4:15pm to 5:15pm

4:15pm to 5:15pm

Traditional nonconvex optimization procedures often rely on good initializations and give locally optimal solutions. On the other hand, methods that learns a the function between and their parameters either cannot handle complicated forms of mapping, or require enormous number of training samples. In particular, to achieve a minimax guarantee of \epsilon error, the sample complexity required is O(1/epsilon^d), an order that is likely to be optimal without knowing any knowledge of the function but its local smoothness.

In this talk, I will demonstrate in the specific case of image deformation, how we can achieve worst-case (minimax) guarantees with lower sample complexity, using more structure specified by the generative model of the data. The sample complexity is reduced to to O(C^d log1/\epsilon) with an iterative regression procedure, if compositional structures of deformation are considered, and further reduced to O(C1^d+C2 log1/\epsilon) with top-down hierarchical procedure, if local patch structure is considered. Following this line, we bridge nonconvex optimization with learning-based approaches, and put many effective heuristics used in previous works into a unified and theoretical framework. In the future, I will continue exploring this direction to achieve a better theoretical understanding how Deep Learning works.

4:15pm to 5:15pm

**This Week's Speaker: Lexing Ying**

**Stanford University**

This seminar series is designed for a student audience. The idea of the seminar is for faculty to describe their research, their philosophy or specific projects in broad terms at a level that students can easily follow. The style is informal to encourage a dialog rather than a one-way flow. Computer projectors are rarely used except to display images. Also the lack of faculty in the audience enables the student to be the ones asking the questions. The less preparation the presenter makes the better. Anything that is not completely familiar to the speaker is not something that the students need to know.Since this is informal for the speaker we hope that there will a large number of faculty that will respond to request that they speak. We also extend the list of speakers to include senior students with a particular focus on those who no longer sit in ICME space and do not have an ICME adviser. The class is also encouraged to make suggestions for speakers and can if they wish invite speakers.

9:30am to 12:30pm

ICME and the Stanford Research Computing Center, in conjunction with SAP, to offer a two-session short course in using SAP HANA, SAP's massively parallel, in-memory database and application server platform. The HANA platform accelerates analytics performance over very large databases. HANA simply makes things possible that weren’t possible before. For example, what could a 10,000-fold improvement in query performance do for you?

Get hands-on training to start prototyping and developing on SAP HANA. Join us and your peers at this two-session short course, which will examine the technological advances and innovations that make SAP HANA so high performing. During the short course, you will install, configure, and use SAP HANA tools over real data. We'll consider the possibilities and potential impact of real-time, big data analytics for science, business, and research.

Topics:

- SAP HANA Technical Overview
- Setting up the SAP HANA Development Environment
- Modeling and SQL Script in SAP HANA
- Native Application Development with SAP HANA
- Text Analytics, Geospatial Visualization, Predictive Analytics, and more

**Dates and Times: **

- January 31
^{st},9:30 AM - 12:30 PM - February 7
^{th}, 9:30 AM - 12:30 PM

**Location: **Turing Auditorium, Polya Hall Room 111**Cost: **No charge to attend. Food and beverages will be served.

**Learn more and register at: **http://stanford.io/hana-at-stanford.

4:15pm to 5:15pm

TBD

4:15pm to 5:15pm

Whether the 3D incompressible Euler equations can develop a singularity in finite time from smooth initial data is one of the most challenging problems in mathematical fluid dynamics. This work attempts to provide an affirmative answer to this long-standing open question from a numerical point of view, by presenting a class of potentially singular solutions to the Euler equations computed in axisymmetric geometries. The solutions satisfy a periodic boundary condition along the axial direction and no-flow boundary condition on the solid wall. The equations are discretized in space using a hybrid 6th-order Galerkin and 6th-order finite difference method, on specially designed adaptive (moving) meshes that are dynamically adjusted to the evolving solutions. With a maximum effective resolution of over $(3 \times 10^{12})^{2}$ near the point of the singularity, we are able to advance the solution up to $\tau_{2} = 0.003505$ and predict a singularity time of $t_{s} \approx 0.0035056$, while achieving a \emph{pointwise} relative error of $O(10^{-4})$ in the vorticity vector $\omega$ and observing a $(3 \times 10^{8})$-fold increase in the maximum vorticity $\norm{\omega}_{\infty}$. The numerical data is checked against all major blowup (non-blowup) criteria, including Beale-Kato-Majda, Constantin-Fefferman-Majda, and Deng-Hou-Yu, to confirm the validity of the singularity. A local analysis near the point of the singularity also suggests the existence of a self-similar blowup. We also discuss a 1D model which can be viewed as a local approximation to the Euler equations near the solid boundary of the cylinder. The finite-time blowup of this 1D model is proved under certain convexity conditions for the velocity field.

4:15pm to 5:15pm

Nothing polarizes opinions in the machine learning world more than the topic of competitions - especially, perhaps, Kaggle competitions. Some people view competitions' razor focus on objective predictive performance as having a pernicious influence, leading to a disregard of computational complexity, real insight, scientific purity, or elegant simplicity.

On the other hand, there are those who view competitions as being the only "true" way to ascertain pragmatically the capabilities of machine learning algorithms. They view researchers who aren't prepared to submit their algorithms to the rigors of competition as hiding from objective scrutiny. They see the success of competitions in schools as showing that only competitions really engage students with the exciting potential of machine learning.

So where is the truth? In this talk, Kaggle's Director of Engineering, will engage with both the pros and cons of competitions, and will explain what has been learnt from the last three years of Kaggle competitions. He will look at case studies both from academic and industrial competitions, and will draw out some of the themes coming from the most successful performers.

-Short Bio - Ben Hamner is the Director of Engineering Kaggle. He has worked with machine learning problems in a variety of different domains, including natural language processing, computer vision, web classification, and neuroscience. Prior to joining Kaggle, he applied machine learning to improve brain-computer interfaces as a Whitaker Fellow at the École Polytechnique Fédérale de Lausanne in Lausanne, Switzerland. He graduated with a BSE in Biomedical Engineering, Electrical Engineering, and Math from Duke University.

8:00am to 5:00pm

Alan George, Michael Saunders, and Jim Varah turn 70 during this year. We'll be celebrating their birthdays and accomplishments at Stanford University on Saturday, January 25, 2014 (with a banquet dinner that evening). Students, friends, and colleagues are all welcome to attend. This meeting also marks the 10-year anniversary of the first SVG Meeting,

For addional information and to register please click here.

4:15pm to 5:15pm

**This Week's Speaker: Andrew Spakowitz**

**Stanford University **

This seminar series is designed for a student audience. The idea of the seminar is for faculty to describe their research, their philosophy or specific projects in broad terms at a level that students can easily follow. The style is informal to encourage a dialog rather than a one-way flow. Computer projectors are rarely used except to display images. Also the lack of faculty in the audience enables the student to be the ones asking the questions. The less preparation the presenter makes the better. Anything that is not completely familiar to the speaker is not something that the students need to know.Since this is informal for the speaker we hope that there will a large number of faculty that will respond to request that they speak. We also extend the list of speakers to include senior students with a particular focus on those who no longer sit in ICME space and do not have an ICME adviser. The class is also encouraged to make suggestions for speakers and can if they wish invite speakers.

4:15pm to 5:15pm

Inverse problems arise in scientific applications such as biomedical imaging, computer graphics, computational biology, and geophysics, and computing accurate solutions to inverse problems can be both mathematically and computationally challenging. We develop a new framework for solving inverse problems by incorporating training data to compute an optimal regularized inverse matrix. In the first part of the talk we describe methods for computing optimal spectral filters, for cases where the SVD is available. In the second part we describe methods for computing an optimal low-rank regularized inverse matrix, for cases where the forward model is not known.

Joint work with Dianne O'Leary (University of Maryland, College Park).

3:30pm to 5:00pm

Abstract: OpenMP 4.0 is the current release of the OpenMP API specification. It added new major features to the OpenMP language to incorporate latest technological trends in HPC and beyond. The new features will significantly increase the expressiveness of OpenMP and its applicability for complex HPC codes. In this presentation, we will provide an in-depth overview of the new features. The presented features include user-defined reductions, support for SIMD instructions, support for accelerators/coprocessors, and affinity.

*Bio: Michael Klemm is part of Intel's Software and Services Group, Developer Relations Division. His focus is on High Performance and Throughput Computing. Michael holds a Doctor of Engineering degree (Dr.-Ing.) in Computer Science from the Friedrich-Alexander-University Erlangen-Nuremberg, Germany. Michael's areas of interest include compiler construction, design of programming languages, parallel programming, and performance analysis and tuning. Michael is Intel representative in the OpenMP Language Committee and leads the efforts to develop error handling features for OpenMP.*

4:15pm to 5:15pm

__Inviscid limits for the stochastic Navier-Stokes equation__

We discuss recent results on the behavior in the infinite Reynolds number limit of invariant measures for the 2D stochastic Navier-Stokes equations. Invariant measures provide a canonical object which can be used to link the fluids equations to the heuristic statistical theories of turbulent flow. We prove that the limiting inviscid invariant measures are supported on bounded vorticity solutions of the 2D Euler equations. This is joint work with N. Glatt-Holtz and V. Sverak.

4:00pm to 5:15pm

**This Week's Speaker: Lexing Ying ****(POSTPONED) **

This seminar series is designed for a student audience. The idea of the seminar is for faculty to describe their research, their philosophy or specific projects in broad terms at a level that students can easily follow. The style is informal to encourage a dialog rather than a one-way flow. Computer projectors are rarely used except to display images. Also the lack of faculty in the audience enables the student to be the ones asking the questions. The less preparation the presenter makes the better. Anything that is not completely familiar to the speaker is not something that the students need to know.Since this is informal for the speaker we hope that there will a large number of faculty that will respond to request that they speak. We also extend the list of speakers to include senior students with a particular focus on those who no longer sit in ICME space and do not have an ICME adviser. The class is also encouraged to make suggestions for speakers and can if they wish invite speakers.

4:15pm to 5:15pm

We study the stationary states of the semi-relativistic Schroedinger-Poisson system in the repulsive (plasma physics) Coulomb case. In particular, we establish the existence and the nonlinear stability of a wide class of stationary states by means of the energy-Casimir method. We generalize our global well-posedness result for the semi-relativistic Schroedinger-Poisson system to spaces with higher regularity

4:15pm to 5:15pm

Modern technology enables collecting data of unprecedented size and complexity. To extract useful information from these data, it is critical to collect high-quality labels (judgments) and annotations. Due to the advent of many online crowdsourcing services (e.g., Amazon Mechanical Turk), an effective way for collecting labels is to ask a distributed crowd of workers to label the instances in exchange for micropayments.

In this talk, we discuss an important statistical decision problem in crowdsourcing: the optimal budget allocation for binary/multi-class labeling task. In particular, we address the problem of allocating a predetermined budget among all instances so that the overall labeling accuracy can be maximized. Since different instances have different ambiguities and different workers have different reliabilities, this problem is very challenging. Using tools from Bayesian statistical decision theory, we formulate the allocation problem as a Bayesian Markov Decision Process (MDP). Then one can obtain the optimal allocation policy for any given budget via the dynamic programming (DP) . However, the DP algorithm is computationally intractable. To address the challenge, we propose a new approximate policy, called optimistic knowledge gradient policy, which is a consistent policy and leads to superior empirical performance. This work is jointly done with Qihang Lin and Dengyong Zhou.

Xi Chen is current a Postdoc in the Group of Prof. Michael Jordan at UC Berkeley and will join Department of Information, Operations, and Management Sciences at Stern School of Business at New York University in Sept. 2014. He obtained his Ph.D. from the Machine Learning Department at Carnegie Mellon University (CMU); and his Masters degree in Industry Administration and Operations Research from the Tepper School of Business at CMU.

He is developing fast and scalable algorithms for parametric and non-parametric structured sparse learning problems with applications to text mining and climate modeling. He also investigates machine learning foundations for collective intelligence, in particular, crowdsourcing.

4:00pm to 5:15pm

4:15pm to 5:15pm

Special Seminar

January 9th, 2014

4:15PM, Bldg. 380, Room 383-N

Lin Lin (Lawrence Berkeley National Laboratory)

**Fast algorithms for electronic structure analysis**

Kohn-Sham density functional theory (KSDFT) is the most widely used electronic structure theory for molecules and condensed matter systems. For a system with N electrons, the standard method for solving KSDFT requires solving N eigenvectors for an O(N) * O(N) Kohn-Sham Hamiltonian matrix. The computational cost for such procedure is expensive and scales as O(N^3). We have developed pole expansion plus selected inversion (PEXSI) method, in which KSDFT is solved by evaluating the selected elements of the inverse of a series of sparse symmetric matrices, and the overall algorithm scales at most O(N^2) for all materials including insulators, semiconductors and metals. The PEXSI method can be used with orthogonal or nonorthogonal basis set, and the physical quantities including electron density, energy, atomic force, density of states, and local density of states are calculated accurately without using the eigenvalues and eigenvectors. The recently developed massively parallel PEXSI method has been implemented in SIESTA, one of the most popular electronic structure software packages using atomic orbital basis sets. The resulting method can allow accurate treatment of electronic structure in a unprecedented scale. We demonstrate the application of the method for solving graphene-like structures with more than 30,000 atoms, and the method can be efficiently parallelized 10,000 - 100,000 processors on Department of Energy (DOE) high performance machines.

2:00pm to 5:00pm

_

4:15pm to 5:15pm

8:30am to 5:00pm

We are pleased to announce that the Bay Area Scientific Computing Day (BASCD) will be hosted by LBNL on December 11, 2013.

BASCD is an annual one-day meeting focused on fostering interactions and collaborations between researchers in the fields of scientific computing and computational science and engineering from the San Francisco Bay Area. The event provides junior researchers a venue to present their work to the local community, and for the Bay Area scientific and computational science and engineering communities at large to interchange views on today’s multidisciplinary computational challenges and state-of-the-art developments.

The speakers at this year’s meeting are Kevin Carlberg (Sandia), Erin Carson (UC Berkeley), Lixin Ge (SLAC), Jeff Irion (UC Davis), Lex Kemper (LBNL), Christian Linder (Stanford University), Ding Lu (UC Davis), Ali Mani (Stanford University), François-Henry Rouet (LBNL), Cindy Rubio-Gonzalez (UC Berkeley), Khachik Sargsyan (Sandia) and Samuel Skillman (SLAC).

This year we will have a combined lunch and poster session. If you are interested in presenting a poster, please indicate this during registration.The schedule, registration, and the title and abstract of the speakers are available on the BASCD website: https://sites.google.com/a/lbl.gov/bay-area-scientific-computing-day/

4:15pm to 5:15pm

4:15pm to 5:15pm

Circulant matrices play a central role in a recently proposed formulation of three-way data computations. In this setting, a three-way table corresponds to a matrix where each "scalar" is a vector of parameters defining a circulant. This interpretation provides many generalizations of results from matrix or vector-space algebra. We derive the power and Arnoldi methods in this algebra. In the course of our derivation, we define inner products, norms, and other notions. These extensions are straightforward in an algebraic sense, but the implications are dramatically different from the standard matrix case. For example, a matrix of circulants has a polynomial number of eigenvalues in its dimension; although, these can all be represented by a carefully chosen canonical set of eigenvalues and vectors. These results and algorithms are closely related to standard decoupling techniques on block-circulant matrices using the fast Fourier transform.

4:15pm to 5:15pm

Universal properties related to the KPZ equation is an extremely active research field right now. It is connected to many different areas in mathematics and mathematical physics: probability theory and statistical mechanics, PDEs, random matrices, integrable systems, and many others. While it was a big recent progress, mainly in deriving exact formulas for correlation functions, the problem of universality is still largely open. We shall discuss some recent results in this direction, as well as possible approaches to establishing of the full universality.

4:15pm to 5:15pm

Examining the electronic structure of a material requires quantum mechanical methods, and in my studies, this is density functional theory (DFT), a computationally intensive technique that originates in traditional physics such as semi-conductor research, but which is increasingly applied in fields such as earth sciences and biology. Studying the behavior of electrons is important because electrons are responsible for all chemical reactivity, and understanding this leads to control over processes. For example, understanding how CO_{2} interacts with the materials present in underground storage environments, contributes to determining the conditions required for secure CO_{2} sequestration and long-term storage.

In this presentation I will show you the tools of my trade and how I have applied DFT to the investigation of clay minerals within gas-shale, within the wider context of CO_{2} sequestration; to the tentative exploration of the formation of fossil oils; to the undesirable behavior of clays within cement and finally, to the preliminary investigation into the formation of peptides within the context of Origins of Life studies.

4:15pm to 5:15pm

Over the past few years, parallel sparse direct solvers have made significant progress [1]. They are now able to solve efficiently real-life three-dimensional problems having several millions of equations. The ongoing hardware evolution exhibits an escalation in the number and the heterogeneity of computing resources. PaStiX is a parallel sparse direct solver based on a dynamic scheduler for modern hierarchical architectures [2].

Recently, a comparative study of the performance of the PaStiX solver over generic DAG-based schedulers has been performed. The analysis demonstrates that these runtimes provide a uniform and portable programming interface across heterogeneous environments, and are therefore a sustainable solution for hybrid architectures (multiple CPUs and GPUs) [3]. Nevertheless, the complexity and the need for a large amount of memory are still a bottleneck in these methods.

The incomplete factorization technique usually relies on scalar implementations and thus does not benefit from superscalar effects provided by modern high performance architectures. Such methods are also difficult to parallelize. We have implemented a method that exploits the parallel blockwise algorithmic approach used in the framework of high performance sparse direct solvers in order to develop robust parallel incomplete factorization based preconditioners for iterative solvers [4]. On the numerical side, we are now studying how the data sparseness that might exist in some dense blocks appearing during the factorization can be exploited using different compression techniques based on H-matrix arithmetic (and variants). Some first attempts have already been investigated. In a recent work, X. S. Li and colleagues have considered the HSS-matrix representation in the context of a sparse multifrontal factorization technique to design an efficient sparse parallel direct solver for the solution of 3D Helmholtz equations [5]. The MUMPS and CHOLMOD solvers [6,7] also investigate the use of low-rank approximations [6,7].

This research activity is conducted in the framework of the FastLA associate team and will naturally irrigate the hybrid solvers that are also investigated and will closely interact with the other research efforts where similar data sparseness might be exploited [8].

[1] Anshul Gupta. Recent progress in general sparse direct solvers.

Lecture Notes in Computer Science, 2073:823-840, 2001.

[2] Pascal Henon, Pierre Ramet, and Jean Roman. PaStiX: A

High-Performance Parallel Direct Solver for Sparse Symmetric Definite

Systems. Parallel Computing, 28(2):301-321, January 2002.

[3] Xavier Lacoste, Pierre Ramet, Mathieu Faverge, Yamazaki Ichitaro, and Jack Dongarra. Sparse direct solvers with accelerators over DAG runtimes. Research Report RR-7972, INRIA, 2012.

[4] Pascal Henon, Pierre Ramet, and Jean Roman. On finding approximate supernodes for an efficient ILU(k) factorization. Parallel Computing, 34:345-362, 2008.

[5] X. S. Li. Towards an optimal-order approximate sparse factorization exploiting data-sparseness in separators. Workshop Celebrating 40 Years of Nested Dissection, July 22-23, 2013, Waterloo.

[6] Patrick Amestoy, Alfredo Buttari, Guillaume Joslin, Jean-Yves

L'Excellent, Mohamed Sid-Lakhdar, Clément Weisbecker, Michele Forzan,

Cristian Pozza, Remy Perrin, Valène Pellissier. Shared memory parallelism and low-rank approximation techniques applied to direct solvers in FEM simulation. To appear in IEEE Transactions on Magnetics, Extended selected short papers from Compumag 2013 conference.

[7] David S. Bindel and Jeffrey N. Chadwick. An Efficient Solver for

Sparse Linear Systems Based on Rank-Structured Cholesky Factorization.

Workshop Celebrating 40 Years of Nested Dissection, July 22-23, 2013,

Waterloo.

[8] FastLA is an associate team between INRIA project-team HiePacs,

Scientific Computing Group in the Computational Research Division in

Lawrence Berkeley National Laboratory and the Institute for

Computational and Mathematical Engineering at Stanford University, funded from 2012 to 2013.

http://people.bordeaux.inria.fr/coulaud/projets/)FastLA_Website/

4:00pm to 5:15pm

4:15pm to 5:15pm

We will describe the hot spots conjecture of J. Rauch, known results and counterexamples. After that I will consider arguably the simplest unknown case: acute triangles. We will discuss recent progress sparked by a new method due to Miyamoto.

4:15pm to 5:15pm

In this presentation I will address some recent results in model reduction of parametric and nonlinear problems by focusing on two classes of methods. The first class of techniques consist of Reduced Basis (RB) methods aiming at complexity reduction for parametric partial differential equations by projection techniques. Basic ingredients are an offline/online decomposition, rigorous and rapid a-°©‐posteriori error estimation for certification of the reduced simulation results and Greedy/POD-°©‐Greedy procedures for construction of provable quasi-°©‐optimal approximation spaces. We show some results for nonlinear problems, including hyperbolic transport problems, two-°©‐phase flow in porous media and contact-°©‐problems involving inequality constraints.

A second class of methods are Kernel Methods for generating approximative models of nonlinear functions. These powerful machine learning techniques can be used for sparse vectorial function approximation, for example by vectorial support vector regression or vectorial kernel greedy procedures. The resulting approximants allow efficient complexity reduction in projection-°©‐based model order reduction or in multiscale problems as demonstrated in applications from biomechanics and porous media flow.

4:15pm to 5:15pm

When a colleague knocks on your door with a mathematical problem, great joy ensues on both sides if you happen to have numerical software that can solve the problem. This is a reason for writing "general-purpose software". The Systems Optimization Laboratory (SOL) was founded by George Dantzig and Richard Cottle at Stanford University in 1974 to encourage algorithm and software development in traditional Operations Research areas. We review the optimization and linear system solvers that have been developed as a result of the SOL initiative long ago. They include MINOS, LSSOL, QPSOL, NPSOL, SQOPT, SNOPT, LUSOL, MINRES, MINRES-QLP, LSQR, LSMR, LSRN, and PDCO. We also review some of the resulting applications. A theme of the talk is that one never knows what general-purpose software will be used for. Presented at the workshop "Computational Linear Algebra and Optimization for the Digital Economy", University of Edinburgh, Scotland, Oct 31-Nov 1, 2013.

Our goal is to bring together scientists from various theoretical and application fields to join forces in solving complex scientific computing problems. It covers recent developments in numerical linear algebra and numerical optimization. The speakers are frequently guests from other institutions and the local industry.

4:00pm to 5:15pm

4:15pm to 5:15pm

The development of drug resistance is a major challenge in the treatment of cancer. In this talk we will overview some of the aspects of drug resistance that have been studied by the mathematical community. We will focus on two examples: 1) Modeling the dynamics of cancer stem cells and their role in developing drug resistance. When combined with clinical and experimental data, our mathematical analysis provides new insights on how to approach treating Chronic Myelogenous Leukemia. 2) Studying the role of cell density and mutations on the dynamics of drug resistance in solid tumors. This is another example in which the mathematical analysis leads to insights on the design of new treatment protocols. *This is a joint work with C. Tomasetti, J. Greene, O. Lavi, and M. Gottesman.*

4:15pm to 5:15pm

Geologic materials are heterogeneous and their macroscopic response is often dominated by changes in the material at lower scales. For this reason, development of predictive modeling capabilities that represent deformation and failure in geologic materials remains a significant scientific challenge. In part this is due to a lack of comprehensive understanding of the complex physical processes that underlie this behavior, and in part due to numerical challenges that prohibit accurate representation

Of the heterogeneities that influence the material response. Recent advances in modeling capabilities coupled with modern high performance computing platforms enable physics-°©‐based simulations of heterogeneous media with unprecedented details, offering a prospect for significant advances in the state of the art. This presentation provides an overview of some of the modern computational approaches under development at LLNL, discusses their advantages and limitations, and presents simulation results demonstrating the application of these numerical methods to practical problems involving heterogeneous geologic materials subjected to extreme dynamic loading environments.

4:15pm to 5:15pm

Location: ICME Lobby, Suite 060 (Huang Engineering Building)

Date: Friday November 8th, 2013 @ 4:15pm

We consider in this work multiphase flows in porous media with complex interfacial dynamics and/or high Bond and capillary numbers preventing the use of classical quasi-static approaches. Current limitations of the modeling of such flows are first highlighted in a brief presentation of my Phd work concerning direct-numerical and pore-network simulations of multiphase flow in packed beds. Then, I present a mixed approach of these two methods considering two distinct flow areas : one-phase domain which can be treated by a classical pore-network approach, and a multiphase domain close to the interfaces simulated by direct simulations on movable DNS-windows. To avoid multiple costly adaptive mesh refinements, a penalization method for two-phase flow is developed to represent local pore-scale geometry on a simple Cartesian mesh.

4:15pm to 5:15pm

Multiphysics and multimodel problems are ubiquitous in many scientific and engineering applications, such as thermal cooling, combustion modeling, fluid-structure interaction, contact dynamics. Traditional numerical methods rely on either Neumann-Neumann/Robin-Robin preconditioners for the discrete Steklov-Poincare operator, or mortar methods and Lagrangian multiplier. As an alternative, in this study we have proposed and developed a new general optimization-based framework where we minimize the difference between scalar and vector fields at the domain interfaces subject to constraining physics in each domain. Two advantages of this approach are that

- an adjoint-based optimization approach allows for the solution of defective boundary conditions and design problems without nested iterations;
- the approach admits different numerical discretizations such as finite elements or finite differences in each physical domain, enabling a wide range of applications.

Joint work with Judith Hill (Oak Ridge National Lab).

Performed under the auspices of the US Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.

3:00pm to 4:00pm

4:15pm to 5:15pm

I will discuss some recent results on developing new factorizations for matrices obtained from discretizing differential and integral operators. A common ingredient of these new factorizations is the interpolative decomposition for numerically low-rank matrices. As we shall see, these factorizations offer efficient algorithms for applying and inverting these operators. This is a joint work with Kenneth Ho.

4:15pm to 5:15pm

Life on earth has evolved utilizing the unique redox chemistry of iron. It is an essential element for virtually all life forms due to its presence in heme-°©‐containing proteins, With biological functions including oxygen transportation, Chemical catalysis, and electron transfer. Electron Exchange between ferrous and ferric iron determines the availability of iron in the biosphere by influencing the form of key iron-°©‐bearing minerals and their interactions with trace polyvalent contaminant metals. Under Environmentally relevant conditions this exchange involves interaction between aqueous ferrous iron and solid-°©‐phase ferrous iron oxides and oxyhydroxides, with complex involvement of solid-°©‐state charge migration. Examples Include Fe(II)-°©‐catalyzed Recrystallization of hematite and goethite, and mixed-°©‐valent spinels such as magnetite acting as a mineralogic source and sink for reactive Fe(II) Due to its topotactic solid-°©‐ solution property. Ferrous-°©‐ferric Electron exchange is also essential for microbial respiration via the evolution-°©‐optimized molecular machinery present in metal-°©‐respiring bacteria. This Machinery transmits current across their cell membranes using redox metalloprotein modules comprised of distinct multiheme c-°©‐type cytochromes. This Presentation explores the dynamics of ferrous-°©‐ferric electron exchange in such systems at the atomic and nanoscale levels from theory, computational molecular simulation, spectroscopy, and microscopy. It Will particularly emphasize use of molecular simulation and modeling strategies for predicting the thermodynamics and kinetics of ferrous-°©‐ferric electron transfer reactions, based on the Marcus Theory framework. More Generally the topic also illustrates the importance of computational molecular science for making fundamental advances in understanding the environmental biogeochemistry of the earth’s near-°©‐surface.

4:15pm to 5:15pm

Deducing the state or structure of a system from partial, noisy measurements is a fundamental task throughout the sciences and engineering. The resulting inverse problems are often ill-posed because there are fewer measurements available than the ambient dimension of the model to be estimated. In practice, however, many interesting signals or models contain few degrees of freedom relative to their ambient dimension. For example, a small number of genes may constitute the signature of a disease, very few parameters may specify the correlation structure of a time series, or a sparse collection of geometric constraints may determine a network configuration. Discovering, leveraging, or recognizing such low-dimensional structure plays an important role in making inverse problems well-posed.

In this talk, I will propose a unified approach to transform notions of simplicity and latent low-dimensionality into convex penalty functions. This approach builds on the success of generalizing compressed sensing to matrix completion and greatly extends the catalog of objects and structures that can be recovered from partial information. I will focus on a suite of data analysis algorithms designed to decompose general signals into sums of atoms from a simple---but not necessarily discrete---set. These algorithms are derived in an optimization framework that encompasses previous methods based on l1-norm minimization and nuclear norm minimization for recovering sparse vectors and low-rank matrices. I will provide sharp estimates of the number of generic measurements required for exact and robust estimation of a variety of structured models. I will contextualize these results in several example applications.

Benjamin Recht is an Assistant Professor in the Department of Electrical Engineering and Computer Sciences and the Department of Statistics at the University of California, Berkeley. Ben was previously an Assistant Professor in the Department of Computer Sciences at the University of Wisconsin-Madison. Ben received his BS in Mathematics from the University of Chicago, and received a MS and PhD from the MIT Media Laboratory. After completing his doctoral work, he was a postdoctoral fellow in the Center for the Mathematics of Information at Caltech. He is the recipient of an NSF Career Award, an Alfred P. Sloan Research Fellowship, and the 2012 SIAM/MOS Lagrange Prize in Continuous Optimization.

4:00pm to 5:15pm

4:15pm to 5:15pm

Many objects in solid state physics have properties that vary dramatically depending on the arithmetics of certain physically relevant and measurable parameters. Quasiperiodic operators is one field where such phenomena are rampant. We will discuss how this can be interpreted in a non-contradictory way and, in particular, will present some recent continuity/discontinuity results.

4:15pm to 5:15pm

An often substantial barrier faced by technical staff in developing and applying new algorithms or addressing novel problems at scale is the work needed to create that part of the code infrastructure that is common to many simulation codes, from the need to handle user input to generating and analyzing output from simulations. The problem is compounded by the manifold concerns inherent in migrating between platforms and in addressing scalable, distributed memory parallelization. This talk will discuss the development of the GEOS framework at LLNL and discuss detailed examples of how the framework has been used to address several different types of computational geosciences problems, from understanding the mechanisms and sensitivity of injection-induced seismicity to the modeling of hydraulic fracture stimulation. The aim here is to demonstrate not only how the process of software design for massively parallel physical simulation can be simplified and modularized but also how such an approach can accelerate research in computational geosciences and facilitate interaction between researchers.

9:00am to 4:00pm

**Register now for Stanford ICME and NVIDIA CUDA on Campus event.**

**What is the event: **Half day scientific computing symposium & half day hands on GPU computing workshop.

** ****Who is it for:** Undergraduate, Graduate Students, Postdocs, Researchers, and Professors.

**Why you should attend:** Hear from fellow researchers focused on best practices, trends, tools, and research findings discovered using GPU computing. Meet NVIDIA leadership, engineers and industry experts.

**Registration:** https://stanford-gpuworkshop.eventbrite.com

**Location: **Huang Building, 475 Via Ortega, Stanford University – McKenzie Room

** **Stanford University, an NVIDIA CUDA Center of Excellence, and NVIDIA are organizing CUDA on Campus. Stanford faculty and researchers will speak about of how heterogeneous computing with GPUs is important to sustaining and advancing the state of the art in scientific and research computing.

NVIDIA GPUs are the world’s fastest and most efficient accelerators delivering world record scientific application performance. NVIDIA’s CUDA Technology is the most pervasive parallel computing model, used by over 250 scientific applications and over 150,000 developers worldwide.

**Agenda:**

**Symposium – 9AM to Noon**

Research Symposium where you will hear from fellow researchers on tools, research discoveries, best practices and trends using GPU computing for computational research.

- Keynote: Accelerated Computing – Past, Present, Future – Ian Buck, Ph.D. VP of GPU Computing, NVIDIA
- Scientific Computing Symposium – Presented by Stanford Faculty and Researchers in Chemistry, Earth Sciences, Medicine, Engineering

**GPU Computing Hands-on Training – 1PM to 4PM**

GPU Computing workshop with hands-on exercises in GPU Computing where you will learn how to program GPUs via the use of libraries, OpenACC compiler directives, and CUDA programming.

- Track 1 – Intro to GPU Computing Hands-on Exercises
- Track 2 – Intermediate/Advanced Hands-on Exercises

**Presenter Biographies:**

**Ian Buck, Ph.D. Vice President GPU Computing, NVIDIA**

Ian Buck is NVIDIA’s Vice President for GPU Computing Software, responsible for all engineering, 3rd party enablement, and developer marketing activities for GPU Computing at NVIDIA. Ian joined NVIDIA in 2004 and created CUDA, which remains the established leading platform for accelerated based parallel computing. Before joining NVIDIA, Ian was the development lead on Brook which was the forerunner to generalized computing on GPUs. He holds a Ph.D. in Computer Science from Stanford University and B.S.E from Princeton University.

*Coffee & **Lunch will be provided. *

*Please bring your laptop to participate in hands-on exercises.*

4:00pm to 5:15pm

4:15pm to 5:15pm

In many inverse problems, the nonlinear least squares optimization requires the solution of a large number of large linear systems per iteration. This is expensive, and the ever increasing number of measurements possible through advances in engineering increases the computational burden further. In this talk, we show how techniques from interpolatory model reduction can drastically reduce the cost of general inverse problems where the reconstructions are parameterized with a relatively modest number of parameters. In particular, we reduce the cost of solving forward problems. We discuss how surrogate models can approximate both the cost functional and the associated Jacobian with little loss of accuracy but significantly reduced cost.

The quality and performance of our method is demonstrated for several synthetic diffuse optical tomography (DOT) problems. An important result is that projection bases for reduced models developed for a single reconstruction can be used without further expensive linear solves for very different reconstructions. We will (try to) explain this result and discuss interesting related (open) questions in parametric inversion, model reduction, optimization, and linear solvers.

This describes joint work with Serkan Gugercin, Chris Beattie, Saifon Chaturantabut (all VT), Misha Kilmer and Meghan O'Connell (Tufts).

4:15pm to 5:15pm

*AAPG Foundation J. Ben Carsey and AAPG/SEG Inter-Society Distinguished Lecturer*

* Joe Stefani received degrees in engineering and geophysics from Cal and Stanford. Since 1984, he has worked for Chevron Energy Technology Company, during which time he has been involved in a range of geophysical R&D, including high fidelity earth and seismic modeling, acquisition, anisotropy, inversion, and general Aki & Richards stuff. Most recently he has helped to build the SEG SEAM Phase 1 and Phase 2 earth models.*

Earth modeling, from the construction of subsurface structure and stratigraphy, to the accurate understanding of rock physics, through the simulation of seismic and nonseismic responses, is an enabling technology to guide decisions in acquisition, processing, imaging, inversion and reservoir property inference, for both static and time-lapse understanding. So it is crucial to capture those earth elements that most influence the geophysical phenomena we seek to study. This is notoriously difficult, probably because we regularly underestimate how clever the earth can be in producing various geophysical phenomena.

The main part of the talk focuses on methods we have used in building complex earth models (both overburden and reservoirs) and their seismic simulations, emphasizing the challenge to reproduce the appropriate features observed in real data. Questions to consider are the quality of the seismic data that will act as a guide in the model building, and that of the well logs used to quantify the rock physics. Another consideration is the amount of physics to include in the geophysical response simulation, which is a tradeoff between computational load and acceptable characterization of the data features.

Finally, the industry workhorse for seismic modeling continues to be the time-domain finite-difference (FD) algorithm, mainly because of its balance between accuracy and efficiency, simple concept and gridding, and ease of programming on various hardware platforms. Because of this simplicity, and the growing interest in time-lapse and geomechanical problems, a short treatment is included of how FD modeling can be adapted to problems in rock physics and geomechanics from core to basin scales.

4:15pm to 5:15pm

The butterfly algorithm efficiently approximates a discrete analogue of an integral transform whose kernel satisfies a certain simple geometric low-rank condition; this class of kernels is quite broad and includes, among others, non-uniform FFT's, generalizations of Radon transforms, and spherical harmonic transforms. This talk discusses a recently introduced distributed-memory parallelization of the algorithm that has been shown to efficiently strong scale to tens of thousands of cores. The implementation has been released within the open source library DistButterfly.

4:00pm to 5:15pm

4:15pm to 5:15pm

The development of high-performance computing and numerical techniques has enabled global and regional tomography to reach high levels of precision, and seismic adjoint tomography has become a state-of-the-art tomographic technique.

Adjoint tomography uses full waveform simulations and back projection to compute finite frequency sensitivity kernels. These kernels describe the variation of the discrepancy (or misfit) between observed seismic data and modeled synthetics as a function of the model parameters. They are used in an iterative inversion aiming at minimizing the misfit function, thereby recovering model parameters.

This inverse approach benefits from an accurate numerical technique to solve the seismic wave equation in complex 3D media, in the first place. Here I use a spectral-element method, which contrary to finite-element methods (FEM), uses high degree Lagrange polynomials, allowing the technique to not only handle complex geometries, like the FEM, but also to retain the strength of exponential convergence and accuracy due to the use of high degree polynomials. After describing spectral-element and adjoint methods, I will discuss two applications: (1) a 3D adjoint tomography for the Middle East to improve seismic waveform predictions in the area, and (2) results on seismic source parameters inversion for seismic monitoring.

4:15pm to 5:15pm

We consider random-walk transition matrices from large social and information networks. For these matrices, we describe and evaluate a fast method to estimate one column of the matrix exponential. The method we employ is a coordinate descent method applied to a large linear system derived from the random walk matrix and customized to the matrix exponential. Under some conditions on the maximum degree of the graph, we can prove a sublinear runtime, and for real-world networks with over 5 million edges, we find it runs in less than a second on a standard desktop machine.

4:00pm to 5:15pm