Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Edge computing”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,405 records · Page 78

Alfalfa

Alfalfa-based testbeds enable building equipment, control products, and workforce development tools to interact with dynamic building simulations representing the desired building, system, weather, and grid configuration. Alfalfa is used to de-risk implementation of load flexibility prior to field deployment, reducing the costs and timelines associated with adoption of decarbonization technology at the grid edge.

building energy modeling↗

INFRARED NARROWBAND TOMOGRAPHY OF THE LOCAL STARBURST NGC 1569 WITH THE LARGE BINOCULAR TELESCOPE/LUCIFER

We used the near-IR imager/spectrograph LUCIFER mounted on the Large Binocular Telescope to image, with subarcsecond seeing, the local dwarf starburst NGC 1569 in the JHK bands and He I 1.08 {mu}m, [Fe II] 1.64 {mu}m, and Br{gamma} narrowband filters. We obtained high-quality spatial maps of He I 1.08 {mu}m, [Fe II] 1.64 {mu}m, and Br{gamma} emission across the galaxy, and used them together with Hubble Space Telescope/Advanced Camera for Surveys images of NGC 1569 in the H{alpha} filter to derive the two-dimensional spatial map of the dust extinction and surface star formation rate (SFR) density. We show that dust extinction (as derived from the H{alpha}/Br{gamma} flux ratio) is rather patchy and, on average, higher in the northwest (NW) portion of the galaxy (E{sub g}(B - V) {approx_equal} 0.71 mag) than in the southeast (E{sub g}(B - V) {approx_equal} 0.57 mag). Similarly, the surface density of SFR (computed from either the dereddened H{alpha} or dereddened Br{gamma} image) peaks in the NW region of NGC 1569, reaching a value of about 4 x 10{sup -6} M{sub sun} yr{sup -1} pc{sup -2}. The total SFR as estimated from the integrated, dereddened H{alpha} (or, alternatively, Br{gamma}) luminosity is about 0.4 M{sub sun} yr{sup -1}, and the total supernova rate from the integrated, dereddened [Fe II] 1.64 {mu}m luminosity is about 0.005 yr{sup -1} (assuming a distance of 3.36 Mpc). The azimuthally averaged [Fe II] 1.64 {mu}m/Br{gamma} flux ratio is larger at the edges of the central, gas-deficient cavities (encompassing the superstar clusters A and B) and in the galaxy outskirts. If we interpret this line ratio as the ratio between the average past star formation (as traced by supernovae) and ongoing activity (represented by OB stars able to ionize the interstellar medium), it would then indicate that star formation has been quenched within the central cavities and lately triggered in a ring around them. The number of ionizing hydrogen and helium photons as computed from the integrated, dereddened H{alpha} and He I 1.08 {mu}m luminosities suggests that the latest burst of star formation occurred about 4 Myr ago and produced new stars with a total mass of {approx_equal}1.8 x 10{sup 6} M{sub sun}.

79 ASTRONOMY AND ASTROPHYSICS↗

A tantalum strength model using a multiscale approach: version 2

A continuum strength model for tantalum was developed in 2007 using a multiscale approach. This was our first attempt at connecting simulation results from atomistic to continuum length scales, and much was learned that we were not able to incorporate into the model at that time. The tantalum model described in this report represents a second cut at pulling together multiscale simulation results into a continuum model. Insight gained in creating previous multiscale models for tantalum and vanadium was used to guide the model construction and functional relations for the present model. While the basic approach follows that of the vanadium model, there are significant departures. Some of the recommendations from the vanadium report were followed, but not all. Results from several new analysis techniques have not yet been incorporated due to technical difficulties. Molecular dynamics simulations of single dislocation motion at several temperatures suggested that the thermal activation barrier was temperature dependent. This dependency required additional temperature functions be included within the assumed Arrhenius relation. The combination of temperature dependent functions created a complex model with a non unique parameterization and extra model constants. The added complexity had no tangible benefits. The recommendation was to abandon the strict Arrhenius form and create a simpler curve fit to the molecular dynamics data for shear stress versus dislocation velocity. Functions relating dislocation velocity and applied shear stress were constructed vor vanadium for both edge and screw dislocations. However, an attempt to formulate a robust continuum constitutive model for vanadium using both dislocation populations was unsuccessful; the level of coupling achieved was inadequate to constrain the dislocation evolution properly. Since the behavior of BCC materials is typically assumed to be dominated by screw dislocations, the constitutive relations were ultimately built using only the screw relations. In light of this, the screw dislocation mobility relation is chosen as the starting point for the present tantalum model. Edge dislocations are not included explicitly in the current model. A significant change from the previous models is in the functional dependence of the dislocation evolution equations. The prior multiscale models assumed that the dislocation evolution rate depended on stress as well dislocation velocity. This crated an implicit dependence on the kinetic relation and required an iterative solution for the dislocation density. In the present model the integration scheme is simplified by casting the dislocation evolution terms of strain rate and current dislocation density. The final notable change was in the transition relation from the thermally activated regime to phonon drag. Historically, and in the prior models, this transition had been through a harmonic average on the strain rates making it impossible to determine the stress directly when given the plastic strain rate. After considering alternative transition relations, it was determined that an equally suitable fit to the molecular dynamics simulation data in the transition region could be constructed by averaging stresses rather than strain rates. The end result of these enhancements is a simpler (fewer parameters) and more straight forward model formulation where the material strength can be evaluated directly when given the plastic strain rate, temperature, pressure and the dislocation density at the beginning of the time step. This greatly improves the computational efficiency and robustness over the prior models where additional iteration loops were required.

36 MATERIALS SCIENCE↗

Atomic-scale Imaging of PGM-free Catalyst Active Sites by 30 keV 4D-STEM

Platinum group metal (PGM)-free catalysts have attracted a large amount of attention due to their potential for enabling low-cost, commercially viable hydrogen fuel cells and electrolyzers [1]. Degradation of cell performance remains a significant challenge, however, currently limiting implementation of devices that utilize PGM-free materials [2]. Controlling degradation in these materials requires a better understanding of the atomic structure of active sites, thought to be FeN4 structures within a graphitic carbon lattice, which would enable more accurate prediction of the associated degradation pathways. The exact structural arrangement of active sites in these materials is still debated since a range of structures have been computationally predicted and methods to directly validate these models are still needed [1,3-5].While scanning transmission electron microscopy (STEM) has provided initial glimpses into the nature of the proposed active sites, detailed evaluation of the structure of these sites is challenging since they involve defects, edges, and nitrogen dopants in the graphitic lattice, which are susceptible to beam damage [6,7]. Conventional STEM imaging and spectroscopy methods exacerbate this problem with dose-inefficiency or nonideal contrast characteristics for imaging light and heavy elements simultaneously. A method that minimizes damage while generating dose-efficient, easily-interpretable contrast for both light and heavy elements is therefore needed to facilitate imaging of sensitive active site atomic structure.Here, we demonstrate direct atomic-scale imaging of PGM-free catalyst active sites by low-voltage four-dimensional (4D)-STEM. To accomplish this, we pair a 30 keV aberration-corrected probe, which minimizes knock-on damage, with a fast pixelated detector that has optimal performance at low beam energies [8]. We show how this setup enables relatively simple center-of-mass (CoM) techniques [9,10] to produce images with an increased signal-to-noise ratio (SNR) and light-element contrast over conventional imaging modes, allowing the entire active site structure to be imaged at the atomic scale. Moreover, we show how electron ptychography [11,12] enables images with further improved characteristics to be obtained, for example by minimizing residual aberrations, which provides a more accurate structural representation. The increased understanding that these low-voltage 4D-STEM techniques will provide about the atomic structure of PGM-free active sites and their associated degradation pathways will facilitate rational design of next-generation materials, promoting development of low-cost hydrogen fuel cells, electrolyzers, and other energy conversion devices [13]. References:[1] U Martinez et al., Adv. Mater. 31 (2019), p. 1806545.[2] Y Shao et al., Adv. Mater. 31 (2019), p. 1807615.[3] J Kneebone et al., J. Phys. Chem. C 121, (2017), p.16283.[4] T Mineva et al., ACS Catal. 9 (2019), p. 9359.[5] J Li et al., Nat. Catal. 4 (2021), p. 10.[6] T Susi et al., ACS Nano 6 (2012), p. 8837.[7] P Zelenay and DJ Myers, US Department of Energy Hydrogen and Fuel Cells Program 2017 Annual Merit Review and Peer Evaluation Meeting, Washington, DC (2017).[8] H Ryll et al., J. Inst. 11 (2016), p. P04006.[9] K Müller et al., Nat. Commun. 5 (2014), p. 5653.[10] I Lazic et al., Ultramicroscopy 160 (2016), p. 265.[11] H Yang et al., Nat. Commun. 7 (2016), p. 12532.[12] Y Jiang et al., Nature 559 (2018), p. 343. [13] Research sponsored by the Hydrogen and Fuel Cell Technologies Office, Office of Energy Efficiency and Renewable Energy, US Department of Energy (DOE). Research was conducted at the Center for Nanophase Materials Sciences, which is a DOE Office of Science User Facility.

Zachman, Michael↗

Validation of a synthetic fast ion loss detector model for Wendelstein 7-X

Abstract We present the first validated synthetic diagnostic for fast ion loss detectors (FILDs) in the Wendelstein 7-X (W7-X) stellarator. This model has been developed on, and validated against experimental data from, a FILD provided by the National Institute for Fusion Science (NIFS-FILD), with potential future applicability to the existing Faraday Cup FILD (FC-FILD) on W7-X as well as the scintillating FILD (S-FILD) currently under development. A workflow combining Monte Carlo codes BEAMS3D and ASCOT5 is used to track fast ions produced by neutral beam injection from the moment of ionization until they are thermalized or lost from the last closed flux surface, and from there to a virtual plane which serves as a projection of the entrance aperture to the FILD. Simulations in ASCOT5 are analyzed via a geometric method to determine the probability of transmission through the FILD aperture and onto the detector as a function of normalized momentum, pitch angle, gyrophase, and position at the virtual plane. This probability is then applied to the simulated ions arriving from the plasma, producing a simulated signal from a computationally tractable number of simulated fast ions. Simulated signals are presented for two W7-X experiments with neutral beam injection and quantitatively compared with experimental measurements from the NIFS-FILD diagnostic. An estimate of the frequency of charge-exchange with neutral particles in the edge is performed, and it is found that this process may have a significant impact on the measured signals.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗

Contributions of the stochastic shape wake model to predictions of aerodynamic loads and power under single wake conditions

The contribution of wake meandering and shape asymmetry to load and power estimates is quantified by comparing aeroelastic simulations initialized with different inflow conditions: an axisymmetric base wake, an unsteady stochastic shape wake, and a large-eddy simulation with rotating actuator-line turbine representation. Time series of blade-root and tower base bending moments are analyzed. We find that meandering has a large contribution to the fluctuation of the loads. Moreover, considering the wake edge intermittence via the stochastic shape model improves the simulation of load and power fluctuations and of the fatigue damage equivalent loads. Furthermore, these results indicate that the stochastic shape wake simulator is a valuable addition to simplified wake models when seeking to obtain higher-fidelity computationally inexpensive predictions of loads and power.

17 WIND ENERGY↗

Solutes that reduce yield strength anisotropies in magnesium from first principles

Using Labusch-type solid solution strengthening models parameterized with DFT-computed solute-dislocation interaction energies, we perform a computational search for 63 solutes across the periodic table to find those that lower anisotropy ratios (non-basal to basal CRSS) of magnesium potentially increasing its ductility per the von Mises criterion. For this purpose, we compute changes in strength for solutes as a function of composition and temperature, and compute anisotropy ratios for solutes that include both rare earth and non-rare earth elements. Here we specifically focus on solute-dislocation interaction energies in the following DFT-optimized dislocations as representative of three non-basal plastic deformation modes: $\langle c + a \rangle$ edge, (10$\bar1$2) tension twinning edge, and the (10$\bar{1}$1) compression twinning edge. We find that solute-induced changes in non-basal deformation modes can be approximated using a second-order polynomial in the size misfit of the solutes, which permits rapid screening of solutes. Our approach to identify solutes known to improve strengthening incorporates solute solubility, and suggests other solutes that not have been previously explored for strengthening. The 8 rare-earth solutes that our method suggests as the best, ordered by increasing anisotropy ratios at their optimal concentrations, are: Gd, Tb, Dy, Nd, Ho, Er, Tm, and Yb. The 12 non-rare-earth solutes that our method suggests as the best, ordered by increasing anisotropy ratios, are: Y, Mn, Sc, Pb, Ca, Ag, Bi, Tl, Zn, Li, Ga, and Al. Of these, Gd, Nd, Er, Yb, Y, Mn, Ca, Zn, Li, and Al are used in commercial Mg alloys.

36 MATERIALS SCIENCE↗

Dotted-line FLEET for two-component velocimetry

Femtosecond laser electronic excitation tagging (FLEET) is a powerful unseeded velocimetry technique typically used to measure one component of velocity along a line, or two or three components from a dot. In this Letter, we demonstrate a dotted-line FLEET technique which combines the dense profile capability of a line with the ability to perform two-component velocimetry with a single camera on a dot. Our set-up uses a single beam path to create multiple simultaneous spots, more than previously achieved in other FLEET spot configurations. We perform dotted-line FLEET measurements downstream of a highly turbulent, supersonic nitrogen free jet. Dotted-line FLEET is created by focusing light transmitted by a periodic mask with rectangular slits of 1.6 × 40 mm 2 and an edge-to-edge spacing of 0.5 mm, then focusing the imaged light at the measurement region. Up to seven symmetric dots spaced approximately 0.9 mm apart, with mean full-width at half maximum diameters between 150 and 350 µm, are simultaneously imaged. Both streamwise and radial velocities are computed and presented in this Letter.

47 OTHER INSTRUMENTATION↗

Porosity in Millimeter-Scale Welds of Stainless Steel: Three-Dimensional Characterization

A variety of edge joints utilizing a continuous wave Nd:YAG laser have been produced and examined in a 304-L stainless steel to advance fundamental understanding of the linkage between processing and resultant microstructure in high-rate solidification events. Acquisition of three-dimensional reconstructions via micro-computed tomography combined with traditional metallography has allowed for qualitative and quantitative characterization of weld joints in a material system of wide use and broad applicability. The presence, variability and distribution of porosity, has been examined for average values, spatial distributions and morphology and then related back to fundamental processing parameters such as weld speed, weld power and laser focal length.

36 MATERIALS SCIENCE↗

Understanding the physics of a possible non-Abelian fractional quantum hall effect state.

We wish to present in this report experimental results from a one-year Senior Council Tier-1 LDRD project that focused on understanding the physics of a possible non-Abelian fractional quantum Hall effect state. We first give a general introduction to the quantum Hall effect, and then present the experimental results on the edge-state transport in a special fractional quantum Hall effect state at Landau level filling {nu} = 5/2 - a possible non-Abelian quantum Hall state. This state has been at the center of current basic research due to its potential applications in fault-resistant topological quantum computation. We will also describe the semiconductor 'Hall-bar' devices we used in this project. Electron physics in low dimensional systems has been one of the most exciting fields in condensed matter physics for many years. This is especially true of quantum Hall effect (QHE) physics, which has seen its intellectual wealth applied in and has influenced many seemingly unrelated fields, such as the black hole physics, where a fractional QHE-like phase has been identified. Two Nobel prizes have been awarded for discoveries of quantum Hall effects: in 1985 to von Klitzing for the discovery of integer QHE, and in 1998 to Tsui, Stormer, and Laughlin for the discovery of fractional QHE. Today, QH physics remains one of the most vibrant research fields, and many unexpected novel quantum states continue to be discovered and to surprise us, such as utilizing an exotic, non-Abelian FQHE state at {nu} = 5/2 for fault resistant topological computation. Below we give a briefly introduction of the quantum Hall physics.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Band-selective shaped pulse for high fidelity quantum control in diamond

High fidelity quantum control of qubits is crucially important for realistic quantum computing, and it becomes more challenging when there are inevitable interactions between qubits. We introduce a band-selective shaped pulse, refocusing BURP (REBURP) pulse, to cope with the problems. The electron spin of nitrogen-vacancy centers in diamond is flipped with high fidelity by the REBURP pulse. In contrast with traditional rectangular pulses, the shaped pulse has almost equal excitation effect in a sharply edged region (in frequency domain). So the three sublevels of host {sup 14}N nuclear spin can be flipped accurately simultaneously, while unwanted excitations of other sublevels (e.g., of a nearby {sup 13}C nuclear spin) is well suppressed. Our scheme can be used for various applications such as quantum metrology, quantum sensing, and quantum information process.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗

Computing Maximum Cardinality Matchings in Parallel on Bipartite Graphs via Tree-Grafting

It is difficult to obtain high performance when computing matchings on parallel processors because matching algorithms explicitly or implicitly search for paths in the graph, and when these paths become long, there is little concurrency. In spite of this limitation, we present a new algorithm and its shared-memory parallelization that achieves good performance and scalability in computing maximum cardinality matchings in bipartite graphs. This algorithm searches for augmenting paths via specialized breadth-first searches (BFS) from multiple source vertices, hence creating more parallelism than single source algorithms. Algorithms that employ multiple-source searches cannot discard a search tree once no augmenting path is discovered from the tree, unlike algorithms that rely on single-source searches. We describe a novel tree-grafting method that eliminates most of the redundant edge traversals resulting from this property of multiple-source searches. We also employ the recent direction-optimizing BFS algorithm as a subroutine to discover augmenting paths faster. Our algorithm compares favorably with the current best algorithms in terms of the number of edges traversed, the average augmenting path length, and the number of iterations. Here, we provide a proof of correctness for our algorithm. Our NUMA-aware implementation is scalable to 80 threads of an Intel multiprocessor and to 240 threads on an Intel Knights Corner coprocessor. On average, our parallel algorithm runs an order of magnitude faster than the fastest algorithms available. The performance improvement is more significant on graphs with small matching number.

97 MATHEMATICS AND COMPUTING↗

Development of a Novel Accelerator for Neutron Transport Solution Using the Galerkin Spectral Element Methods (Final Report)

This report summarizes the results of a three-year research project sponsored by the U.S. Department of Energy (DOE) Nuclear Energy University Program (NEUP) to develop and implement advanced acceleration schemes for the DOE NEAMS neutronics code PROTEUSSN. The project team included the University of Massachusetts Lowell, The Ohio State University, University of Michigan, and Argonne National Laboratory. The PROTEUS code is a high-fidelity capable deterministic neutron transport code based on unstructured finite element meshes, which solves the steady-state and transient neutron transport problem using the 2nd-order discrete ordinate method (SN2ND), the method of characteristics (MOC), and the advanced nodal transport method (NODAL). The DSA scheme has been implemented in PROTEUS to speed up SN transport calculations. The existing DSA scheme employs a consistent finite element formulation on the same fine mesh structure as the SN solution While the use of consistent discretizations makes the DSA method effective, it does not necessarily make it efficient because the numerical solution of a discretized elliptic diffusion problem itself can be costly, particularly when the problem size becomes large. We developed a new discontinuous Galerkin (DG) discretization of the diffusion equation, called DG-DSA, which can effectively and efficiently accelerate the SN transport iterations. As compared with the previous work, the novelty of our method is that the diffusion equation is solved on a coarse-mesh grid using the DG methods, and the DG diffusion discretization incorporates local hp adaptation, i.e., local adaptation of mesh size and/or polynomial degree, based on local total cross section (or optical thickness). Therefore, the resulting number of degrees of freedom (DOF) of the DG discretization is much less than the conventional consistent DSA discretizations, and thus DG-DSA can achieve significant improvement in computational efficiency. We implemented this scheme in PROTEUS-SN. In addition, we developed a new stabilization scheme named linear prolongation CMFD (lpCMFD). A novel feature of this scheme is that the conventional flat flux ratio–based scaling approach is replaced with a linear interpolation of the scalar flux differences at the coarse-mesh cell edges between the neutron transport and CMFD calculations. Fourier convergence analysis and numerical results show that lpCMFD is unconditionally stable and more effective than other CMFD based on acceleration schemes such as pCMFD, odCMFD. We also developed a new nonlinear diffusion acceleration scheme for solving neutron transport equations. This scheme, called LR-NDA, employs a local refinement approach on the framework of CMFD by solving a local boundary value problem of the scalar flux on the coarse-mesh structure to replace the piecewise constant scalar flux obtained by CMFD. The refined flux is then used to update the scalar flux in the neutron transport source iteration. It has been demonstrated that that LR-NDA is much more effective and stable than CMFD for a wide range of optical thicknesses. LR-NDA is a local adaptive method, which means LR-NDA does not necessarily require local refinement for all the coarse-mesh cells on the problem domain, i.e., it can be used only for relatively optically thick regions where the standard CMFD scheme would encounter the convergence problem.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗

Quantifying Barotrauma Risk to Juvenile Fish during Hydro-turbine Passage

We introduce a method for hydro turbine biological performance assessment (BioPA) to bridge the gap between field and laboratory studies on fish injury and turbine engineering design. Using this method, a suite of biological performance indicators is computed based on simulated data from a computational fluid dynamics (CFD) model of a proposed hydro turbine design. Each performance indicator is a measure of the probability of exposure to a certain dose of an injury mechanism. If the relationship between the dose of an injury mechanism (stressor) and frequency of injury (dose-response) is known from laboratory or field studies, the likelihood of fish injury for a turbine design can be computed from the performance indicator. By comparing the values of the indicators from various turbine designs, engineers and biologists can identify the more-promising designs and operating conditions to minimize hydraulic conditions hazardous to passing fish. In this paper, the BioPA method is applied to estimate barotrauma induced mortal injury rates for Chinook salmon exposed to rapid pressure changes in Kaplan-type hydro turbines. Following the description of the general method, application of the BioPA to estimate the probability of mortal injury from exposure to rapid decompression is illustrated using a Kaplan hydro turbine at the John Day Dam on the Columbia River in the Pacific Northwest region of the USA. The estimated rates of mortal injury increased from 0.3% to 1.7% as discharge through the turbine increased from 334 to 564 m3/s for fish assumed to be acclimated to a depth of 5 m. The majority of pressure nadirs occurred immediately below the runner blades, with the lowest values in the gap at the blade tips and just below the leading edge of the blades. Such information can help engineers focus on problem areas when designing new turbine runners to be more fish-friendly than existing units.

computational fluid dynamics (CFD)↗

On eigenvalues of a P T-symmetric operator in a thin layer

We consider an elliptic operator with variable coefficients in a thin three-dimensional layer with P T-symmetric boundary conditions. We study the effect of the appearance of isolated eigenvalues at the edges of the gaps in the essential spectrum. We obtain sufficient conditions that guarantee that such eigenvalues either exist or are absent near a given edge of a gap. In the case of existence, the first terms in the asymptotic expansion of these emerging eigenvalues are calculated. Bibliography: 34 titles. (paper)

97 MATHEMATICS AND COMPUTING↗

Gap solitons and Bloch waves in nonlinear periodic systems

We comprehensively investigate gap solitons and Bloch waves in one-dimensional nonlinear periodic systems. Our results show that there exists a composition relation between them: Bloch waves at either the center or edge of the Brillouin zone are infinite chains composed of fundamental gap solitons (FGSs). We argue that such a relation is related to the exact relation between nonlinear Bloch waves and nonlinear Wannier functions. With this composition relation, many conclusions can be drawn for gap solitons without any computation. For example, for the defocusing nonlinearity, there are n families of FGS in the nth linear Bloch band gap; for the focusing case, there are infinite number of families of FGSs in the semi-infinite gap and other gaps. In addition, the stability of gap solitons is analyzed. In literature, there are numerical results showing that some FGSs have cutoffs on propagation constant (or chemical potential), i.e., these FGSs do not exist for all values of propagation constant (or chemical potential) in the linear band gap. We develop an intuitive picture to describe this cutoff.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Machine Learning Based Software Package for Reproducible Data Analysis for Electron Microscopy

Advances in hardware and computation technologies are enabling large hyperspectral imaging during Electron Microscopy (EM), resulting in datasets being produced at rapidly increasing spatial and temporal resolutions. This presents exciting opportunities for microstructure analysis, but also poses great challenges in data analysis. Under this STTR program, QuesTek Innovations LLC, a leader in the field of integrated computational materials engineering (ICME), collaborated with Lawrence Livermore National Laboratory and Argonne National Laboratory. The team developed an open-source machine learning (ML) powered tool for EM data analysis of multiple systems and data types, and collected EM data for ML model development using the cutting-edge PicoProbe equipment at Argonne National Laboratory.

Lin, Yu↗

FPGA Acceleration of GCN in Light of the Symmetry of Graph Adjacency Matrix

Graph Convolutional Neural Networks (GCNs) are widely used to process large-scale graph data. Different from deep neural networks (DNNs), GCNs are sparse, irregular, and unstructured, posing unique challenges to hardware acceleration with regular processing elements (PEs). In particular, the adjacency matrix of a GCN is extremely sparse, leading to frequent but irregular memory access, low spatial/temporal data locality and poor data reuse. Furthermore, a realistic graph usually consists of unstructured data (e.g., unbalanced distributions), creating significantly different processing times and imbalanced workload for each node in GCN acceleration. To overcome these challenges, we propose an end-to-end hardware-software co-design to accelerate GCNs on resource-constrained FPGAs with the features including: (1) A custom dataflow that leverages symmetry along the diagonal of the adjacency matrix to accelerate feature aggregation for undirected graphs. We utilize either the upper or the lower triangular matrix of the adjacency matrix to perform aggregation in GCN to improve data reuse. (2) Unified compute cores for both aggregation and transform phases, with full support to the symmetry-based dataflow. These cores can be dynamically reconfigured to the systolic mode for transformation or as individual accumulators for aggregation in GCN processing. (3) Preprocessing of the graph in software to rearrange the edges and features to match the custom dataflow. This step improves the regularity in memory access and data reuse in the aggregation phase. Moreover, we quantize the GCN precision from FP32 to INT8 to reduce the memory footprint without losing the inference accuracy. We implement our accelerator design in Intel Stratix10 MX FPGA board with HBM2, and demonstrate 1.3x-110.5x improvement in end-to-end GCN latency as compared to the state-of the-art FPGA implementations, on the graph datasets of Cora, Pubmed, Citeseer and Reddit.

Nair, Gopikrishnan R.↗