Title: Predictive Models of Inertial Confinement Fusion

Abstract:

Multiphysics hydrodynamic simulations are used to understand inertial confinement fusion (ICF) experiments and are used to design future experiments at the National Ignition Facility. These simulations are computationally expensive to run and do not capture all of the physics observed in experiments. Deep neural networks can be used to build powerful predictive models, mapping the simulation inputs (e.g. ICF physics parameters and laser inputs) to various quantities of interest, such as neutron yield, nuclear bang time, and x-ray images. However, most deep neural networks yield point estimates with no information on how certain the network is in its prediction. Furthermore, deep neural networks trained only on simulation data cannot fully capture the physics observed in experiments. In this talk I will present current efforts to: identify the various sources of uncertainty in deep neural networks trained on ICF simulation data and equip the networks with physics from experiments using transfer learning.

Title: Capturing near-fields in plasmonic structures with corners

Abstract:

Plasmonic structures are commonly made of dielectrics and metals. At optical frequencies metals exhibit unusual electromagnetic properties like a dielectric permittivity with a negative real part whereas dielectrics have a positive one. This configuration allows the propagation of electromagnetic surface waves strongly oscillating at the metal-dielectric interface, and hyper-oscillating if the interface presents corners. Standard numerical methods to study surface plasmons excitation do not always take into account the multiple scales inherent in electromagnetic problems which may lead to inaccurate predictions. In this presentation we present some techniques to accurately compute and efficiently take into account the multiple scales of 2D light scattering problems in plasmonic structures with corners.

Title: Learning model parameters with uncertainty using statistical inference of hypothetical data sets

Abstract:

Closure models employed in predictive simulations of physical systems often rely on comparison of model outputs to noisy experimental data to constrain parameter values. The noise and bias in these experimental data result in uncertainty in the model parameters, which must be propagated through the predictive simulations for results to be meaningful. Information on model parameters and their uncertainties is often incomplete or inconsistent with the context of the predictive model. These issues could be corrected if the original data is available for re-analysis, which is often not the case for legacy experiments. Using concepts from entropic inference and approximate Bayesian computation, we explore a space of hypothetical data consistent with the available information reported from the experiments. Analysis of this ensemble of consistent data delivers the desired joint parameter uncertainties that may have been missing from the original fitting, and also allows for the combination of information across experiments using a data centric approach.

Title: Scalable shallow learning: Optimal complexity kernel ridge regression via hierarchical matrices for high-energy physics datasets

Abstract:

We present an optimal complexity approximation of the dense kernel ridge regression (KRR) matrix that is extensively used in machine learning classification tasks. The latest iteration of this work features a novel compression scheme based on approximate nearest neighbors, as opposed to randomized sampling, within the hierarchical semiseparable matrices (HSS) framework. Additionally, we discuss a hybrid MPI+OpenMP parallel implementation which is benchmarked against alternative methods using a set of well-known scientific datasets coming from high-energy physics with the goal of identifying signatures of exotic particles.

Title: Automated Audience Segmentation Using Reputation Signals

Abstract:

Selecting the right audience for an advertising campaign is one of the most challenging, time-consuming and costly steps in the advertising process. To target the right audience, advertisers usually have two options: a) market research to identify user segments of interest and b) sophisticated machine learning models trained on data from past campaigns. In this paper we study how demand-side platforms (DSPs) can leverage the data they collect (demographic and behavioral) in order to learn reputation signals about end user convertibility and advertisement (ad) quality. In particular, we propose a reputation system that learns interest scores about end users, as an additional signal of ad conversion, and quality scores about ads, as a signal of campaign success. Then our model builds user segments based on a combination of demographic, behavioral and the new reputation signals. And it recommends transparent targeting rules that are easy for the advertiser to interpret and refine. We perform an experimental evaluation on industry data that showcases the benefits of our approach for both new and existing advertiser campaigns.

Title: Shallow Learning for Flow Reconstruction with Limited Sensors

Abstract:

Numerous applications across the physical, biological, and engineering sciences generate image like data and prominent examples are fluid flows. These type of data are known to be sparse in some domain and it should be possible to reconstruct the high-dimensional state-space vectors from only a small set of measurements. In this talk, we discuss a neural network-based learning methodology to train an inverse operator from sensor measurements to reconstructed state-space through limited example data. More concretely, we treat the problem as a shallow inversion problem. No prior information is assumed to be available and the estimation method is purely data-driven. The shallow network structure we advocate is flexible, fast and robust to train, and accommodates nonlinear measurements. We demonstrate the performance of this approach against popular modal approximations techniques on two examples in fluid mechanics and oceanography.

Title: Deep learning for probability density and flow map approximation of high dimensional nonlinear systems

Abstract:

State-of-the-art computational optimal control methods for uncertain dynamical systems rely on sample trajectories to evaluate performance metrics. As the number of such trajectories increases, the dimension of the optimal control problem rapidly becomes intractable. In order to address this challenge, we propose a new approach for probability density function (PDF) approximation that emphasizes the role of data-driven PDF equations over extensive numerical integration. Specifically, we use deep neural nets to efficiently and accurately approximate the time-evolving PDF and flow map of high-dimensional nonlinear systems, as well as propose improvements to physics informed neural nets and efficient data-gathering schemes.

Title: Real-world benefits of Machine Learning in healthcare

Abstract:

This talk will survey on all of our efforts in Applied Machine Learning lab in Computer Science and Engineering department of University of Santa Cruz. We will particularly discuss state-of-the-art machine learning healthcare applications and our efforts in those areas, different data sources we are using and how data drives our investigations, ethics of using algorithms in healthcare, and future applications and direction of the lab.

Title: Machine Learning Closure Modeling for Reduced-Order Models of Dynamical Systems

Abstract:

Projection-based reduced-order models are a popular tool to generate computationally efficient approximations to complex systems. It is important to both quantify and reduced the epistemic uncertainty generated by such approximations. This work investigates the use of machine learning techniques along with the variational multiscale method and Mori-Zwanzig formalism to quantify and reduce errors in projection-based models of dynamical systems. The presented approach utilizes equation-based features derived from the variational multiscale method and Mori-Zwanzig formalism as inputs to neural and recurrent neural networks with the aim of developing local subgrid-scale closures and global error models of reduced-order models of dynamical systems. Results are shown for time-dependent problems relevant to fluid mechanics.

Title: Machine learning for X-ray Lasers

Abstract:

X-ray Free Electron Lasers (XFELs) are among the most complex accelerator projects in the world today. With large parameter spaces, sensitive dependence on beam quality, huge data rates, and challenging machine protection, there are expanding opportunities to apply machine learning (ML) to XFEL operation. In this talk I will summarize some promising ML methods for XFELs, and highlight recent examples of successful applications at the Linac Coherent Light Source at SLAC.

Title: Building Physically Consistent Surrogates using Generative Models

Abstract:

Computational modeling of complex dynamical systems is prevalent in a wide variety of scientific applications. Driven by the widespread adoption of machine learning in surrogate modeling, deep neural networks have become prominent in scientific applications. Despite the success of deep learning, an important, yet often overlooked, consideration is that the networks must be conditioned to produce scientifically accurate predictions. While in some cases, this conditioning can be achieved by placing constraints on the space of outputs, in practice, it is often non-trivial to translate prior knowledge into mathematical constraints that can be integrated into the optimization process. Hence, there is a strong disconnect between highly parameterized surrogate models and the simulator. In this talk, I will present a data-driven approach for enabling DNNs to be consistent with the physical processes governing the data. In particular, I will motivate the use of deep generative models in this context and describe how utilizing multiple disparate outputs from the simulator helps in producing physically consistent surrogates.

Title: Using Deep Learning for the solution of Partial Differential Equations

Abstract:

Deep learning is a state-of-the-art technique that is widely used in various fields like computer vision and natural language processing. However, only recently has it been utilized to solve engineering problems. In engineering, many problems involve a low-dimensional input but low-dimensional output (the quantity of interest). The classical example is solving the solution of partial differential equations (PDEs). We present an implementation and experimental results using a novel recurrent neural network (RNN) with customized cells that solve systems described by PDEs. In such systems, given low-dimensional inputs, we use an RNN with customized cells to directly solve for the low-dimensional output.

Benchmarks will be presented for the following three problems: (1) Solve for the highest temperature in a two-dimensional time-dependent heat transfer problem. (2) Find the location where the maximum speed occurs in Burgers equation. (3) Determine the position of the wave source in a wave propagation problem. In all three problems, the error rate achieved (with a smaller number of parameters than long-short term memory (LSTM)) is lower than with LSTM and Gated Recurrent Unit (GRU).