Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “tensor networks”

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 379 records · Page 21
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Machine-Learning-Based Multiscale Methods for 3D Modelling of Granular Materials by Incorporating History-Dependent State Variables

Over the past decades, the prevalence of machine learning (ML) methods has made the development of ML-based constitutive models for granular materials undoubtedly a popular subject. Numerous studies have been made to feature the loading path or history-dependent stress-strain response of granular media using neural networks. In this work, a novel finite element method (FEM)–ML multiscale approach was developed by incorporating internal variables to improve the simulation accuracy of 3D history-dependent granular materials for the first time. To this end, a surrogate constitutive model based on the single-step-based multi-layer perceptron (MLP) neural network was used to replace representative volume element (RVE) simulations conducted by the discrete element method (DEM) in the multiscale FEM–DEM approach. Although the prediction principle of the MLP aligns with the FEM algorithm, artificially added internal variables are required to differentiate the loading history. To address this issue, history variables associated with the Frobenius norm are proposed to be fed into the MLP coupled with the strain tensor to extract the history-dependent behaviour of granular assemblies. The developed FEM–ML approach was demonstrated in 3D conventional triaxial compression (CTC) simulations. Compared to the multiscale FEM–DEM approach, the proposed FEM–ML method exhibits a significantly improved computational efficiency.

granular materials↗
DOE OSTI2021

Regional Moment Tensor Inversion Using Rotational Observations

There are benefits from the addition of rotational motions to translational displacements for moment tensor (MT) inversions. The rotational radiation pattern is orthogonal to the shear radiation pattern, thus incorporating rotations is equivalent to gaining another observation point on the focal sphere. We demonstrated this by simulating curl and displacement wavefields for a regional distance station. Thus, one 6-C station (3-Component translational + 3-Component rotational) gathers the same information on radiation pattern as two 3-C stations at 90° azimuth from one another along the focal plane axis, which is sometimes difficult to obtain when restricted to surface sensors. We added rotational Green's functions to a regional MT inversion scheme (long-period, time-domain, and linear inversion) by computing spatial gradients from f-κ reflectivity synthetics. For rotational data, we used Array Derived Rotations (ADR) from Piñon Flats Observatory Array and Golay array deployed during IRIS Community Wavefield Demonstration Experiment. The hope is to ultimately use compact and field deployable broadband rotational seismometers instead of ADRs. Rotational motions were predicted from well-constrained deviatoric MT solutions of nine earthquakes recorded by the arrays and other seismic networks. Additionally, we formed three station sparse datasets with (3-C and 6-C) and without rotational ground motions (3-C only) for eight earthquakes testing the benefits of including rotational waveforms in MT inversions when they are equally weighted with translational waveforms. Adding rotational motions improved the double-couple components and reduced compensated-linear-vector-dipole and isotropic components in full-MT solutions using sparse datasets with poor coverage.

58 GEOSCIENCES↗
DOE OSTI2021

Machine learning elastic constants of multi-component alloys

The present manuscript explores application of machine learning methods for determining elastic constants and other derived mechanical properties of multi-component alloys. Here, a number of machine learning models, including linear regression, neural network and random forest based models, are trained and tested on a dataset of binary alloys generated using density functional theory (DFT) calculations and spanning over a large number of elemental species in the periodic table. Starting with a wide range of simple and easily accessible compositionally-averaged elemental features, a correlation-based feature selection strategy was used to systematically down-select a set of most relevant features towards the prediction of the elasticity tensor components. The true predictive performance and the associated uncertainties of the models were established by testing on unseen data and bootstrapping, respectively. A single and pair-wise feature partial dependence analysis was performed to visualize the average property trends in the multi-dimensional feature space in order to further understand the achieved predictive performance. The utility of the trained model is further demonstrated by obtaining sufficiently accurate yet highly efficient approximations for bulk modulus, Young’s modulus, shear modulus and Poisson’s ratio for alloys beyond the binary space (i.e., two-component alloys) on which the model was originally trained. More importantly, we test and validate the predictive performance of the developed model directly against the experimentally measured elastic constants of technologically relevant multi-component alloys (such as, Ni- and Ti-based alloys). Finally, utility of such a data-enabled route is demonstrated by predicting the possible range of various elastic properties for vast composition space available within the five component Ni-Cr-Fe-Mo-W alloy system in a high-throughput manner.

36 MATERIALS SCIENCE↗
DOE OSTI2026

Lunar accelerometer network gravitational observatory (LANGO)

With ground-based interferometers detecting hundreds of gravitational-wave (GW) events, GW astronomy has continued to blossom. U.S. and European scientists are developing plans to construct third-generation ground-based interferometers. ESA has proceeded through the mission formulation phase of space-based interferometer in a lower-frequency band, 10 –4 –0.1 Hz. Despite all these exciting developments, there is still a missing frequency band, 0.1–10 Hz. This mid-frequency band is rich with interesting astrophysical events. Coalescence and merger of intermediate-mass black holes (IMBHs) will occur in this frequency band. Coalescing stellar-mass BHs will pass through this frequency band days before they reach the frequency band of LIGO and Virgo. Detection of such signals would enable a mid-frequency detector to issue an advance notice to the high-frequency GW detectors, as well as to optical, x-ray and γ-ray telescopes. We propose Lunar Accelerometer Network Gravitational Observatory (LANGO) to detect GWs in this frequency band. In the first phase (LANGO 1), we propose to deploy four ambient-temperature (250 K) accelerometers in the tetrahedral or in a square configuration on the hemisphere facing the Earth. After successful operation at 250 K, LANGO would be upgraded to a cryogenic (4 K) version with over two orders of magnitude increased sensitivity (LANGO 2), with coherent rejection of seismic noise implemented. LANGO is a full-tensor detector, capable of determining the source direction and wave polarization. Each test mass (TM) is suspended as a pendulum with resonance frequency ∼ 0.01 Hz, thus is only weakly coupled to the lunar surface horizontally. LANGO is designed to detect the relative motion of globally separated, nearly free, TMs by using the Moon as a large quiet platform. LANGO 1 and 2 accelerometers aim at sensitivities ⩽ 10 –11 m s –2 Hz –1/2 and ⩽ 10 –13 m s –2 Hz –1/2 in the horizontal axes over the frequency band of 1 mHz–10 Hz, which yield GW sensitivities 1.1 x 10 -21 Hz -1/2 and 3.8 x 10 -24 Hz -1/2 at 1 Hz, respectively. The LANGO accelerometers will be 10 3 –10 5 times more sensitive than Apollo seismometers. With such sensitivity, LANGO will also make great contribution to the advancement of lunar geophysics.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2023

Plasma production and ion heating systems for the Material Plasma Exposure eXperiment

Plasma production and ion heating in the Material Plasma Exposure eXperiment (MPEX), whose design is nearing completion, is accomplished using continuous wave RF power with average power density up to 1.6 MW/m2 at the plasma interface. Plasma is produced using helicon waves coupled through a single helical antenna at 13.56 MHz with power supplied by three 100 kW fixed-frequency RF generators feeding a power combiner network followed by the matching network and launcher. The helicon source can utilize various gasses including hydrogen, deuterium, and helium, with a magnetic field strength in the source region up to 0.2 T, and maximum device |B| of 2.5 T. Power is coupled to ions via ion cyclotron heating at the fundamental resonance using a pair of phased helical antennas operating in the frequency range 4-9 MHz, that launch waves towards the resonance from the antenna region where ω > ωci. ICH power is supplied by a single 500 kW tunable RF transmitter through a 90° power splitter and matching/decoupling network. In the case of both the helicon and ICH systems the antennas are located external to the vacuum, with power transferred through novel water-cooled coaxial vacuum windows consisting of fused quartz outer cylinders and silicon nitride inner cylinders with forced convection water cooling between them. The antenna enclosures are pressurized with dry air to 3 bar absolute for voltage standoff.Several 3-D COMSOL models have been created to simulate the two systems. A model of the helicon region utilizing a cold plasma dielectric tensor with accurate magnetic field and realistic plasma density profiles has been used to calculate the plasma loading/complex antenna input impedance at the launcher feed for various ne and |B| values, for the purpose of estimating power handling. It also incorporates the geometry of all launcher structures relevant to this determination. A still more detailed model of the launcher together with a lossy dielectric plasma surrogate has been used to determine RF electric field values and power losses in the device components.Similar models have been produced to predict the performance and power handling of the ICH launcher. For this device the impedance matrix of the two-element antenna array is calculated using a warm plasma model, necessary to properly determine the wave propagation and absorption.

Goulding, Richard↗
DOE OSTI2020

Research Introduction [Slides]

In the topic of Reverse Time Imaging, we proposed a new IC to reduce computation cost but reserve image resolution for distributed sensor networks. For Induced Seismicity in Oklahoma, we analyzed fault stress state analysis at state scale, and we applied machine learning techniques to polarity picking and seismicity rate forecasting. The results provide better understanding of fault properties, stress field, and the relationships among fault, stress state, injections, and potential seismic hazards. Lastly, for Microseismic Monitoring, we detected and located 770 low-frequency events (5-50 Hz): (1) Shallow events are highly clustered, consistent with the pathway from injection well 13-10A to monitoring well; moment tensor inversion shows dominant tensile cracking; (2) Deep events are scattered and show migration pattern to the basement; moment tensors show that most events are shear cracks.

58 GEOSCIENCES↗
DOE OSTI2022

Precursory Off-Fault Deformation in Restraining and Releasing Step Overs: Insights From Discrete Element Method Models

Accelerating geophysical activity is detected preceding some, but not all, large earthquakes. This observation may indicate that no precursors occur before some earthquakes, or that the instrumentation lacks the required sensitivity. Here, to aid crustal monitoring efforts, we use discrete element method models to identify the locations and styles of deformation that may provide useful information about approaching fault reactivation. We model the reactivation of two healed rough faults in a variety of step over configurations, embedded in a host rock with varying amounts of damage subject to shear velocity loading parallel to the faults. Both the fault geometry and ratio of fault to host rock strength control the amount of off-fault deformation. Consistent with field observations, models with larger steps and more preexisting host rock damage produce higher amounts of off-fault deformation. We assess the size of the continuous regions of high velocities and strains to compare the value of the precursory information of each velocity and strain component. Comparing the three components of the velocity vector suggests that the fault-parallel velocity produces the largest and most temporally continuous regions of elevated velocity. The size of these regions increases toward failure, indicating the usefulness of tracking this component. Comparing the volumetric and shear components of the three-dimensional strain tensor suggests that during most of the interseismic period, the shear strain provides more information about approaching fault slip than the volumetric strain. However, in the days and months preceding fault reactivation, both the shear and volumetric strains provide similarly valuable information.

58 GEOSCIENCES↗
DOE OSTI2024

Bridging the Gap Between LLMs and LNS with Dynamic Data Format and Architecture Codesign

Deep neural networks (DNNs) have achieved tremendous success in the past few years. However, their training and inference demand exceptional computational and memory resources. Quantization has been shown as an effective approach to mitigate the cost, with the mainstream data types reduced from FP32 to FP16/BF16 and recently FP8 in the latest NVIDIA H100 GPUs. With increasingly aggressive quantization, however, the conventional floating-point formats suffer from limited precision in representing numbers around zero. Recently, NVIDIA demonstrated the potential of using a Logarithmic Number System (LNS) for the next generation of tensor cores. While LNS mitigates the hurdles in representing small numbers, in this work we observed a mismatch between LNS and the emerging Large Language Models (LLM), where LLM exhibits significant outliers when directly adopting the LNS format. In this paper, we present a data-format/architecture codesign to bright this gap. On the format side, we propose a dynamic LNS format to flexibly represent outliers at a higher precision, by exploiting asymmetry in the LNS representation and identifying outliers through a per-vector basis. On the architecture side, for demonstration, we realize the dynamic LNS format in a systolic array, which can handle the irregularity of the outliers at runtime. We implement our approach on an Alveo U280 FPGA as a prototype. Experimental results show that our design can effectively handle the outliers and resolve the mismatch between LNS and LLM, contributing to an accuracy improvement of 15.4% and 16% over the floating-point and the original LNS baselines, using four state-of-the-art LLM models. Our observation and design lay a solid foundation for the large-scale adoption of the LNS format in the next-generation deep learning hardware.

Haghi, Pouya↗
DOE OSTI2024

Hutchinson Trace Estimation for high-dimensional and high-order Physics-Informed Neural Networks

Physics-Informed Neural Networks (PINNs) have proven effective in solving partial differential equations (PDEs), especially when some data are available by seamlessly blending data and physics. However, extending PINNs to high-dimensional and even high-order PDEs encounters significant challenges due to the computational cost associated with automatic differentiation in the residual loss function calculation. Herein, we address the limitations of PINNs in handling high-dimensional and high-order PDEs by introducing the Hutchinson Trace Estimation (HTE) method. Starting with the second-order high-dimensional PDEs, which are ubiquitous in scientific computing, HTE is applied to transform the calculation of the entire Hessian matrix into a Hessian vector product (HVP). This approach not only alleviates the computational bottleneck via Taylor-mode automatic differentiation but also significantly reduces memory consumption from the Hessian matrix to an HVP’s scalar output. We further showcase HTE’s convergence to the original PINN loss and its unbiased behavior under specific conditions. Comparisons with the Stochastic Dimension Gradient Descent (SDGD) highlight the distinct advantages of HTE, particularly in scenarios with significant variability and variance among dimensions. We further extend the application of HTE to higher-order and higher-dimensional PDEs, specifically addressing the biharmonic equation. By employing tensor-vector products (TVP), HTE efficiently computes the colossal tensor associated with the fourth-order high-dimensional biharmonic equation, saving memory and enabling rapid computation. The effectiveness of HTE is illustrated through experimental setups, demonstrating comparable convergence rates with SDGD under memory and speed constraints. Additionally, HTE proves valuable in accelerating the Gradient-Enhanced PINN (gPINN) version as well as the Biharmonic equation. Overall, HTE opens up a new capability in scientific machine learning for tackling high-order and high-dimensional PDEs.

Curse of dimensionality↗
NASA NTRSDate not supplied

Towards Surrogate Modeling of Subgrid Turbulent Transport for 3D Radiative Hydrodynamic Simulations of the Quiet Sun

In this work, we investigate the use of deep learn-ing techniques as surrogate models, to enhance the estimationof effects of subgrid turbulent transport for 3D radiatuve hy-drodynamic simulations of the quiet Sun. We develop two dis-tinct 3D Convolutional Neural Networks (3DCNNs) to capturespatio-temporal dependencies in 3D velocity fields, leveragingdifferent activation functions and architectural designs. Thesemodels integrate both averaged velocity vector components andscalar features such as plasma density to enhance predictionaccuracy. Additionally, a Multilayer Perceptron (MLP) modelis employed to approximate complex nonlinear relationships,offering a comparison in performance between convolutionaland fully connected architectures. Logarithmic transformationis applied to the targets to handle heavily skewed data, im-proving model performance. All models are compared againsta physics-based Gradient Model. Results show that the 3DCNNmodels excel at approximating Reynolds stress tensors, makingthem a candidate for assisting in producing reduced resolutionsimulations, and thereby reducing computational overheadwhile maintaining higher accuracy than the baseline. Thesefindings demonstrate the potential of deep learning, particu-larly CNNs, to advance scalable and accurate simulations ofsolar dynamics, offering a promising alternative to traditionalturbulence models.

Heliophysics↗
DOE OSTI2020

Accelerating Scientific Computing in the Post-Moore’s Era

Novel uses of graphical processing units for accelerated computation revolutionized the field of high-performance scientific computing by providing specialized workflows tailored to algorithmic requirements. As the era of Moore’s law draws to a close, many new non–von Neumann processors are emerging as potential computational accelerators, including those based on the principles of neuromorphic computing, tensor algebra, and quantum information. While development of these new processors is continuing to mature, the potential impact on accelerated computing is anticipated to be profound. We discuss how different processing models can advance computing in key scientific paradigms: machine learning and constraint satisfaction. Significantly, each of these new processor types utilizes a fundamentally different model of computation, and this raises questions about how to best use such processors in the design and implementation of applications. While many processors are being developed with a specific domain target, the ubiquity of spin-glass models and neural networks provides an avenue for multi-functional applications. Furthermore, this also hints at the infrastructure needed to integrate next-generation processing units into future high-performance computing systems.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Transversity PDFs of the proton from lattice QCD with physical quark masses

We present a lattice QCD calculation of the transversity isovector- and isoscalar-quark parton distribution functions (PDFs) of the proton utilizing a perturbative matching at next-to-leading-order (NLO) accuracy. Additionally, we determine the isovector and isoscalar tensor charges for the proton. In both calculations, the disconnected contributions to the isoscalar matrix elements have been ignored. The calculations are performed using a single ensemble of N f = 2 + 1 highly improved staggered quarks simulated with physical-mass quarks and a lattice spacing of a = 0.076 fm . The Wilson-clover action, with physical quark masses and smeared gauge links obtained from one iteration of hypercubic smearing, is used in the valence sector. Using the NLO operator product expansion, we extract the lowest four to six Mellin moments and the PDFs via a neural network from the matrix elements in the pseudo-PDF approach. In addition, we calculate the PDFs in the quasi-PDF approach with hybrid-scheme renormalization and the recently developed leading-renormalon resummation technique, at NLO with the resummation of leading small- x logarithms. Published by the American Physical Society 2024

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗
DOE OSTI2021

Sneutrino tribrid inflation, metastable cosmic strings and gravitational waves

Here, we present a successful realization of sneutrino tribrid inflation model based on a gauged U(1)_B-L extension of Minimal Supersymmetric Standard Model (MSSM). A single interaction term involving the B-L Higgs field and the right-handed neutrinos serves multiple purposes. These include the generation of heavy Majorana masses for the right-handed neutrinos to provide an explanation for the tiny neutrino masses via the seesaw mechanism, a realistic scenario for reheating and non-thermal leptogenesis with a reheat temperature as low as 10 6 GeV, and a successful realization of inflation with right-handed sneutrino as the inflaton. The matter parity which helps avoid rapid proton decay survives as a Z 2 subgroup of a U(1) R-symmetry. Depending on the choice of model parameters yields the following predicted range of the tensor to scalar ratio, 3 x 10 -11 ≲ r≲ 7x 10 -4 (6 x 10 -7 ≲ r ≲ 0.01), and the running of the scalar spectral index, -0.00022 ≲ dn_s/dln k ≲ -0.0026 (-0.00014 ≲ dn_s/dln k ≲ 0.005), along with the B-L breaking scale, 3 x 10 14 ≲ M/ GeV≲ 5 x 10 15 (6 x 10 15 ≲ M/ GeV≲ 2 x 10 16 ), calculated at the central value of the scalar spectral index, n_s =0.966, reported by Planck 2018. The possibility of realizing metastable cosmic strings in a grand unified theory setup is briefly discussed. The metastable cosmic string network admits string tension values in the range 10 -8 ≲ Gμ s ≲ 10 -6 , and predicts a stochastic gravitational wave background lying within the 2-σ bounds of the recent NANOGrav 12.5-yr data.

supersymmetry and cosmology↗
DOE OSTI2022

Exploring the Use of Novel Spatial Accelerators in Scientific Applications

Driven by the need to find alternative accelerators which can viably replace GPUs in next-generation Supercomputing systems, this paper proposes a methodology to enable agile application/hardware co-design. The application-first methodology provides the ability to come up with design of accelerators while working with real-world workloads, available accelerators, and system software. The iterative design process targets a set of kernels in a workload for performance estimates that can prune the design space for later phases of detailed architectural evaluations. To this effect, in this paper, a novel data-parallel device model is introduced that simulates the latency of performance-sensitive operations in an accelerator including data transfers and kernel computation using multi-core CPUs. The use of off-the-shelf simulators, such as pre-RTL simulator Aladdin or multiple tools available for exploring the design of deep neural network accelerators (e.g., Timeloop) is demonstrated for evaluation of various accelerator designs using applications with realistic inputs. Examples of multiple device configurations that are instantiable in a system are explored to evaluate the performance benefit of deploying novel accelerators. The proposed device is integrated with a programming model and system software to potentially explore the impacts of high-level programming languages/compilers and low-level effects such as task scheduling on multiple accelerators. We analyze our methodology for a set of applications that represent high-performance computing (HPC) and graph analytics. The applications include a computational chemistry kernel realized using tensor contractions, triangle counting, GraphSAGE and Breadth-first Search. These applications include kernels such as dense matrix-dense matrix multiplication, sparse matrix-spare matrix multiplication, and sparse matrix-dense vector multiplication. Our results indicate potential performance benefits and insights for system design by including accelerators that realize these kernels along-side general purpose accelerators.

AI, codesign, Accelerated Computing, Modeling and ↗
DOE OSTI2025

Design Choices in Anomaly Detection for Industrial Control Systems: Insights from Gas Pipeline Data

Industrial control systems (ICS) remain vulnerable to increasingly sophisticated cyberattacks, yet evaluating anomaly detection models in these environments is challenging due to temporal dependencies, missing-not-at-random patterns, and extremely imbalanced datasets. These factors make common practices—especially random data splits and naïve imputation—prone to severe temporal leakage, which can inflate reported performance and obscure real-world limitations. In this work, we systematically examine classical machine learning models, temporal deep learning architecture, and tensor-decomposition–based methods on a gas-pipeline dataset using a fully temporally separated evaluation pipeline designed to mimic realistic deployment conditions. Our findings show that proper temporal handling and MNAR-aware preprocessing significantly alter the relative performance of popular anomaly-detection methods, providing practical guidance for designing reliable, leakage-resistant ICS intrusion-detection systems.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Ocelot: An Interactive, Efficient Distributed Compression-As-a-Service Platform With Optimized Data Compression Techniques

Large volumes of data generated by scientific simulations, genome sequencing, and other applications need to be moved among clusters for data collection/analysis. Data compression techniques have effectively reduced data storage and transfer costs. However, users' requirements on interactively controlling both data quality and compression ratios are non-trivial to fulfill. Here, we propose a novel Compression-as-a-Service (CaaS) platform called Ocelot with four important contributions: (1) It offers real-time visualization, interactive compression, and transfer of scientific datasets. (2) It incorporates new strategies for compressing diverse types of datasets more effectively than traditional methods. (3) It provides an effective method for estimating the compression ratio and execution time of compression tasks. (4) Experiments on multiple real-world datasets on geographically distributed computers show that Ocelot can significantly improve data transfer efficiency with a performance gain of more than 10x in computing clusters with relatively slow networks.

compression as a service (CaaS)↗
DOE OSTI2024

Direct numerical simulation of open-channel flow over a heterogeneous particle bed at low relative submergence

In this article, this study investigates turbulent open-channel flows over beds of irregularly arranged particles, using direct numerical simulations at a friction Reynolds number of Re T = 300⁠. Two distinct cases are examined: a polydisperse bed (P800) composed of multiple layers of randomly distributed spheres of varying sizes, and a monodisperse bed (M1015) formed by a random distribution of uniform sized spheres, with a bottommost single layer of varied-sized particles to introduce realistic randomness. Our investigation unveils a rich network of low- and high-speed streaks within the flow field, exhibiting distinctive behaviors in different bed configurations. The P800 case presents a poorly organized flow pattern induced by the varied particle sizes and arrangements, while the M1015 case shows a more regular flow pattern, marked by larger streaks. We also observe that total wall shear stress is substantially influenced by surface roughness-induced drag, extending beyond the effects documented in existing studies of open-channel flows. The present study reveals intricate secondary flow patterns over irregular particle beds. Large-scale circulations are discerned around particle crests in the P800 case and localized circulations with increased turbulence in the M1015 case. Furthermore, analysis of Reynolds stress tensor components indicates that roughness disrupts coherent turbulent eddies, consequently mitigating peak stress. We quantify correlations between drag force and local fluid velocity fluctuations. Notably, a larger deviation in drag is observed in the P800 case compared to M1015, accentuating the influence of particle size and distribution on fluid–particle interactions.

42 ENGINEERING↗
DOE OSTI2024

Distributed-Memory Sparse Deep Neural Network Inference Using Global Arrays

Partitioned Global Address Space (PGAS) models exhibit tremendous promise in developing efficient and productive distributed-memory parallel applications. They have been used extensively in scientific computations due to conveniently offering a ``shared-memory''-like model and convenient interfaces that separate communication with synchronization. Traditionally, PGAS communication models have been applied to dense/contiguously distributed data, but most modern applications depict varied levels of sparsity. Existing PGAS models require certain adaptations to support distributed sparse computations, since associated computations often require matrix arithmetic, in addition to data movement. The Global Arrays toolkit from Pacific Northwest National Laboratory (PNNL) is one of the earliest PGAS models to combine one-sided data communication and distributed matrix operations and is still used in the popular NWChem quantum chemistry suite. Recently, we have expanded the Global Arrays toolkit to support common sparse operations, like sparse matrix-dense matrix multiplies (SpMM), sparse matrix-sparse matrix multiplication (SpGEMM) and Sampled Dense-Dense Matrix Multiplication (SDDMM). As it turns out, these operations are the bedrock of sparse Deep Learning (DL); sparse deep neural networks and Graph Neural Networks (GNNs) have gained increasing attention recently in achieving speedups on training and inference with reduced memory footprints. Unlike scientific applications in High Performance Computing (HPC), modern (distributed-memory capable) DL toolkits often rely on non-standardized and closed-source vendor software optimizations, creating challenges in software-hardware co-design at scale. Our goal is to support a variety of distributed-memory sparse matrix operations and helper functions in the newly created Sparse Global Arrays (SGA), such that it is possible to build portable and productive Machine Learning scenarios for algorithm/software and hardware codesign purposes. Contemporary data-parallel schemes for training/inference are undergoing a major overhaul since model replication limits scalability and causes resource inefficiencies. As such, we have adopted tensor parallelism in decomposing the model and inputs, to mitigate memory issues. Current implementation is built on top of MPI and uses CPUs to maximize the portability across the platforms.

Distributed computing, machine learning↗