Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Kernel learning”

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 271 records · Page 15
All sourcesNASA NTRSDOE OSTI
DOE OSTI2021

Physics-Infused AI/ML Based Digital-Twin Framework for Flow-Induced-Vibration Damage Prediction in a Nuclear Reactor Heat Exchanger

This report summarizes some of the ongoing work related to the development of an expert-elicitation-digital-twin framework for real time damage state prediction in heat exchanger components of a nuclear reactor. The framework is targeted towards predicting damage associated with coupled low cycle fatigue (associated with regular heat-up, cool-down and power operation transients) and high cycle fatigue (associated with flow induced vibration transients). The overall framework will be based on a NoSQL based database, physics-infused-geometry-dependent virtual-sensor data, different AI/ML techniques-based data-driven-predictive-model applications (Apps) and real-time plant sensor measurements available through few existing sensors. Towards this overall goal, this report updates some of the ongoing work, such as on implementation of a NoSQL Database (such as MongoDB), FE based heat transfer analysis of a heat exchanger (e.g. of a PWR steam generator) for generating geometry-dependent virtual sensor data and evaluation of various AI/ML models such as based on multivariate linear regression, ensembled decision-tree based Random-Forest and Gradient-Boosting regression and high-dimensional-kernel-function-transformation based Support-Vector-Machine regression models. The AI/ML models were evaluated for predicting multi-time-series thermal states at thousands of 3D point-clouds

22 GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2026

RU-net for automatic characterization of TRISO fuel cross sections

During irradiation, phenomena such as kernel swelling and buffer densification may impact the performance of tristructural isotropic (TRISO) particle fuel. Post-irradiation microscopy is often used to identify these irradiation-induced morphologic changes. However, each fuel compact generally contains thousands of TRISO particles. Manually performing the work to get statistical information on these phenomena is cumbersome and subjective. Here, to reduce the subjectivity inherent in that process and to accelerate data analysis, we used convolutional neural networks (CNNs) to automatically segment cross-sectional images of microscopic TRISO layers. CNNs are a class of machine-learning algorithms specifically designed for processing structured grid data. They have gained popularity in recent years due to their remarkable performance in various computer vision tasks, including image classification, object detection, and image segmentation. In this research, we generated a large irradiated TRISO layer dataset with more than 2,000 microscopic images of cross-sectional TRISO particles and the corresponding annotated images. Based on these annotated images, we used different CNNs to automatically segment different TRISO layers. These CNNs include RU-Net (developed in this study), as well as three existing architectures: U-Net, Residual Network (ResNet), and Attention U-Net. The preliminary results show that the model based on RU-Net performs best in terms of Intersection over Union (IoU). Using CNN models, we can expedite the analysis of TRISO particle cross sections, significantly reducing the manual labor involved and improving the objectivity of the segmentation results.

11 - NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗
DOE OSTI2022

A comprehensive and fair comparison of two neural operators (with practical extensions) based on $\mathrm{FAIR}$ data

Neural operators can learn nonlinear mappings between function spaces and offer a new simulation paradigm for real-time prediction of complex dynamics for realistic diverse applications as well as for system identification in science and engineering. Herein, we investigate the performance of two neural operators, which have shown promising results so far, and we develop new practical extensions that will make them more accurate and robust and importantly more suitable for industrial-complexity applications. The first neural operator, DeepONet, was published in 2019 (Lu et al., 2019), and its original architecture was based on the universal approximation theorem of Chen & Chen (1995). The second one, named Fourier Neural Operator or FNO, was published in 2020, and it is based on parameterizing the integral kernel in the Fourier space. DeepONet is represented by a summation of products of neural networks (NNs), corresponding to the branch NN for the input function and the trunk NN for the output function; both NNs are general architectures, e.g., the branch NN can be replaced with a CNN or a ResNet. According to Kovachki et al. (2021), FNO in its continuous form can be viewed conceptually as a DeepONet with a specific architecture of the branch NN and a trunk NN represented by a trigonometric basis. In order to compare FNO with DeepONet computationally for realistic setups, we develop several extensions of FNO that can deal with complex geometric domains as well as mappings where the input and output function spaces are of different dimensions. We also develop an extended DeepONet with special features that provide inductive bias and accelerate training, and we present a faster implementation of DeepONet with cost comparable to the computational cost of FNO, which is based on the Fast Fourier Transform. Here we consider 16 different benchmarks to demonstrate the relative performance of the two neural operators, including instability wave analysis in hypersonic boundary layers, prediction of the vorticity field of a flapping airfoil, porous media simulations in complex-geometry domains, etc. We follow the guiding principles of FAIR (Findability, Accessibility, Interoperability, and Reusability) for scientific data management and stewardship. The performance of DeepONet and FNO is comparable for relatively simple settings, but for complex geometries the performance of FNO deteriorates greatly. We also compare theoretically the two neural operators and obtain similar error estimates for DeepONet and FNO under the same regularity assumptions.

42 ENGINEERING↗
DOE OSTI2021

Solar Transposition Modeling via Deep Neural Networks With Sky Images

This article presents a notable advance toward the development of a new method of increasing the single-axis tracking photovoltaic (PV) system power output by improving the determination and near-term prediction of the optimum module tilt angle. The tilt angle of the plane receiving the greatest total irradiance changes with Sun position and atmospheric conditions including cloud formation and movement, aerosols, and particulate loading, as well as varying albedo within a module's field of view. In this article, we present a multi-input convolutional neural network that can create a profile of plane-of-array irradiance versus surface tilt angle over a full 180° arc from horizon to horizon. As input, the neural network uses the calculated solar position and clear-sky irradiance values, along with sky images. The target irradiance values are provided by the multiplanar irradiance sensor (MPIS). In order to account for varying irradiance conditions, the MPIS signal is normalized by the theoretical clear-sky global horizontal irradiance. Using this information, the neural network outputs an N -dimensional vector, where N is the number of points to approximate the MPIS curve via Fourier resampling. The output vector of the model is smoothed with a Gaussian kernel to account for error in the downsamping and subsequent upsampling steps, as well as to smooth the unconstrained output of the model. Furthermore, these profiles may be used to perform near-term prediction of angular irradiance, which can then inform the movement of a PV tracker.

42 ENGINEERING↗
DOE OSTI2026

IRMA

IRMA (In)elastic Representation of Materials As S(α,β) evaluations IRMA turns one phonon model into three outputs that usually require three separate tool chains: an evaluated nuclear-data file, predicted neutron-scattering spectra, and scattering kernels for Monte Carlo transport. The three outputs draw on a single, consistent description of the material, so the evaluation, the spectroscopy that can validate it, and the transport that uses it always agree about the physics. Nuclear data. IRMA writes ENDF-6 File 7 thermal scattering evaluations on automatically constructed (α, β) grids. This part reimplements and generalizes NJOY's LEAPR: the classic kernels reproduce freshly generated NJOY2016 tapes digit for digit and published reference tapes to about 1e-4, and the generalized paths add the exact coherent one-phonon term, anisotropic Debye-Waller tensors, coherent elastic for arbitrary crystals, and a per-species partition for polyatomic materials. The tapes feed NJOY, AMPX, FUDGE, and every transport code downstream of them. Neutron spectroscopy. The irma.spectra forward model projects the same physics onto an instrument's kinematics and resolution: INS spectra for VISION and generic indirect geometries, and 2-D S(Q,E) powder maps for direct-geometry spectrometers, from a phonopy model or straight from a phonon DOS. It can be used to predict a proposed measurement before beam time; in analysis, it supplies the calculated single-scattering counterpart of a measured spectrum, from the same material description the evaluation was built from. Monte Carlo transport. The irma.ncrystal exporter writes per-temperature scattering kernels for the companion NCrystal plugin, so McStas, OpenMC, and other NCrystal-aware codes sample the same physics. The exported kernels carry the per-site anisotropic Debye-Waller tensors, keeping directional coherent-elastic physics that NCrystal's standard scalar treatment does not represent. With the same physics inside a transport code, an entire beamline becomes a virtual experiment: IRMA's end-to-end validation ran a custom McStas implementation of the ARCS spectrometer, assembled from the existing McVine and McStas models, against measured data. From a bare crystal structure. The irma mlip front end builds the phonon model itself: a structure file and a choice of potential are enough. Nine pretrained machine-learned interatomic potentials are supported, on a laptop CPU, with no first-principles calculation; an approximate phonon model for a new material costs minutes, not a DFT campaign, and the build emits prefilled inputs for all three outputs. The result is a good starting point rather than a finished evaluation: survey-quality physics with every parameter exposed for review. A converged atomistic calculation enters the same way, as a phonopy model, when higher fidelity is needed.

Ramic, Kemal [Oak Ridge National Laboratory (ORNL)↗
DOE OSTI2024

Machine learning materials properties with accurate predictions, uncertainty estimates, domain guidance, and persistent online accessibility

One compelling vision of the future of materials discovery and design involves the use of machine learning (ML) models to predict materials properties and then rapidly find materials tailored for specific applications. However, realizing this vision requires both providing detailed uncertainty quantification (model prediction errors and domain of applicability) and making models readily usable. At present, it is common practice in the community to assess ML model performance only in terms of prediction accuracy (e.g. mean absolute error), while neglecting detailed uncertainty quantification and robust model accessibility and usability. Here, we demonstrate a practical method for realizing both uncertainty and accessibility features with a large set of models. We develop random forest ML models for 33 materials properties spanning an array of data sources (computational and experimental) and property types (electrical, mechanical, thermodynamic, etc). All models have calibrated ensemble error bars to quantify prediction uncertainty and domain of applicability guidance enabled by kernel-density-estimate-based feature distance measures. All data and models are publicly hosted on the Garden-AI infrastructure, which provides an easy-to-use, persistent interface for model dissemination that permits models to be invoked with only a few lines of Python code. We demonstrate the power of this approach by using our models to conduct a fully ML-based materials discovery exercise to search for new stable, highly active perovskite oxide catalyst materials.

domain of applicability↗
DOE OSTI2021

Implementing a neural network interatomic model with performance portability for emerging exascale architectures

The two main thrusts of computational science are increasingly accurate predictions and faster calculations; to this end, the zeitgeist in molecular dynamics (MD) simulations is pursuing machine learned and data driven interatomic models, e.g. neural network potentials, and novel hardware architectures, e.g. GPUs. Current implementations of neural network potentials are orders of magnitude slower than traditional interatomic models and while looming exascale computing offers the ability to run large, accurate simulations with these models, achieving portable performance for MD with new and varied exascale hardware requires rethinking traditional algorithms, using novel data structures, and library solutions. We re-implement a neural network interatomic model in CabanaMD, an MD proxy application, built on libraries developed for performance portability. Our implementation shows significantly improved thread scaling in this complex kernel as compared to a current LAMMPS implementation, across both strong and weak scaling. Our single-source solution enables simulations up to 20 million atoms on a single CPU node and 4 million atoms with improved performance on a single GPU. Furthermore, we also explore parallelism and data layout choices (using flexible data structures called AoSoAs) and their effect on performance, seeing up to ~50% and ~5% improvements in performance on a GPU by choosing the right level of parallelism and data layout respectively.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Detecting Anomalies in Time Series Using Kernel Density Approaches

This paper introduces a novel anomaly detection approach tailored for time series data with exclusive reliance on normal events during training. Our key innovation lies in the application of kernel-density estimation (KDE) to scrutinize reconstruction errors, providing an empirically derived probability distribution for normal events post-reconstruction. This non-parametric density estimation technique offers a nuanced understanding of anomaly detection, differentiating it from prevalent threshold-based mechanisms in existing methodologies. In post-training, events are encoded, decoded, and evaluated against the estimated density, providing a comprehensive notion of normality. In addition, we propose a data augmentation strategy involving variational autoencoder-generated events and a smoothing step for enhanced model robustness. The significance of our autoencoder-based approach is evident in its capacity to learn normal representation without prior anomaly knowledge. Through the KDE step on reconstruction errors, our method addresses the versatility of anomalies, departing from assumptions tied to larger reconstruction errors for anomalous events. Our proposed likelihood measure then distinguishes normal from anomalous events, providing a concise yet comprehensive anomaly detection solution. The extensive experimental results support the feasibility of our proposed method, yielding significantly improved classification performance by nearly 10% on the UCR benchmark data.

Frehner, Robin↗
DOE OSTI2020

A Data-Driven Nonparametric Approach for Probabilistic Load-Margin Assessment Considering Wind Power Penetration

A modern power system is characterized by an increasing penetration of wind power, which results in large uncertainties in its states. These uncertainties must be quantified properly; otherwise, the system security may be threatened. Facing this challenge, here we propose a cost-effective, data-driven approach to assessing a power system's load margin probabilistically. Using actual wind data, a kernel density estimator is applied to infer the nonparametric wind speed distributions, which are further merged into the framework of a vine copula. The latter enables us to simulate complex multivariate and highly dependent model inputs with a variety of bivariate copulae that precisely represent the tail dependence in the correlated samples. Furthermore, to reduce the prohibitive computational time of traditional Monte-Carlo simulations that process a large amount of samples, we propose to use a nonparametric, Gaussian-process-emulator-based reduced-order model to replace the original complicated continuation power-flow model through a Bayesian-learning framework. To accelerate the convergence rate of this Bayesian algorithm, a truncated polynomial chaos surrogate, which serves as a highly efficient, parametric Bayesian prior, is developed. This emulator allows us to execute the time-consuming continuation power-flow solver at the sampled values with a negligible computational cost. Results of simulations that are performed on several test systems reveal the impressive performance of the proposed method in the probabilistic load-margin assessment.

17 WIND ENERGY↗
DOE OSTI2023

A Multi-Objective Bayesian Optimized Human Assessed Multi-Target Generated Spectral Recommender System for Rapid Pareto Discoveries of Material Properties

Optimization for different tasks like material characterization, synthesis, and functional properties for desired applications over multi-dimensional control parameter and function spaces need a rapid strategic search through active learning. However, in all cases prior to optimization, the target material properties are assumed known and fixed, which mostly deviates from real-world scenarios in material synthesis. This can be critical for running expensive experiments on new materials, when the experimental results are fuzzy for any scientific outcomes due to improper target setting, ultimately wasting time and cost. The failure rate and cost are even higher over exploring on multi-target space, where we want to learn the pareto among multiple properties, to jointly optimize during material synthesis for desired applications. To address the challenge, here we introduce the human-operator attempt flexibility in the active learning based automated experiment framework, with generating multiple human assessed targets through a voting-based recommender system during real-time microscope measurements over the large material image space, sequentially learn/update multiple desired targets through a weighting system, and adaptively search in multiple material properties functional space for non-dominated pareto discoveries to maximize the custom structural similarity based acquisition function. We term this a multi-objective Bayesian optimized human assessed multi-target generated spectral recommender systems (MOBO-HAM-SRS). The approach has been demonstrated to peizoresponse force spectroscopy of a ferroelectric thin film, exploring with different kernels and acquisition functions. This work shows an advancement towards human-AI collaborated automated experiments, steering optimization trajectories through human overpowering AI at the early stage when uncertainty is high and AI overpowering human at the later stage with rapid exploration towards optimal goal, following human-assessed multiple targets properties.

Biswas, Arpan↗
DOE OSTI2022

Predicting the rheology of limestone calcined clay cements (LC 3 ): Linking composition and hydration kinetics to yield stress through Machine Learning

The physicochemical characteristics of calcined clay influence yield stress of limestone calcined clay cements (LC 3 ), but the independent influences the clay's physical and chemical characteristics as well as the effect of other variables on LC 3 rheology are less well-understood. Further, a relationship between LC 3 hydration kinetics and yield stress – important for informing mixture design – has not yet been established. Here, rheological properties were determined in pastes with varying water-to-solid ratio (w/s), constituent mass ratios (PC:metakaolin:limestone), limestone particle size and gypsum content. From these data, an ML model developed allowed the independent examination of the different mechanisms by which metakaolin fraction influences yield stress of LC 3 , identifying four predictors – packing index, Al 2 O 3 /SO 3 , total particle density and metakaolin fraction relative to limestone (MK/LS) – most significant for predicting LC 3 yield stress. A methodology based on kernel smoothing also identified hydration kinetics parameters best correlated with yield stress.

36 MATERIALS SCIENCE↗
DOE OSTI2021

Gaussian process regression constrained by boundary value problems

We develop a framework for Gaussian processes regression constrained by boundary value problems. The framework may be applied to infer the solution of a well-posed boundary value problem with a known second-order differential operator and boundary conditions, but for which only scattered observations of the source term are available. Scattered observations of the solution may also be used in the regression. The framework combines co-kriging with the linear transformation of a Gaussian process together with the use of kernels given by spectral expansions in eigenfunctions of the boundary value problem. Furthermore, it benefits from a reduced-rank property of covariance matrices. We demonstrate that the resulting framework yields more accurate and stable solution inference as compared to physics-informed Gaussian process regression without boundary condition constraints.

42 ENGINEERING↗
DOE OSTI2024

Scalable Risk Assessment of Rare Events in Power Systems With Uncertain Wind Generation and Loads

Risk assessment of rare events has become increasingly important in power system planning and operation with the increasing integration of renewable energy and the presence of system uncertainties. However, quantifying the risk posed by rare events via the traditional method, i.e., Monte Carlo sampling (MCS), incurs substantial computational expense stemming from the vast ensemble of power flow simulations. To accelerate the assessment, this paper proposes a Deep Neural Network (DNN)-kernelized vector-valued Gaussian Process (VVGP) approach with excellent computational efficiency while maintaining high accuracy. Consequently, serving as a surrogate model for the power flow solver, the DNN-kernelized VVGP enables significantly faster but accurate risk assessment compared to the power flow solver. The developed surrogate model evaluates low-order N - k events that contain more than 90% instances by adeptly capturing the topological features while the high-order N - k events are assessed via a power flow solver, thereby striking a balance between computational efficiency and uncertainty quantification accuracy. Moreover, the model incorporates a Support Vector Machine (SVM) classifier to resample concerning low-probability tail events to counteract the biases potentially introduced during the DNN-kernelized VVGP evaluations. Simulations conducted on the modified IEEE 24-bus, 118-bus, and European 1354-bus systems demonstrate that the proposed method maintains the accuracy benchmark set by MCS while significantly reducing computational demands in large-scale power systems as compared to other state-of-the-art methods.

17 WIND ENERGY↗
NASA NTRS2009

Simulating and Detecting Radiation-Induced Errors for Onboard Machine Learning

Spacecraft processors and memory are subjected to high radiation doses and therefore employ radiation-hardened components. However, these components are orders of magnitude more expensive than typical desktop components, and they lag years behind in terms of speed and size. We have integrated algorithm-based fault tolerance (ABFT) methods into onboard data analysis algorithms to detect radiation-induced errors, which ultimately may permit the use of spacecraft memory that need not be fully hardened, reducing cost and increasing capability at the same time. We have also developed a lightweight software radiation simulator, BITFLIPS, that permits evaluation of error detection strategies in a controlled fashion, including the specification of the radiation rate and selective exposure of individual data structures. Using BITFLIPS, we evaluated our error detection methods when using a support vector machine to analyze data collected by the Mars Odyssey spacecraft. We found ABFT error detection for matrix multiplication is very successful, while error detection for Gaussian kernel computation still has room for improvement.

data analysis↗
DOE OSTI2021

Cholesky-based experimental design for Gaussian process and kernel-based emulation and calibration.

Gaussian processes and other kernel-based methods are used extensively to construct approximations of multivariate data sets. The accuracy of these approximations is dependent on the data used. This paper presents a computationally efficient algorithm to greedily select training samples that minimize the weighted L p error of kernel-based approximations for a given number of data. The method successively generates nested samples, with the goal of minimizing the error in high probability regions of densities specified by users. The algorithm presented is extremely simple and can be implemented using existing pivoted Cholesky factorization methods. Training samples are generated in batches which allows training data to be evaluated (labeled) in parallel. For smooth kernels, the algorithm performs comparably with the greedy integrated variance design but has significantly lower complexity. Numerical experiments demonstrate the efficacy of the approach for bounded, unbounded, multi-modal and non-tensor product densities. We also show how to use the proposed algorithm to efficiently generate surrogates for inferring unknown model parameters from data using Bayesian inference.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

RAP: Resource-aware Automated GPU Sharing for Multi-GPU Recommendation Model Training and Input Preprocessing

Ensuring high-quality recommendations for newly onboarded users requires the continuous retraining of Deep Learning Recommendation Models (DLRMs) with freshly generated data. To serve the online DLRM retraining, existing solutions use hundreds of CPU computing nodes designated for input preprocessing, causing significant power consumption that surpasses even the power usage of GPU trainers. To this end, we propose RAP, an end-to-end DLRM training framework that supports Resource-aware Automated GPU sharing for DLRM input Preprocessing and Training. The core idea of RAP is to accurately capture the remaining GPU computing resources during DLRM training for input preprocessing, achieving superior training efficiency without requiring additional resources. Specifically, RAP utilizes a co-running cost model to efficiently assess the costs of various input preprocessing operations, and it implements a resource-aware horizontal fusion technique that adaptively merges smaller kernels according to GPU availability, circumventing any interference with DLRM training. In addition, RAP leverages a heuristic searching algorithm that jointly optimizes both the input preprocessing graph mapping and the co-running schedule to maximize the end-to-end DLRM training throughput. The comprehensive evaluation shows that RAP achieves 78.3× speedup on average over CPU-based DLRM input preprocessing frameworks. In addition, the end-to-end training throughput of RAP is only 2.04% lower than the ideal case, which has no input preprocessing overhead.

Wang, Zheng↗
DOE OSTI2025

Many-body expansion based machine learning models for octahedral transition metal complexes

Abstract Graph-based machine learning (ML) models for material properties show great potential to accelerate virtual high-throughput screening of large chemical spaces. However, in their simplest forms, graph-based models do not include any 3D information and are unable to distinguish stereoisomers such as those arising from different orderings of ligands around a metal center in coordination complexes. In this work we present a modification to revised autocorrelation descriptors, a molecular graph featurization method, for predicting spin state dependent properties of octahedral transition metal complexes (TMCs). Inspired by analytical semi-empirical models for TMCs, the new modeling strategy is based on the many-body expansion (MBE) and allows one to tune the captured stereoisomer information by changing the truncation order of the MBE. We present the necessary modifications to include this approach in two commonly used ML methods, kernel ridge regression and feed-forward neural networks. On a test set composed of all possible isomers of binary TMCs, the best MBE models achieve mean absolute errors (MAEs) of 2.75 kcal mol −1 on spin-splitting energies and 0.26 eV on frontier orbital energy gaps, a 30%–40% reduction in error compared to models based on our previous approach. We also observe improved generalization to previously unseen ligands where the best-performing models exhibit MAEs of 4.00 kcal mol −1 (i.e. a 0.73 kcal mol −1 reduction) on the spin-splitting energies and 0.53 eV (i.e. a 0.10 eV reduction) on the frontier orbital energy gaps. Because the new approach incorporates insights from electronic structure theory, such as ligand additivity relationships, these models exhibit systematic generalization from homoleptic to heteroleptic complexes, allowing for efficient screening of TMC search spaces.

Meyer, Ralf (ORCID:0000000322360261)↗
DOE OSTI2023

Software-Hardware Co-design of Heterogeneous SmartNIC System for Recommendation Models Inference and Training

Deep Learning Recommendation Models (DLRMs) are critical applications in various domains and have evolved as one of the single largest machine learning applications. Trillions of DLRM parameters exceed the on-chip memory capacity of GPUs. Large-scale multi-node systems are required for distributed DLRM inference and training, which suffer from the all-to-all communication bottleneck, mainly limiting the scalability of ever-growing DLRMs. In recent years, SmartNICs have evolved with coupled computation and communication capabilities providing opportunities for a powerful heterogeneous device in the system. However, there isn't such a distributed system that fully leverages the abundant smartNIC resources that resolve the scalability issue of DLRMs. In this work, we proposed a software-hardware co-design of a heterogeneous smartNIC system that resolves the communication bottleneck of distributed DLRMs, mitigates the memory bandwidth pressure, and improves computation efficiency. We provide a set of smartNIC designs of cache systems (including local cache and remote cache) and smartNIC computation kernels which reduce data movement, relieve memory lookup intensity, and improve the GPU's computation efficiency. In addition, we propose a graph algorithm that improves the data locality of queries within batches which optimizes the overall system performance with higher data reuse. Our evaluation shows that our system achieves 2.1x latency speedup for inference and 1.6x throughput speedup for training.

Guo, Anqi↗