Publications

We present a new strategy for automatically exploring the design space of key CUDA + MPI programs and providing design rules that discriminate slow from fast implementations. In such programs, the order of operations (e.g., G PU kernels, MPI communication) and assignment of operations to resources (e.g., G PU streams) makes the space of possible designs enormous. Systems experts have the task of redesigning and reoptimizing these programs to effectively utilize each new platform. This work provides a prototype tool to reduce that burden. In our approach, a directed acyclic graph of CUDA and MPI operations defines the design space for the program. Monte-Carlo tree search discovers regions of the design space that have large impact on the program's performance. A sequence-to-vector transformation defines features for each explored im-plementation, and each implementation is assigned a class label according to its relative performance. A decision tree is trained on the features and labels to produce design rules for each class; these rules can be used by systems experts to guide their implementations. We demonstrate our strategy using a key kernel from scientific computing - sparse-matrix vector multiplication - on a platform with multiple MPI ranks and GPU streams.

Abstract not provided.

Iterative methods for solving linear systems serve as a basic building block for computational science. The computational cost of these methods can be significantly influenced by the round-off errors that accumulate as a result of their implementation in finite precision. In the extreme case, round-off errors that occur in practice can completely prevent an implementation from satisfying the accuracy and convergence behavior prescribed by its underlying algorithm. In the exascale era where cost is paramount, a thorough and rigorous analysis of the delay of convergence due to round-off should not be ignored. In this paper, we use a small model problem and the Jacobi iterative method to demonstrate how the Coq proof assistant can be used to formally specify the floating-point behavior of iterative methods, and to rigorously prove the accuracy of these methods.

Calibrating a finite element model to test data is often required to accurately characterize a joint, predict its dynamic behavior, and determine fastener fatigue life. In this work, modal testing, model calibration, and fatigue analysis are performed for a bolted structure, and various joint modeling techniques are compared. The structure is designed to test a single bolt to fatigue failure by utilizing an electrodynamic modal shaker to axially force the bolted joint at resonance. Modal testing is done to obtain the dynamic properties, evaluate finite element joint modeling techniques, and assess the effectiveness of a vibration approach to fatigue testing of bolts. Results show that common joint models can be inaccurate in predicting bolt loads, and even when updated using modal test data, linear structural models alone may be insufficient in evaluating fastener fatigue.

Detailed analysis of both the line-intensity ratios and line shapes of the K-lines of elements of different abundances (Fe, Cr, Ni, and Mn) emitted from the stagnation of a steel wire-array implosion on Z, were used to determine the line opacities. While the opacities at the early time of stagnation appear to be consistent with a nearly uniform hot-plasma cylinder on-axis surrounded by a colder annulus, the opacities during the peak K-emission strongly suggest that the main K-emission is due to small hot regions (spots) spread over the stagnating column. The spots are shown to be at least 4× denser than expected based on a uniform-cylinder emission (namely, ni > 3 ×1020 cm-3 ), are of diameters of about 200 μ or less (where the smaller the spots the higher are the densities), and are thousands in number. The total mass of the spots was determined to be 3-10 % of the load mass, and their total volume 3-15 % of the O 1.2-mm stagnation-column volume, both are less than the respective values for the earlier period of lower K power.

This work evaluates the benefits of using a 'smart' network interface card (SmartNIC) as a compute accelerator for the example of the MiniMD molecular dynamics proxy application. The accelerator is NVIDIA's BlueField-2 card, which includes an 8-core Arm processor along with a small amount of DRAM and storage. We test the networking and data movement performance of these cards compared to a standard Intel server host using microbenchmarks and MiniMD. In MiniMD, we identify two distinct classes of computation, namely core computation and maintenance computation, which are executed in sequence. We restructure the algorithm and code to weaken this dependence and increase task parallelism, thereby making it possible to increase utilization of the BlueField-2 concurrently with the host. We evaluate our implementation on a cluster consisting of 16 dual-socket Intel Broadwell host nodes with one BlueField-2 per host-node. Our results show that while the overall compute performance of BlueField-2 is limited, using them with a modified MiniMD algorithm allows for up to 20% speedup over the host CPU baseline with no loss in simulation accuracy.

High-performing teams learn intelligent and efficient communication and coordination strategies to maximize their joint utility. These teams implicitly understand the different roles of heterogeneous team members and adapt their communication protocols accordingly. Multi-Agent Reinforcement Learning (MARL) seeks to develop computational methods for synthesizing such coordination strategies, but formulating models for heterogeneous teams with different state, action, and observation spaces has remained an open problem. Without properly modeling agent heterogeneity, as in prior MARL work that leverages homogeneous graph networks, communication becomes less helpful and can even deteriorate the cooperativity and team performance. We propose Heterogeneous Policy Networks (HetNet) to learn efficient and diverse communication models for coordinating cooperative heterogeneous teams. Building on heterogeneous graph-attention networks, we show that HetNet not only facilitates learning heterogeneous collaborative policies per existing agent-class but also enables end-to-end training for learning highly efficient binarized messaging. Our empirical evaluation shows that HetNet sets a new state of the art in learning coordination and communication strategies for heterogeneous multi-agent teams by achieving an 8.1% to 434.7% performance improvement over the next-best baseline across multiple domains while simultaneously achieving a 200× reduction in the required communication bandwidth.

The development of the High-Resolution Wavelet Transform (HRWT) is driven by the need of increasing the high-frequency resolution of widely used discrete Wavelet Transforms (WTs). Based on the Stationary Wavelet Transform (SWT), which is a modification of the Discrete Wavelet Transform (DWT), a novel WT that increases the number of decomposition levels (therefore increasing the previously mentioned frequency resolution) is proposed. In order to show the validity of the HRWT, this paper encompasses a theoretical comparison with other discrete WT methods. First, a summary of the DWT and the SWT, along with a brief explanation of the WT theory, is provided. Then, the concept of the HRWT is presented, followed by a discussion of the adherence of this new method to the WT's common properties. Finally, an example of the application is performed on a transient waveform analysis from a power system fault event, outlining the benefits that can be obtained from its usage compared to the SWT.

With the increase in penetration of inverter-based resources (IBRs) in the electrical power system, the ability of these devices to provide grid support to the system has become a necessity. With standards previously developed for the interconnection requirements of grid-following inverters (GFLI) (most commonly photovoltaic inverters), it has been well documented how these inverters 'should' respond to changes in voltage and frequency. However, with other IBRs such as grid-forming inverters (GFMIs) (used for energy storage systems, standalone systems, and as uninterruptable power supplies) these requirements are either: not yet documented, or require a more in deep analysis. With the increased interest in microgrids, GFMIs that can be paralleled onto a distribution system have become desired. With the proper control schemes, a GFMI can help maintain grid stability through fast response compared to rotating machines. This paper will present an experimental comparison of commercially available GFMIand GFLI ' responses to voltage and frequency deviation, as well as the GFMIoperating as a standalone system and subjected to various changes in loads.

Metal additive manufacturing allows for the fabrication of parts at the point of use as well as the manufacture of parts with complex geometries that would be difficult to manufacture via conventional methods (milling, casting, etc.). Additively manufactured parts are likely to contain internal defects due to the melt pool, powder material, and laser velocity conditions when printing. Two different types of defects were present in the CT scans of printed AlSi10Mg dogbones: spherical porosity and irregular porosity. Identification of these pores via a machine learning approach (i.e., support vector machines, convolutional neural networks, k-nearest neighbors’ classifiers) could be helpful with part qualification and inspections. The machine learning approach will aim to label the regions of porosity and label the type of porosity present. The results showed that a combination approach of Canny edge detection and a classification-based machine learning model (k-nearest neighbors or support vector machine) outperformed the convolutional neural network in segmenting and labeling different types of porosity.

In the summer of 2020, the National Aeronautics and Space Administration (NASA) launched a spacecraft as part of the Mars 2020 mission. The rover on the spacecraft uses a Multi-Mission Radioisotope Thermoelectric Generator (MMRTG) to provide continuous electrical and thermal power for the mission. The MMRTG uses radioactive plutonium dioxide. NASA prepared a Supplemental Environmental Impact Statement (SEIS) for the mission in accordance with the National Environmental Policy Act. The SEIS provides information related to updates to the potential environmental impacts associated with the Mars 2020 mission as outlined in the Final Environmental Impact Statement (FEIS) for the Mars 2020 Mission issued in 2014 and associated Record of Decision (ROD) issued in January 2015. The Nuclear Risk Assessment (NRA) 2019 Update includes new and updated Mars 2020 mission information since the publication of the 2014 FEIS and the updates to the Launch Approval Process with the issuance of Presidential Memorandum on Launch of Spacecraft Containing Space Nuclear Systems, National Security Presidential Memorandum 20 (NSPM-20). The NRA 2019 Update addresses the responses of the MMRTG to potential accident and abort conditions during the launch opportunity for the Mars 2020 mission and the associated consequences. This information provides the technical basis for the radiological risks discussed in the SEIS. This paper provides a summary of the methods and results used in the NRA 2019 Update.

A survey of cadmium plated field return hardware showed ubiquitous cadmium whisker growth. The most worn and debris-covered hardware showed the densest whisker growth. Whiskers were often found growing in agglomerates of nodules and whiskers. The hardware was rinsed with alcohol to transfer whiskers and debris from the hardware to a flat stub. Fifty whiskers were studied individually by scanning electron microscopy (SEM), including energy dispersive spectroscopy (EDS) and electron backscatter diffraction (EBSD). Most of the whiskers were single crystal, though three were found to contain grain boundaries at kinks. The whiskers ranged from 5 to 600 μm in length and 80 pct showed a <1 ¯ 2 1 ¯ 0> type growth direction. This growth direction facilitates the development of low energy side faces of the whisker, (0001) and {1010}.

A high-speed, two-color pyrometer was developed and employed to characterize the temperature of the ejecta from pyrotechnic igniters. The pyrometer used a single objective lens, beamsplitter, and two high-speed cameras to maximize the spatial and temporal resolutions. The pyrometer used the integrated intensity of under-resolved particles to maintain a large region of interest to capture more particles. The spectral response of the pyrometer was determined based on the response of each optical component and the total system was calibrated using a black body source to ensure accurate intensity ratios over the range of interest.

Periodic arrays of nanoholes perforated in metallic thin films interact strongly with light and produce large electromagnetic near-field enhancements in their vicinity. As a result, the optical response of these systems is very sensitive to changes in their dielectric environment, thus making them an exceptional platform for the development of compact optical sensors. Given that these systems already operate at the shot-noise limit when used as optical sensors, their sensing capabilities can be enhanced beyond this limit by probing them with quantum light, such as squeezed or entangled states. Motivated by this goal, here, we present a comparative theoretical analysis of the quantum enhanced sensing capabilities of metallic nanohole arrays with one and two holes per unit cell. Through a detailed investigation of their optical response, we find that the two-hole array supports resonances that are narrower and stronger than its one-hole counterpart, and therefore have a higher fundamental sensitivity limit as defined by the quantum Cramér-Rao bound. We validate the optical response of the analyzed arrays with experimental measurements of the reflectance of representative samples. The results of this work advance our understanding of the optical response of these systems and pave the way for developing sensing platforms capable of taking full advantage of the resources offered by quantum states of light.

Despite their susceptibility to hydrogen-assisted fracture, ferritic steels make up a large portion of the hydrogen infrastructure. It is impractical and too costly to build large scale components such as pipelines and pressure vessels out of more hydrogen-resistant materials such as austenitic stainless steels. Therefore, it is necessary to understand the fracture behavior of ferritic steels in high-pressure hydrogen environments to manage design margins and reduce costs. Quenched and tempered (Q&T) martensite is the predominant microstructure of high-pressure hydrogen pressure vessels, and higher strength grades of this steel type are more susceptible to hydrogen degradation than lower strength grades. In this study, a single heat of 4340 alloy was heat treated to develop alternative microstructures for evaluation of fracture resistance in hydrogen gas. Fracture tests of several microstructures, such as lower bainite and upper bainite with similar strength to the baseline Q&T martensite, were tested at 21 and 105 MPa H2. Despite a higher MnS inclusion content in the tested 4340 alloy which reduced the fracture toughness in air, the fracture behavior in hydrogen gas fit a similar trend to other previously tested Q&T martensitic steels. The lower bainite microstructure performed similar to the Q&T martensite, whereas the upper bainite microstructure performed slightly worse. In this paper, we extend the range of high-strength microstructures evaluated for hydrogen-assisted fracture beyond conventional Q&T martensitic steels.

Abstract not provided.

A new small-scale pressure vessel with a 5×5 fuel assembly and axially truncated PWR hardware was created to simulate commercial vacuum drying processes. This test assembly, known as the Dashpot Drying Apparatus, was built to focus on the drying of a single PWR dashpot and surrounding fuel. Drying operations were simulated for three tests with the DDA based on the pressure and temperature histories observed in the HBDP. All three tests were conducted with an empty guide tube. One test was performed with deionized water as the fill fluid. The other two tests used 0.2 M boric acid as the fill fluid to accurately simulate spent fuel pool conditions. These tests proved the capability of the DDA to mimic commercial drying processes on a limited scale and detect the presence of bulk and residual water. Furthermore, for all tests, pressure remained below the 0.4 kPa (3 Torr) rebound threshold for the final evacuation step in the drying procedure. Results indicate that after bulk fluid is removed from the pressure vessel, residual water is verifiably measured through confirmatory measurements of pressure and water content using a mass spectrometer. The final pressure rebound behaviors for the three tests conducted were well below the established regulatory limit of less than 0.4 kPa (3 Torr) within 30 minutes of isolation. The water content measurements across all tests showed that despite observing high water content within the DDA vessel at the beginning of the vacuum isolations, the water content drastically drops to below 1,200 ppmv after the isolations were conducted. The data and operational experience from these tests will guide the next evolution of experiments on a prototypic-length scale with multiple surrogate rods in a full 17×17 PWR assembly. The insight gained through these investigations is expected to support the technical basis for the continued safe storage of spent nuclear fuel into long term operations.

Firmware emulation is useful for finding vulnerabil-ities, performing debugging, and testing functionalities. However, the process of enabling firmware to execute in an emulator (i.e., re-hosting) is difficult. Each piece of the firmware may depend on hardware peripherals outside the microcontroller that are inaccessible during emulation. Current practices involve painstakingly disentangling these dependencies or replacing them with developed models that emulate functions interacting with hardware. Unfortunately, both are highly manual and error-prone. In this paper, we introduce a systematic graph-based approach to analyze firmware binaries and determine which functions need to be replaced. Our approach is customizable to balance the fidelity of the emulation and the amount of effort it would take to achieve the emulation by modeling functions. We run our algorithm across a number of firmware binaries and show its ability to capture and remove a large majority of hardware dependencies.

Stochasticity is ubiquitous in the world around us. However, our predominant computing paradigm is deterministic. Random number generation (RNG) can be a computationally inefficient operation in this system especially for larger workloads. Our work leverages the underlying physics of emerging devices to develop probabilistic neural circuits for RNGs from a given distribution. However, codesign for novel circuits and systems that leverage inherent device stochasticity is a hard problem. This is mostly due to the large design space and complexity of doing so. It requires concurrent input from multiple areas in the design stack from algorithms, architectures, circuits, to devices. In this paper, we present examples of optimal circuits developed leveraging AI-enhanced codesign techniques using constraints from emerging devices and algorithms. Our AI-enhanced codesign approach accelerated design and enabled interactions between experts from different areas of the micro-electronics design stack including theory, algorithms, circuits, and devices. We demonstrate optimal probabilistic neural circuits using magnetic tunnel junction and tunnel diode devices that generate an RNG from a given distribution.

This paper describes an efficient reverse-mode differentiation algorithm for contraction operations for arbitrary and unconventional tensor network topologies. The approach leverages the tensor contraction tree of Evenbly and Pfeifer (2014), which provides an instruction set for the contraction sequence of a network. We show that this tree can be efficiently leveraged for differentiation of a full tensor network contraction using a recursive scheme that exploits (1) the bilinear property of contraction and (2) the property that trees have a single path from root to leaves. While differentiation of tensor-tensor contraction is already possible in most automatic differentiation packages, we show that exploiting these two additional properties in the specific context of contraction sequences can improve eficiency. Following a description of the algorithm and computational complexity analysis, we investigate its utility for gradient-based supervised learning for low-rank function recovery and for fitting real-world unstructured datasets. We demonstrate improved performance over alternating least-squares optimization approaches and the capability to handle heterogeneous and arbitrary tensor network formats. When compared to alternating minimization algorithms, we find that the gradient-based approach requires a smaller oversampling ratio (number of samples compared to number model parameters) for recovery. This increased efficiency extends to fitting unstructured data of varying dimensionality and when employing a variety of tensor network formats. Here, we show improved learning using the hierarchical Tucker method over the tensor-train in high-dimensional settings on a number of benchmark problems.

Abstract not provided.

The Multi-Fidelity Toolkit (MFTK) is a simulation tool being developed at Sandia National Laboratories for aerodynamic predictions of compressible flows over a range of physics fidelities and computational speeds. These models include the Reynolds-Averaged Navier–Stokes (RANS) equations, the Euler equations, and modified Newtonian aerodynamics (MNA) equations, and they can be invoked independently or coupled with hierarchical Kriging to interpolate between high-fidelity simulations using lower-fidelity data. However, as with any new simulation capability, verification and validation are necessary to gather credibility evidence. This work describes formal model validation with uncertainty considerations that leverages experimental data from the HIFiRE-1 wind tunnel tests. The geometry is a multi-conic shape that produces complex flow phenomena under hypersonic conditions. A thorough treatment of the validation comparison with prediction error and validation uncertainty is also presented.

A first-of-its-kind model calibration was performed using Sandia National Laboratories' high-throughput initiation (HTI) experiment for two types of vapor-deposited explosive films consisting of hexanitrostilbene (HNS) or pentaerythritol tetranitrate (PETN). These films exhibit prompt initiation, and they reach steady detonation at sub-millimeter length scales. Following prior work on HNS, we test the hypothesis of approximating these explosive films as fine-grained homogeneous solids with simple Arrhenius kinetics burn models. The model calibration process is described herein using a single-step as well as a two-step Arrhenius rate law, and it consists of systematic parameter sampling leading to a reduction in the model degrees of freedom. Multiple local minima are observed; results are given for seven different optimized parameter sets. Each model set is further evaluated in a two-dimensional simulation of the critical failure thickness for a sustained detonation. Overall, the two-step Arrhenius kinetics model captures the observed behavior for HNS; however, neither model produces a good fit to the PETN data. We hypothesize that the HTI results for PETN correspond to a heterogeneous response, owing to the smaller reaction zone of PETN compared to HNS (i.e., it does not homogenize the fine-grained hot spots as well). Future work should consider using the ignition and growth model for PETN, as well as other reactive burn models such as xHVRB, AWSD, PiSURF, and CREST.

Carbon sequestration is a growing field that requires subsurface monitoring for potential leakage of the sequestered fluids through the casing annulus. Sandia National Laboratories (SNL) is developing a smart collar system for downhole fluid monitoring during carbon sequestration. This technology is part of a collaboration between SNL, University of Texas at Austin (UT Austin) (project lead), California Institute of Technology (Caltech), and Research Triangle Institute (RTI) to obtain real-time monitoring of the movement of fluids in the subsurface through direct formation measurements. Caltech and RTI are developing millimeter-scale radio frequency identification (RFID) sensors that can sense carbon dioxide, pH, and methane. These sensors will be impervious to cement, and as such, can be mixed with cement and poured into the casing annulus. The sensors are powered and communicate via standard RFID protocol at 902-928 MHz. SNL is developing a smart collar system that wirelessly gathers RFID sensor data from the sensors embedded in the cement annulus and relays that data to the surface via a wired pipe that utilizes inductive coupling at the collar to transfer data through each segment of pipe. This system cannot transfer a direct current signal to power the smart collar, and therefore, both power and communications will be implemented using alternating current and electromagnetic signals at different frequencies. The complete system will be evaluated at UT Austin's Devine Test Site, which is a highly characterized and hydraulically fractured site. This is the second year of the three-year effort, and a review of SNL's progress on the design and implementation of the smart collar system is provided.

This work presents a high-speed laser-absorption-spectroscopy diagnostic capable of measuring temperature, pressure, and nitric oxide (NO) mole fraction in shock-heated air at a measurement rate of 500 kHz. This diagnostic was demonstrated in the High-Temperature Shock Tube (HST) facility at Sandia National Laboratories. The diagnostic utilizes a quantum-cascade laser to measure the absorbance spectra of two rovibrational transitions near 5.06 µm in the fundamental vibration bands (v" = 0 and 1) of NO in its ground electronic state (X2 Π1/2 ). Gas properties were determined using scanned-wavelength direct absorption and a recently established fitting method that utilizes a modified form of the time-domain molecular free-induction-decay signal (m-FID). This diagnostic was applied to acquire measurements in shock-heated air in the HST at temperatures ranging from approximately 2500 to 5500 K and pressures of 3 to 12 atm behind both incident and reflected shocks. The measurements agree well with the temperature predicted by NASA CEA and the pressure measured simultaneously using PCB pressure sensors. The measurements presented demonstrate that this diagnostic is capable of resolving the formation of NO in shock-heated air and the associated temperature change at the conditions studied.