr. Chao Yang, both rely on the correct determination of correspondences between individual 2D diffraction images; the low signal-to-noise ratio of the X-ray imaging data currently prevents reliable detection of these correspondences. During the summer, I attacked this problem using computer vision and statistical learning techniques to group 2D diffraction images that are of similar orientations. By averaging the images within each group, we were able to boost the signal-to-noise ratio to a certain degree. This work is still ongoing and ho

Developing a Fast Algorithm for Classifying Unknown Proteins

Irene Kaplow, Stanford University

**Practicum Year:**2011

**Practicum Supervisor:**Zhong Wang, , Joint Genome Institute, Genomic Technologies Dept., Lawrence Berkeley National Laboratory

The DOE Joint Genome Institute recently sequenced the cow rumen metagenome. The rumen contained about two million genes, one million of which had strong sequence homology to known genes and one million of which were completely unidentifiable. This summer, we developed a method that will help classify unidentifiable genes from this and other metagenomic data-sets. In addition, it will speed up the classification of known genes.

Global constraints on atmospheric CO2 concentrations and CO2 sources

Scot Miller, Harvard University

**Practicum Year:**2011

**Practicum Supervisor:**Inez Fung, Faculty staff scientist, Earth Sciences Division, Lawrence Berkeley National Laboratory

I use a global-scale meteorological model and a statistical optimization framework in order to improve our understanding of global CO2 concentrations in the atmosphere. We also use this modeling and statistical framework to improve understanding of carbon dioxide sources globally. Our statistical framework allows us to use a much wider spectrum of surface, satellite, and meteorological observations than was previously possible, meaning that our framework will lead to a better understand of carbon dioxide globally than we have had in the past.

Detecting genomic-level regulatory strategies in Shewanella MR-1

Troy Ruths, Rice University

**Practicum Year:**2011

**Practicum Supervisor:**Adam Arkin, Faculty Scientist, Physical Biosciences, Lawrence Berkeley National Laboratory

The concentration levels of proteins -- dictated in part by gene expression -- determine the behavior of the cell; so it is common practice in microbiology to understand functional differences between two conditions using differential gene expression. Recent work by the Arkin lab, however, revealed a surprising non-correlation between gene expression and functional importance in the model bacterium Shewanella MR-1. In this project, I developed and applied statistical techniques to mine regulatory strategies made by Shewanella MR-1 on the genomic level that explains this non-correlation between expression and function. Specifically, I detected signatures of constitutive and anticipatory regulatory control across a compendium of over 100 growth and stress conditions.

Modeling stratified tidal flow with adaptive mesh refinement

Sean Vitousek, Stanford University

**Practicum Year:**2011

**Practicum Supervisor:**Phillip Colella, Dr., Applied Numerical Algorithms Group, Lawrence Berkeley National Laboratory

During my practicum I worked on simulating tidal (oscillatory) flow in a stratified estuary with complex bathymetry using an adaptive mesh refinement with embedded boundaries. The goal of this project was to understand the role turbulent mixing in a stratified estuary with a model that was capable of providing enough grid resolution to capture the flow features and the bathymetry (rather than using existing models which parametrize or under-resolve such features).

Stochastic Population Dynamics in Structured Populations

Carl Boettiger, University of California, Davis

**Practicum Year:**2010

**Practicum Supervisor:**Adam Arkin, Professor, The Division of Physical Biosciences, Lawrence Berkeley National Laboratory

This project investigated to what extent variation in population dynamics in lab populations of the tribolium flour beetle could be attributed to intrinsic factors such as demographic stochasticity rather than environmental variation. Previous research has only asserted stochastic differential equation forms which find minimal contribution, whereas I sought to derive the noise effects directly from the master equation for the population dynamics. I found that demographic stochasticity is insufficient to explain the variation without other mechanisms also being involved. I reached this conclusion partway through my practicum, but had become inspired to pursue a new question (also unrelated to my thesis work) on how such fluctuations could be used as early warning signals of a population catastrophe due to a bifurcation. I developed a set of computational methods that demonstrates the strengths and limitations of this approach in informing conservation policy.

Exploration of Novel Parallel Algorithms for Short Read Assembly

Scott Clark, Cornell University

**Practicum Year:**2010

**Practicum Supervisor:**Zhong Wang, Staff Scientist, DOE Joint Genome Institute, Lawrence Berkeley National Laboratory

(from an abstract of a talk I gave at the end)
I will talk about two projects that were developed over the summer as part of my DOE Computational Science Graduate Fellowship practicum at JGI. Both algorithms were designed to be parallel and scalable from the onset, allowing us to efficiently tackle the exponential growth of sequencing data. In the first project we explore growing assemblies through maximum likelihood scoring. We develop the formalism and background and discuss the future implementation. The second project will overview a mated read assembler model that uses the paired read information up front, distinguishing it from the existing assemblers like Velvet. Because the model builds the assembly locally it can be readily implemented in parallel, allowing for significant speed increases as well as allowing the algorithm to scale with the data. I will discuss the core algorithm and ongoing implementation.

Driving biofuel production with a synthetic oscillator

Tal Danino, University of California, San Diego

**Practicum Year:**2010

**Practicum Supervisor:**Jay Keasling, Dr., Berkeley National Labs, Lawrence Berkeley National Laboratory

In synthetically constructed pathways used to produce chemical compounds, biofuels, etc., inducible promoters are often used to drive various genes in a pathway. These promoters are constantly induced and are always producing mRNA transcripts and proteins during the lifetime of a cell. Constant production of transcripts can cause a high metabolic burden and therefore be detrimental to organism growth, and decrease yield of downstream products. Furthermore, certain products produced in a pathway can also be growth inhibiting or toxic, and therefore a switching between ON and OFF (such in an oscillatory system) of the responsible gene could improve pathway yield. If this is true, the optimal period of oscillations used to drive these genes will be interesting to discover and may lead to insight on design of synthetic gene circuits.

Creating hybrid climate-models

Virgil Griffith, California Institute of Technology

**Practicum Year:**2010

**Practicum Supervisor:**Gunther H. Weber, Computer Research Scientist, Computational Research Division, Lawrence Berkeley National Laboratory

There are numerous (~23) frequently used models of the global climate, some of which are better than others. My project was to create a hybrid climate model that is a hybrid model from the existing 23 that out-performs all of them in predicting global climate.

Electronic structure calculations to screen solar PV materials

Anubhav Jain, Massachusetts Institute of Technology

**Practicum Year:**2010

**Practicum Supervisor:**Jeff Neaton, Facility Director, Theory of Nanostructured Materials Facility, Lawrence Berkeley National Laboratory

The purpose of this study was to determine whether computations could aid in the identification of new solar PV materials. Previously, an analysis by LBL researchers Wadia

Because experimental materials performance data is limited to less than 100 materials (and probably closer to about 50), the analysis of Wadia et al. could not be performed on a large scale over a wide number of materials. The goal of this project was to determine whether materials performance could be found computationally, thus allowing (in theory) a limitless number of materials to be evaluated for their potential cost/watt and annual electricity production. Such a large data set could be screened for next-generation solar cells which were previously overlooked, like FeS

*et al.*identified several materials which were more promising than crystalline silicon (the dominant solar PV material) in terms of potential cost/watt and total potential energy generation capacity/year. This analysis had evaluated roughly 20 materials in total and relied both on published cost and materials availability estimates and experimental data on materials performance. One of their results was that FeS_{2}was particularly promising, and this led to renewed interest in this material by many researchers around the world.Because experimental materials performance data is limited to less than 100 materials (and probably closer to about 50), the analysis of Wadia et al. could not be performed on a large scale over a wide number of materials. The goal of this project was to determine whether materials performance could be found computationally, thus allowing (in theory) a limitless number of materials to be evaluated for their potential cost/watt and annual electricity production. Such a large data set could be screened for next-generation solar cells which were previously overlooked, like FeS

_{2}found by the limited test set of Wadia. Our short-term goal was to reproduce all of Wadia's results from a computational standpoint, thus providing the proof-of-concept for this approach. The challenge is that computations of band gap and absorption spectra, the quantities important for solar PV performance, are particularly challenging because they are excited state calculations whereas standard DFT, the bread and butter of electronic structure calculations, is technically a ground state method. We hoped to determine the most minimal computational techniques which could reproduce these quantities accurately so that our computational analysis could later be scaled to thousands of compounds for solar PV screening.A Comparison of Helicos and Illumina RNA-sequencing platforms

Eric Chi, Rice University

**Practicum Year:**2009

**Practicum Supervisor:**Paul T Spellman, Computational Scientist, Life Sciences, Lawrence Berkeley National Laboratory

A collection of cancer cells can be distinguished by the relative abundance of its transcribed mRNA (transcriptome). Being able to distinguish such heterogeneity among Ovarian cancer cells for example would be an important step in developing targeted therapies to the various subtypes of ovarian cancer. It's possible to measure how much of each kind of mRNA is present in a cell through direct sequencing. There are different technologies with their own benefits and downsides. We compared two different sequencing approaches on a common set of ovarian cell and cell line samples to assess the technical effects (e.g. measurement biases) introduced by each vendor's approach.

Engineering gold nanostructures for surface-enhanced Raman spectroscopy

Ying Hu, Rice University

**Practicum Year:**2009

**Practicum Supervisor:**James Schuck & Jeff Neaton, Staff Scientist & Facility Director, Molecular Foundry, Lawrence Berkeley National Laboratory

Title: Engineering gold nanostructures for surface-enhanced Raman spectroscopy
In this project, we applied finite-element method to investigate the far-field and near-field optical properties of several nanostructures, including gold-silica-gold multilayer nanoshells, gold bowtie antennas, and gold substrates with ordered nanoclusters. The goal of the study is to explore effective designs of plasmonic nanostructures to achieve spectral and spatial localization of light at nanometer scale. The investigation bears potential applications in mind such as a color sorter and substrates for surface-enhancement Raman spectroscopy.

Overall group: Chombo
my work: Artificial Viscosity for 4th order accurate Finite Volume Schemes

Eric Liu, Massachusetts Institute of Technology

**Practicum Year:**2009

**Practicum Supervisor:**Phillip Colella, Group Leader, Applied Numerical Algorithms Group, Lawrence Berkeley National Laboratory

Chombo is a set of tools designed for solving partial differential equations in a fast, scalable manner using finite volume methods. Some of their major selling points include higher-order (4th) accurate solution methods, (automated) adaptive mesh refinement, multiblock mesh capability, cut-cells, and a high degree of scalability. Chombo is designed to be usable by scientists and engineers with complex modeling problems regardless of their level of experience with numerical methods. The package provides a framework for solving a large variety of problems ranging from fluid dynamics to electrodynamics and many things in between.

DeepSky

Douglas Mason, Harvard University

**Practicum Year:**2009

**Practicum Supervisor:**Peter Nugent, Staff Scientist, Computational Cosmology Center, Lawrence Berkeley National Laboratory

In response to the needs of several astrophysics projects hosted at NERSC, Peter Nugent's group at LBNL is creating an all-sky digital image based upon the point-and-stare observations taken via the Palomar-QUEST Consortium and the SN Factory + Near Earth Asteroid Team. This data spans 9 years and almost 20,000 square degrees, with typically 10-100 pointing on a particular part of the sky. In total there are 11 million images. This data will be used to image and analyze stars with high proper motion, and to retrieve images of the stellar objects which precede supernovae and gamma-ray bursts.

Atmospheric Modeling using Chombo

Britton Olson, Stanford University

**Practicum Year:**2009

**Practicum Supervisor:**Phillip Colella, Senior Staff Scientist, Computational Research Division, Lawrence Berkeley National Laboratory

The Applied Numerical Algorithms Group (ANAG) at Lawrence Berkeley National Lab has been developing a library of Partial Differential Equation (PDE) software tools for use in a wide variety of applications. This software, CHOMBO, traditionally employed a 2nd order finite volume discretization coupled with a sophisticated Adaptive Mesh Refinement (AMR) scheme. More recently, this order has been extended to 4th order and the traditional rectilinear Cartesian grid, has been generalized to support structured curvilinear meshes. Finally, a multi-block capability is being added which enables non-aligning grids to interact. With the emergence of these new technologies, group leader Phillip Colella wanted to explore an application which employed all three.
Atmospheric models of the stratified atmosphere traditionally solve whats known as the Shallow Water Equations. We set out to solve these equations on the surface of a sphere and determine the advantages for this particular model, or using a high-order scheme.

Calculation of Side Chain Conformational Entropy and the Discrimination of Native Protein Structures from Decoy Sets

Jenelle Bray, California Institute of Technology

**Practicum Year:**2008

**Practicum Supervisor:**Teresa Head-Gordon, Professor, Physical Biosciences Division, Lawrence Berkeley National Laboratory

I developed a monte carlo algorithm to calculate the conformational entropy of side chains by determining the number of non-clashing side chain conformations for a protein backbone. This is being used to help find the native protein state from a set of decoys.

Can yields of a biofuel candidate Isopentenol be increased using oscillatory promoters?

Tal Danino, University of California, San Diego

**Practicum Year:**2008

**Practicum Supervisor:**Jay Keasling, Dr., Physical Biology, Lawrence Berkeley National Laboratory

In synthetically constructed pathways used to produce chemical compounds, biofuels, etc., inducible promoters are often used to drive various genes in a pathway. These promoters are constantly induced and are always producing mRNA transcripts and proteins during the lifetime of a cell. Constant production of transcripts can cause a high metabolic burden and therefore be detrimental to organism growth, and decrease yield of downstream products. Furthermore, certain products produced in a pathway can also be growth inhibiting or toxic, and therefore a switching between ON and OFF (such in an oscillatory system) of the responsible gene could improve pathway yield. If this is true, what will be the optimal period of oscillations used to drive these genes? I am working on answering this question by constructing various genes in a pathway under control of inducible and oscillatory promoters.

Computational Analysis of CRISPR Mechanism, Function, and Evolution

Robin Friedman, Massachusetts Institute of Technology

**Practicum Year:**2008

**Practicum Supervisor:**Jennifer A. Doudna, Professor, Physical Biosciences, Lawrence Berkeley National Laboratory

CRISPRs (Clustered, Regularly Interspaced Short Palindromic Repeats) are arrays of direct sequence repeats surrounded by spacers, flanked by CRISPR-associated (cas) genes, that provide acquired resistance to phage. This system is widespread in both Archea and Bacteria, but little is known about its mechanism of action or the function of individual cas proteins. We set out to better characterize its function and mechanism by computational analysis of spacer targets, of the evolutionary properties of CRISPR-containing genomes, and of the sequence and structural homologs of cas proteins.

The Design of Orthogonal Components for a Transcription Attenuation System

Sarah Richardson, Johns Hopkins University School of Medicine

**Practicum Year:**2008

**Practicum Supervisor:**Adam Arkin, Faculty Scientist, Physical Biosciences Division, Lawrence Berkeley National Laboratory

The Staphylococcus aureus plasmid pT181 uses an RNA-based transcription attenuation system to control plasmid replication and copy number. The Arkin lab is studying and modifying this system for use as a synthetic and inducible transcription control system in Escherichia coli. Such a nucleotide-based system would allow a rapid, mutable, and fine-grained control of gene expression that is impossible to achieve with protein-based methods like transcription factors. If successful, their system will be characterized by modularity (it will be made up of defined, discrete and flexible pieces), transferability (the attenuator will work inside different organisms) predictability (its behavior under different conditions will be well understood), orthogonality (pieces of one attenuator will not interfere with pieces of another), and composability (several attenuators may be effectively chained in series or stacked in a cascade). These four points are axiomatic attributes of every component in an electrical circuit, and it is the Arkin group's hope that by reproducing them biologically they will be able to take advantage of the great body of knowledge and tools available to electrical engineers and apply it directly to the field of synthetic biology.

High Performance 3-D Image Reconstruction for Molecular Structure Determination

Julianne Chung, Emory University

**Practicum Year:**2007

**Practicum Supervisor:**Chao Yang, Scientific Computing Group (SCG) Staff, Computational Research Division, Lawrence Berkeley National Laboratory

The single particle reconstruction process from cryo-electron microscopy (cryo-EM) consists of taking a collection of 2-D projection images from various angular orientations and recovering a 3-D volume representation. Accurate volume reconstruction can provide important information on complex molecular structures and their assemblies. However, the reconstruction process can be computationally challenging for large-volume structures, due to massive data and memory requirements. Current parallel implementations of reconstruction algorithms are inadequate for computing large-volume macromolecule structures, even on today's state-of-the-art high performance supercomputers. We propose an MPI parallel implementation which allows the volume data to be distributed among processors of a parallel computer, thereby overcoming the current per-processor memory limitations. In addition, we propose using a Lanczos-based reconstruction algorithm for cryo-EM data and show that this algorithm computes better reconstructions in fewer iterations.

Application of adaptive mesh refinement to particle-in-cell methods

Peter Norgaard, Princeton University

**Practicum Year:**2007

**Practicum Supervisor:**Phillip Colella, , Computing Sciences Directorate, Lawrence Berkeley National Laboratory

The applied numerical algorithms group at LBL has developed a software library known as Chombo for solving PDEs using the adaptive mesh refinement technique. The goal of my practicum project was to extend Chombo's capabilities to include the particle-in-cell method. In PIC, discrete charge carrying particles are allowed to move continuously in space, while the fields are solved at discrete grid points. As a necessary step, the charge from each particle must be assigned to the grid, and the force interpolated back to each particle. The presence of grid irregularity resulting from Chombo's block cartesian adaptive mesh requires special treatment of the charge assignment / force interpolation steps. We developed a correction method based on Green's functions which leads to a physically consistent method. Several test problems were developed to verify our technique.

Extremum-Preserving Limiters for the Piecewise Linear Method and Piecewise Parabolic Method

Michael Sekora, Princeton University

**Practicum Year:**2007

**Practicum Supervisor:**Phillip Colella, Senior Staff Scientist, Computational Research Division, LBNL, Lawrence Berkeley National Laboratory

The overarching project was to develop a Godunov-type method for solving multi-dimensional hyperbolic conservation laws that preserves fourth-order spatial accuracy. Such a method can be constructed by combining a technique for computing fourth-order face quadratures and the fourth-order spatial differencing of the unsplit Piecewise Parabolic Method. However, to ensure higher-order accuracy it was critical to reformulate van Leer and Parabolic Profile Limiters so that they did not reduce the problem to first-order accuracy at smooth extrema.

Automated Sorting of Large All Versus All BLAST Datasets

Benjamin Smith, Harvard University

**Practicum Year:**2007

**Practicum Supervisor:**Ernest Szeto, , Computational Research Div [CRBD], Lawrence Berkeley National Laboratory

Recent advances in rapid genome sequencing techniques have allowed the generation of terabytes worth of biological data which must be stored, processed, and distributed. Particularly important to the analysis of this data is the study of similarities in genome sequences via tools such as the Basic Local Alignment Search Tool (BLAST). Running "All versus All" BLAST searches has been made possible, even for very large datasets, via the use of supercomputers. However, the output of such a search is a large collection of files each containing many BLAST matches. To be useful, an efïficient method for sorting the output of the BLAST search by query and subject taxon, and then further sorting these results by score, must be found. A highly parallel, binary search tree based algorithm was tested here on ~475 GB of data, comprising the results of an "All versus All" BLAST search on 2788 taxons. Running on a cluster of 35 dual CPU nodes, the complete sort took just under 14 hours.

Modeling Ideal MHD collimation from differential rotation using an upwinding scheme with adaptive mesh refinement

Christopher Carey, University of Wisconsin

**Practicum Year:**2006

**Practicum Supervisor:**Phillip Colella, , Computational Research Division, Lawrence Berkeley National Laboratory

This project involved building a model of an astrophysical jet system using an existing magnetohydrodynamics (MHD) code built in the Chombo framework. Chombo is a framework for using upwinding methods for hyperbolic partial differential equations with adaptive mesh refinement. The model for the astrophysical jet system treats the accretion disk as a boundary condition, by applying a differentially rotating flow to fluid velocity on the bottom boundary. This flow winds up a small scale magnetic arcade which is set as the initial condition, and expands is to large length scales. This model will run with a computational domain with a length scale which is much larger than the characteristic length scale of the initial magnetic field. Thus, the boundary conditions on the outer boundaries will not have an influence on the evolution of the system. The adaptive mesh will allow for small scale structures in the solution fields to be resolved in this large domain.

Quaternary structure and stability of glycine mutated Amyloid Beta fibrils

Kevin Kohlstedt, Northwestern University

**Practicum Year:**2006

**Practicum Supervisor:**Teresa Head-Gordon, Assistant Professor, Physical Biosciences Division, Lawrence Berkeley National Laboratory

Amyloid beta (AB) fibril plaques have been recently studied due to their role in Alzheimers and Parkinson's disease. Although many recent structural studies have been done on the fibrils, still the arrangement of AB(1-40) monomers in the fibril remains a mystery. Using two quaternary structures proposed by Tycko, et al. for Wild Type AB(1-40) fibrils we studied the stability of the mutations in the two structures. A coarse grained model developed in the Teresa Head-Gordon was used in the dynamical simulations of the fibrils. The model is a four-bead group model, which groups the atoms in an amino acid into one bead. The flavor of the bead, which determines its interaction potential, is derived from its side chain properties. There are four flavors, strong hydrophobic, weak hydrophobic, neutral, and hydrophilic. Using this model we did simulation experiments using Langevin dynamics. We found the structure symmetric with the z-axis (axial to fibril) to be the more stable structure. We propose our model is able to capture the salient properties of the mutated fibril's structure and stability, which gives us insight into the behavior and morphology of the fibrils in vivo.

Stochastic modeling of HIV Tat-transactivation

Etay Ziv, Columbia University

**Practicum Year:**2006

**Practicum Supervisor:**Adam Arkin, Faculty Scientist, Physical Biosciences, Lawrence Berkeley National Laboratory

The long-term goal of the project is to build a stochastic model of gene expression in the HIV LTR (Long Terminal Repeat) which includes various mechanisms of transcriptional control. HIV establishes a long-lived latent reservoir in infected CD4 resting cells, and this is thought to be the major mechanism by which HIV thwarts eradication in patients. It has been hypothesized that the stochastic expression of the TAT protein controls the viral life-cycle decision.
Experimental collaborators have constructed strains containing the LTR promoter along with the Tat protein, tied to flourescent reporters. Using flow cytometer data from these constructs and various mutants, we seek to build a quantitative model of Tat expression.

Algorithm Refinement for the Stochastic Burgers' Equation

Jasmine Foo, Brown University

**Practicum Year:**2005

**Practicum Supervisor:**John Bell, Dr., Center for Computational Sciences and Engineering, Lawrence Berkeley National Laboratory

This summer we designed and implemented a numerical method using adaptive algorithm refinement for Burgers' equation with a stochastic flux term. This hybrid method couples a particle simulation of an asymmetric exclusion process to a godunov-type finite difference solver for the stochasic PDE (SPDE). The asymmetric exclusion process is a system of random walker particles whose behavior in the hydrodynamic limit matches the solution to the SPDE. The hybrid method thus utilizes this particle model in some regions of the physical domain, and a continuum model elsewhere. The solver adaptively chooses to use the particle method in regions of high regional gradient. Using this hybrid solver, we studied the dynamics of the system by investigating properties such as the drift of the shock location over time and the effect of turning off the random component in the SPDE region.

A Numerical Study of the Kelvin-Helmholtz Instability on Large Amplitude Internal Waves

Amber Jackson, University of North Carolina

**Practicum Year:**2005

**Practicum Supervisor:**John Bell, Group Leader, CCSE, Computational Research Division, Lawrence Berkeley National Laboratory

The project looked at large internal gravity waves in density stratified fluid. Adaptive Mesh Refinement Code (AMR) developed by LBNL was used to numerically simulate the experimental findings of a paper by Grue et al which presented the wave profiles, velocity profiles, wave propagation speeds, and instabilities associated with large internal solitary waves. Once satisfied that the numerics correctly captured the experiment, the code was also used to provide comparisons to new theoretical models purposed by Camassa and Choi. The final stage of the project involved looking specifically at Kelvin Helmholtz (KH) instabilities to characterize when and based on what parameters we should expect such instability to occur.

Towards a Mechanistic Model of Agonist-Induced Calcium Signaling in Macrophage Cells

Mala Radhakrishnan, Massachusetts Institute of Technology

**Practicum Year:**2005

**Practicum Supervisor:**Adam Arkin, Professor, Departments of Bioengineering and Chemistry, Lawrence Berkeley National Laboratory

Working with experimentalists in the Alliance for Cellular Signaling (AfCS) our goal is to construct a computational model that describes and predicts the time-dependent cytoplasmic calcium levels in cells as a function of external stimuli. The experimentalists generate large amount of time-dependent, dose-dependent, and knockdown-dependent calcium data, and we are using such data, as well as knowledge of the biological system, to create our model. In turn, our model allows us to suggest new experiments that can further refine the model. The cycle of experiment and model refinement is the hallmark of this project and will culminate in a model that captures the relevant effectors involved in Calcium signaling in the cell.
This project has great utility; such a model, if accurate and quantitative enough, may someday help inform drug designers about the efficacy and the cellular-level impact of an agonist, thus saving time, energy, and experimental resources.

Tomography of molecular machines central to mechanotransduction in hair cells

William Triffo, Rice University

**Practicum Year:**2005

**Practicum Supervisor:**Manfred Auer, Staff Scientist, Life Sciences, Lawrence Berkeley National Laboratory

Hair cells are mechanically sensitive cells that transduce mechanical signals into electrical information through displacements of hair-like stereocilia at their apex; they are medically important in hearing and balance disorders. Electron tomography utilizes transmission electron microscopy (TEM) to generate 3D density maps with nanometer resolution, enabling us to resolve molecular complexes within their native cellular environment. This project focused on two structures related to mechanotransduction: the rootlet at the base of the stereocilium, and the lateral wall of the outer hair cell (OHC), which plays a pivotal role in mammalian hearing. Tomography was used to identify the 3D architecture and potential molecular composition of macromolecular machines, and to use this structural information to deduce mechanical function.

Transfer-matrix approach to phonon defect scattering in three-dimensional carbon structures with thermoelectric applications

Brandon Wood, Massachusetts Institute of Technology

**Practicum Year:**2005

**Practicum Supervisor:**Joel Moore, Research Scientist, Materials Science, Lawrence Berkeley National Laboratory

The study was motivated by the possibility of using carbon nanotubes for direct thermal-to-electric energy conversion. Their theoretically high thermal conductivity and electrical tunability make them ideal candidates for thermoelectrics. In practice, however, defects in the nanotubes cause scattering of phonons (the primary thermal conduction mechanism) and inhibit their usefulness. The work I did involved using computational models to understand how randomly distributed defects (isotopic mass defects, in our prototypical example, although others are possible) affect the phonon dispersion spectrum of nanotubes. Having a better physical understanding of the problem could then lead to overcoming manufacturing barriers.

"Prediction of Regulatory Motifs Controlling the Expression and Processing of MicroRNA Genes in Drosophila"
"Prediction of Regulatory Motifs Controlling the Expression and Processing of MicroRNA Ge

Benjamin Lewis, Massachusetts Institute of Technology

**Practicum Year:**2004

**Practicum Supervisor:**Michael Eisen, Dr., Life Sciences, Lawrence Berkeley National Laboratory

By comparing genome sequences from multiple species, researchers may uncover functional regulatory elements that have been preserved in evolution. The genome sequences of related fruitfly species are an excellent resource for the identification and analysis of regulatory elements by multi-genome comparison. MicroRNA genes, a class of noncoding RNA genes, are an excellent subject for comparative genomics analyses of functional regulatory elements for several reasons: 1) microRNA genes are highly-conserved among divergent species 2) microRNA genes are small in size and may be easily-located in draft genome sequence 3) the expression of many microRNA genes is thought to be tightly-regulated in a cell-specific and developmental stage-specific manner. In this project, I have worked on the assembly and annotation of genome sequence data for 6 species of fruitfly. I have mapped all known microRNA genes in these genomes and I used comparative genomics methods to predict possible regulatory sequences involved in the expression of microRNA genes.

Improving the Performance of the Scaled Matrix/Vector Multiplication with Vector Addition in Tau3P, an Electromagnetic Solver.

Michael Wolf, University of Illinois at Urbana-Champaign

**Practicum Year:**2004

**Practicum Supervisor:**Esmond Ng, Scientific Computing Group Leader, National Energy Research Scientific Computing, Lawrence Berkeley National Laboratory

At the Stanford Linear Accelerator Center, I developed a parallel 3-D distributed-memory time domain electromagnetic solver (Tau3P) that uses unstructured meshes to model large particle accelerator structures. This code has been successful in solving accelerator structures consisting of millions of elements. However, during the development of Tau3P, I found that it was very difficult to obtain good parallel efficiency when a large number of processors were used in an attempt to reduce the runtime sufficiently for running large problems. Some preliminary work has been done on trying to find more optimal mesh partitioning using Zoltan from Sandia National Laboratories in order to obtain better parallel efficiency.
For my summer practicum, I focused on studying the communication patterns in Tau3P and trying different communication methods to reduce the communication overhead and increase parallel performance. In particular, I attempted to improve the parallel performance of the scaled matrix/vector multiplication with vector addition algorithm. I implemented over thirty different variations of this algorithm, exploring several different communication/computation orderings as well as many different modes of communication. I generated several different meshes so that I could examine the scalability of each algorithm as the problem size increases. I then ran the simulations on the IBM SP at NERSC. I found several communication schemes that were significantly better than the others, some surprisingly so.

A Fourth Order Accurate Adaptive Mesh Refinement Method for Poisson's Equation

Michael Barad, University of California, Davis

**Practicum Year:**2003

**Practicum Supervisor:**Phillip Colella, , Applied Numerical Algorithms Group, Lawrence Berkeley National Laboratory

In this project we developed a fourth order accurate numerical method for solving Poisson's equation, with adaptive mesh refinement (AMR) in 2 and 3 spatial dimensions, with either periodic, Neumann or Dirichlet boundary conditions. Our approach uses a conservative finite volume discretization combined with an efficient Multigrid elliptic solver. We use a cell-centered discretization, and using the divergence theorem, evaluate compact fourth order accurate fluxes at the faces of the cells.

Characterization of a Pulse-Generating Gene Circuit Using Quantitative Real-Time PCR

Michael Driscoll, Boston University

**Practicum Year:**2003

**Practicum Supervisor:**Adam Arkin, Faculty Scientist, Physical Biosciences Division, Lawrence Berkeley National Laboratory

Dr. Adam Arkin, a Faculty Scientist at the Lawrence Berkeley National
Laboratories, is presently engaged in a project to forward engineer
genetic regulatory circuits in E.Coli. Cells are multistable systems
which can reliably switch between states, for example during
development or cell division, based on environmental signals. This
capacity for switching states and the ability to maintain a given
state is a product of a cell's complex genetic regulatory circuitry.
I collaborated with a member of Dr. Arkin's group, Michael Cantor, in
building a pulse-generating genetic circuit in E.Coli. This gene
circuit consists of four genes coupled with promoter sequences
arranged on a plasmid vector. These four genes regulate one another
in such a way as to form a simple pulse-generating circuit. This
circuit is induced by the addition of the sugar arabinose, and
responds by transiently expressing a green fluorescent protein.
By providing a means of introducing a transient pulse of a gene
product into a system, the pulse generating circuit represents an
important tool for the study of gene networks.

Automatic segmentation and labeling of images of different cell types

Kristen Grauman, Massachusetts Institute of Technology

**Practicum Year:**2003

**Practicum Supervisor:**Bahram Parvin, Group leader, staff scientist, Computing Sciences, Imaging and Informatics Group, Lawrence Berkeley National Laboratory

The goal of this project was to use image processing and computer vision techniques to analyze images of breast tissue in order to segment and label the different cell types present in a given image. The eventual goal is to collect a large volume of data in this way, and then potentially establish correlations or patterns in the data that might explain the link between the tissue's biological condition and its appearance.

Learning RNA Secondary Structure using Marginalized Kernels on Graphs

Yan Karklin, Carnegie Mellon University

**Practicum Year:**2003

**Practicum Supervisor:**Stephen Holbrook, Dr., Physical Biosciences Division, Lawrence Berkeley N, Lawrence Berkeley National Laboratory

The goal of this project was to apply several new techniques in bioinformatics and machine learning to finding non-coding RNAs in genomes.
Recent work has identified an increasing number of small RNA molecules that do not code for proteins but are nevertheless transcribed from DNA. These "non-coding" RNAs have many different funtional roles, and the facilitation of their discovery by computational analysis of genome sequences is a useful undertaking.
There exist large databases of sequenced, functional family-annotated RNA. Using the primary sequence, it is possible to predict the secondary structure (base-pairing and folding) fairly reliably, and the shape of the folded non-coding RNAs may indicate the functional role of the molecules.
This project attempted to build a classifier to learn the typical shapes of secondary structures of RNA molecules. RNAs were represented as graphs, and (the recently developed) kernels between labeled graphs were used in the SVM classifier. Preliminary results suggest that this approach provides a decent alternative to sequence-based methods for finding non-coding RNAs.

Inference of Protein-DNA Binding Specificity From Structure

Julian Mintseris, Boston University

**Practicum Year:**2003

**Practicum Supervisor:**Michael Eisen, Dr., Life Sciences, Lawrence Berkeley National Laboratory

The goal of the project is to take advantage of the increasing amount of structural data available to try to improve our understanding of protein-DNA recognition and regulation. Specifically, we are interested in inferring protein-DNA binding profiles from structure databases to enable better prediction of cis-regulatory motifs in genomes.

Numerical Simulations of the Lyman Alpha Forest

Gregory Novak, University of California, Santa Cruz

**Practicum Year:**2003

**Practicum Supervisor:**Martin J. White, Professor, Physics, Lawrence Berkeley National Laboratory

The Lyman Alpha Forest (LAF) is a dense set of absorption lines seen
in the spectra of distant quasars due to clouds of neutral hydrogen
along the line of sight. The detailed structure of the LAF gives
information about both the spectrum and the amplitude of density
fluctuations in the early universe that gave rise to all present-day
structure in the universe.
The goal of the project was to produce a code to numerically simulate
the evolution of intergalactic gas from the big bang to a time when
the universe was 3-4 times smaller than its present size. The state
of this gas determines the statistical properties of the LAF that
would be seen by an observer who "lived in" the simulation.
By comparing the observed properties of the LAF to the properties of
the simulated LAF, it is possible to answer a number of questions
about the cosmological parameters of the universe. For example, it
has been shown (Croft 1998) that under a number of simplifying
assumptions, it is possible to use simple prescription for the state
of intergalactic gas in the early universe (rather than full fledged
simulation) to to measure the amplitude of the initial density
fluctuations in the universe to an accuracy of about ten percent.
However, one of the simplifying assumptions in this analysis is that
the ultraviolet background light in the early universe is
homogeneous. Direct simulation can quantify the effect of
inhomogeneities in the ultraviolet background, which can be expected
to degrade the accuracy of the measurement.
Croft, Weinberg, Katz and Hernquist, Astrophysical Journal 495:44
(1998).

Polymorphisms in Ciona intestinalis

Joshua Waterfall, Cornell University

**Practicum Year:**2003

**Practicum Supervisor:**Daniel Rokhsar, Faculty Scientist/Department Head, Physical Biosciences Division/JGI, Lawrence Berkeley National Laboratory

Analyzed shotgun genome sequence from several Ciona intestinalis individuals for polymorphisms, genetic differences between members of the same species. Studied correlations with and evidence for evolutionary selection on different features of the genome such as coding regions, intergenic regions, and regulatory sequences. Also developed tools to analyze DNA methylation in several of the genomes sequenced at the JGI, including Ciona intesinalis, Fugu rubripes and Human chromosome 17.

Investigation of quantum well states in Copper-Cobalt thin films.

Mary Ann Leung, University of Washington

**Practicum Year:**2002

**Practicum Supervisor:**Dr Andrew Canning, Research Scientist, Scientific Computing, Lawrence Berkeley National Laboratory

Project Description: We investigaged the properties of quantum well states in Copper/Cobalt thin fim systems using density functional theory. The Copper/Cobalt quantum well states are believed to be responsible for the "Giant Magneto-Resistance", or GMR effect that is responsible for the development of super dense, high-capacity hard disk drives. We investigated the properties of the quantum well states found in the Copper thin films by running simulations, using Paratec, a code that implements density functional theory and runs on parallel processign super computers. We inserted a Nickel atom at various locations in the thin film to determine effects on the quantum well states by the Nickel atom.

Implementing an embedded boundary algorithm for gas dymamics in EBChombo.

Benjamin Keen, University of Michigan

**Practicum Year:**2001

**Practicum Supervisor:**Dr. Phil Colella, , Applied Numerical Algorithms Research Group, Lawrence Berkeley National Laboratory

I implemented an embedded reflecting boundary algorithm for hyperbolic conservation laws (Modiano and Colella 2000). This is useful because it is a more general implementation of the algorithm than was done in the paper, and because it is one of the first applications using the EBChombo embedded boundary C++ application framework currently under development at ANAG.

Parallelization of the Ewald Summation Method in Molecular Dynamics

Heather Netzloff, Iowa State University

**Practicum Year:**2001

**Practicum Supervisor:**Teresa Head-Gordon, , Physical Biosciences and Life Science, Lawrence Berkeley National Laboratory

My project focused on parallelizing various methods to account for long-range forces in molecular dynamic (MD) simulations. The evaluation of forces is the major bottleneck to any MD simulation. Since the basic minimum image convention cutoff scheme does not accurately account for long-range interactions, new methods have been developed. These include the Ewald sum method in which the potential energy calculation is divided into a real-space sum and a reciprocal space sum. Since we desire to accurately simulation events such as protein folding in the condensed phase, the size of our system requires the use of parallel, efficient algorithms. There are several methods available to parallelize a MD code with various scalings; the focus of my research was to implement and benchmark these methods for future reference in protein simulations. The first two methods described divide the work between real and reciprocal space. In the atom decomposition method, each processor is given a static number of molecules; each processor computes the forces and new positions for its local molecules and then must share new positions with the rest of the processors. The force decomposition method involves assigning each process a portion of the force matrix. Each processor again contains a static number of local molecules, but will only compute a portion of the force matrix. The pertinent pieces of the force matrix are then shared among a smaller subset of processors to update total forces on local molecules. The particle mesh Ewald method puts emphasis on the reciprocal space sum; it is evaluated using fast Fourier transform with convolutions on a grid where charges are interpolated to the grid points. The fast multipole method, theoretical an O(N) method, calculates all forces in real space. The first two methods were parallelized during my summer practicum and research into the third was started. Reasonab

Modeling Multiphase Fluid Flow with Level Set Methods

Catherine Norman, Northwestern University

**Practicum Year:**2001

**Practicum Supervisor:**Ann Almgren, , Center for Computational Sciences and Engineering, Lawrence Berkeley National Laboratory

I worked with a variable density Navier Stokes solver which the Center for Computational Sciences and Engineering at LBL has written and used to model a variety of fluids problems. As part of ongoing efforts to improve and augment this model, I added code to track interfaces between fluids using level set methods.

RNA Gene Finding In Eukaryote

Catherine Quist, Cornell University

**Practicum Year:**2001

**Practicum Supervisor:**Dr. Stephen R. Holbrook, Staff Scientist, Physical Biosciences Division, Lawrence Berkeley National Laboratory

Identifying the "coding" regions of the human genome is one of the chief aims of the human genome project, or of any genome project for that matter. While this is a difficult problem, it is by no means impossible, especially if the search for coding regions is restricted to protein genes, i.e. stretches of DNA which are transcribed into mRNAs and then translated into proteins. Since all regions of DNA that code for a protein are flanked on one side by a start codon (three base pairs which signal to the ribosome to use a methionine to initiate construction of an amino acid chain) and on the other side by a stop codon (three base pairs which signal to the ribosome to release the mRNA and thereby halt translation), they may be relatively easily identified by a computer program using a hidden markov model.
If the search for "coding" regions is restricted to RNA coding regions instead of protein coding regions the problem of gene finding becomes much harder by comparison. While RNA genes are still transcribed, they are not translated and thus lack start and stop codons. A computer program which finds potential RNA genes cannot do so by searching for these markers. It must instead search for adjacent trascription regulatory regions and conserved structural motifs. Due to the complexities and subtleties of the patterns associated with these markers, along with our lack of a deeper understanding of either the protein-DNA complexes that initiate transcription or RNA folding, recognizing them is a very hard machine learning problem. Dr. Stephen R. Holbrook at LBNL has written a program, which uses neural networks to do RNA gene finding in prokaryotes. I spent the summer modifying this program to do RNA gene finding in Yeast, the simplest Eukaryotic organism.

Project 1 (A) "Archiving the 1999 AMANDA Data Set" Project 2 (B) "Enlarging the Gamma-Ray Burst Sample for AMANDA"

Rellen Hardtke, University of Wisconsin

**Practicum Year:**2000

**Practicum Supervisor:**Stewart Loken and George Smoot, , , Lawrence Berkeley National Laboratory

(A) Each year data collected by the Antarctic Muon and Neutrino Detector Array (AMANDA)is copied to DLT tapes at the South Pole experimental site. This set of -70 DLT tapes contains more than 1 terabyte of data and is hand-carried out of Antartica when the site becomes accessible in November. Once these tapes reach LBNL, they must be first be uploaded to the National Energy Research Scientific Computing Center (NERSC) mass storage facility before the data can be filtered and utilized by scientists. Quality checks and monitoring must be undertaken, as well as annual improvements and updates made in the archiving scrips.
(B) One of AMANDA's scienfic projects is the search for neutrinos from gamma-ray bursts (GRBs). These extremely energetic phenomena are little understood and the detection of neutrinos from these events would be a major discovery. Prior to this summer, only 78 GRBs had been studied for coincident neutrinos. My goal was to increase the number of GRBs that AMANDA could examine by finding a new source of detected GRBs and extracting the necessary data from AMANDA archives at LBNL.

Computational Neuroscience of C. Elegans and Mammalian Retinal Ganglion Cells During Development

Asohan Amarasingham, Brown University

**Practicum Year:**1999

**Practicum Supervisor:**Daniel Rokhsar, , , Lawrence Berkeley National Laboratory

Measuring the Extensional Elasticity of the Red Blood Cell Using Micropipette Aspiration

William Marganski, Boston University

**Practicum Year:**1999

**Practicum Supervisor:**Jamie Butler, , , Lawrence Berkeley National Laboratory

The extensional elasticity of the red blood cell represents the major force of recovery after large cell deformations. The source of the extensional elasticity is the cytoskeleton underneath the cell membrane. This protein scaffolding consists of a spectrin network that is interconnected via short actin filaments and protein 4.1. The elasticity of the cytoskeletal network is dependent upon its linkage to the lipid membrane via band 3 and ankyrin and its association with other transmembrane proteins such as glycophorin A. The elasticity of the red cell's protein scaffolding can be quantitated by mechanical aspiration into a suction pipette. An elastic modulus for the cytoskeleton of the red blood cell can be calculated by measuring the aspiration length of the cell within the micropipette as the suction pressure is increased.

Implementation Issues in Haptic Visualization

J. Dean Brederson, University of Utah

**Practicum Year:**1998

**Practicum Supervisor:**Dr. Edward Bethel, , Visualization Group, Lawrence Berkeley National Laboratory

The goal of this research is to solve several of the problems inherent in integrating haptic interfaces with scientific visualization environments. In particular, a PHANTOM haptic interface was used as the device of interaction. Two specific problems that were the goal of the practicum were to develop a closed-loop calibration scheme for the device and a set of low-level device drivers to enable accurate haptic rendering within a calibrated workspace at high servo rates.

Numerical Experiments with Optimal Predictions Method

Eugene Ingerman, University of California, Berkeley

**Practicum Year:**1998

**Practicum Supervisor:**Dr. Alexandre Chorin, , , Lawrence Berkeley National Laboratory

My project consisted of studying the applicability of the optimal predictions method to several types of the differential equations.

Image Recovery Using Multiple Holograms

Brandoch Calef, University of California, Berkeley

**Practicum Year:**1997

**Practicum Supervisor:**Dr. Malcolm Howells, , AFRD, Lawrence Berkeley National Laboratory

Two or more holograms are made of an object at different distances. The holograms are digitized and loaded into a computer. The problem then is to reconstruct an image of the object. Iterative techniques were devised to achieve this.

Using the Space Transformations to solve the Three Dimensional Flow Equation

Marc Serre, University of North Carolina

**Practicum Year:**1997

**Practicum Supervisor:**Dr. Garrison Sposito, , Earth Sciences Division, Lawrence Berkeley National Laboratory

New formulations of the space transformations have been derived and implemented for bounded flow domains. The new formulations are more accurate than previous methods. Additionally this new implementation was tested on the T3E supercomputer.

Integration of physical and genetic genome maps

John Guidi, University of Maryland

**Practicum Year:**1995

**Practicum Supervisor:**Dr. Manfred Zorn, , Information and Computing Sciences, Lawrence Berkeley National Laboratory

The delineation of the ordering of chromosomal loci is an important objective of biomedical research. A variety of experimental techniques exist that provide order information at different levels of resolution, using various coordinate systems. This project focuses on the issues involved in integrating disparate ordering information, taking into consideration the uncertainties and ambiguities that are present.

Slug Tests Under Saturated/Unsaturated Flow: A Richard's Equation Perspective

Philip Weeber, University of North Carolina

**Practicum Year:**1995

**Practicum Supervisor:**Dr. T.N. Narasimhan, , Earth Sciences Division, Lawrence Berkeley National Laboratory

Applying computer model to physical problem of groundwater flow through porous media.

Parallel Computation using OVERFLOW: Overlapping Grid Navier-Stokes Flow Solver

Edwin Blosch, University of Florida

**Practicum Year:**1993

**Practicum Supervisor:**Dr. James Sethian, , Physics Division, Lawrence Berkeley National Laboratory

An overlapping grid scheme based on the Chimera approach has been developed by NAS for solving the N-S eqn's on the iPSC/860 MIMD computer. The code was benchmarked on three model flow problems: (1) supersonic flow over a wedge, (2) transonic flow over a wing at angle-of-attack, and (3) obliques shock wave/turbulent boundary layer interaction with bleed to prevent separation. Parallel computing issues have also been investigated, includiing the effect of domain decomposition of parallel efficiency.

Examination of Image Reconstruction from Transmission Data

Jack Lemmon, Georgia Institute of Technology

**Practicum Year:**1993

**Practicum Supervisor:**Dr. Thomas Budinger, , Center for Functional Imaging, Lawrence Berkeley National Laboratory

Examined the reconstruction of medical images from position emissions using existing algorithms.