Search NASA⌕ Search

SEARCH · Search NASA

Results for “Neural Operator 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 613 records · Page 34
All sourcesNASA NTRSDOE OSTI
NASA NTRS2006

Case Study: Test Results of a Tool and Method for In-Flight, Adaptive Control System Verification on a NASA F-15 Flight Research Aircraft

Adaptive control technologies that incorporate learning algorithms have been proposed to enable autonomous flight control and to maintain vehicle performance in the face of unknown, changing, or poorly defined operating environments [1-2]. At the present time, however, it is unknown how adaptive algorithms can be routinely verified, validated, and certified for use in safety-critical applications. Rigorous methods for adaptive software verification end validation must be developed to ensure that. the control software functions as required and is highly safe and reliable. A large gap appears to exist between the point at which control system designers feel the verification process is complete, and when FAA certification officials agree it is complete. Certification of adaptive flight control software verification is complicated by the use of learning algorithms (e.g., neural networks) and degrees of system non-determinism. Of course, analytical efforts must be made in the verification process to place guarantees on learning algorithm stability, rate of convergence, and convergence accuracy. However, to satisfy FAA certification requirements, it must be demonstrated that the adaptive flight control system is also able to fail and still allow the aircraft to be flown safely or to land, while at the same time providing a means of crew notification of the (impending) failure. It was for this purpose that the NASA Ames Confidence Tool was developed [3]. This paper presents the Confidence Tool as a means of providing in-flight software assurance monitoring of an adaptive flight control system. The paper will present the data obtained from flight testing the tool on a specially modified F-15 aircraft designed to simulate loss of flight control faces.

Jacklin, Stephen A.↗
DOE OSTI2020

Deep Learning for Automated Identification of Eels in Sonar Data

Freshwater eels, such as the American eel (Anguilla rostrata) present numerous challenges related to safe downstream fish passage at hydroelectric facilities. One of those challenges is effective monitoring of their abundance, movements, and behavior to facilitate design and operation of eel protection and passage facilities. A previous EPRI study documented the ability of human analysts to reliably identify American eels in data obtained with a 1100/1800 kHz, multibeam sonar. This report describes a project to develop deep learning (a subset of artificial intelligence) tools to automate the time-consuming, subjective process of eel identification in multibeam sonar data. The project exploited new data collected in the laboratory and the existing data from the prior EPRI field study to develop and test deep learning and other data analytic tools, including wavelet filtering, differencing for static object removal, and convolutional neural network analysis. The analysis of the laboratory data demonstrated feasibility of the approach, revealed object characteristics observed with the sonar that distinguish eels from similarly sized and shaped acoustic targets, and provided additional data for algorithm selection and training. Deep learning algorithms trained and tested on the laboratory data alone achieved accuracy rates of greater than 98% when classifying acoustic images of eels and similar-sized neutrally buoyant sticks. The algorithm trained and tested on the pre-existing field data alone, and yielded classification accuracy of 9.3% false positives and 13.3% false negatives when distinguishing between eels and sticks/PVC pipes based on video clips (i.e., multiple, consecutive images). This performance is comparable to the classification accuracy achieved by human analysts in the prior study. The deep learning algorithm trained on a combination of video clips obtained in the laboratory and the field and tested on video clips from the field, was able to distinguish eels from sticks and PVC pipes (a river debris analog) of similar size with 100% accuracy. Outreach to the hardware, software, and end-user communities early in the project helped to identify needs and specify the application space. Outreach to those communities at the end of the project communicated project results and opportunities for further development. The project achieved proof of concept for automated identification of eel in multibeam sonar data. Future work should focus on acquisition of additional data for more robust algorithm training and testing; modification of the software tools to accommodate multiple acoustic targets in the acoustic field at a given time; identification of additional object classes; incorporation of motion in the object identification and classification algorithms; operationalizing the software tools, including integration with other existing sonar data analysis tools; and partnering with hardware and software providers for distribution of the software tools with their commercial products.

13 HYDRO ENERGY↗
DOE OSTI2020

An Exploratory Approach Using Regression and Machine Learning in the Analysis of Mass Absorption Cross Section of Black Carbon Aerosols: Model Development and Evaluation

Mass absorption cross-section of black carbon (MAC BC ) describes the absorptive cross-section per unit mass of black carbon, and is, thus, an essential parameter to estimate the radiative forcing of black carbon. Many studies have sought to estimate MAC BC from a theoretical perspective, but these studies require the knowledge of a set of aerosol properties, which are difficult and/or labor-intensive to measure. We therefore investigate the ability of seven data analytical approaches (including different multivariate regressions, support vector machine, and neural networks) in predicting MAC BC for both ambient and biomass burning measurements. Our model utilizes multi-wavelength light absorption and scattering as well as the aerosol size distributions as input variables to predict MAC BC across different wavelengths. We assessed the applicability of the proposed approaches in estimating MAC BC using different statistical metrics (such as coefficient of determination (R 2 ), mean square error (MSE), fractional error, and fractional bias). Overall, the approaches used in this study can estimate MAC BC appropriately, but the prediction performance varies across approaches and atmospheric environments. Based on an uncertainty evaluation of our models and the empirical and theoretical approaches to predict MAC BC , we preliminarily put forth support vector machine (SVM) as a recommended data analytical technique for use. We provide an operational tool built with the approaches presented in this paper to facilitate this procedure for future users.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2021

Short-term nodal load forecasting based on machine learning techniques

This paper introduces an advanced Short-term Nodal Load Forecasting (STNLF) method that forecasts nodal load profiles for the next day in power systems, based on the combined use of three machine learning techniques. Least Absolute Shrinkage and Selection Operator (LASSO) is employed to reduce the number of features for a single nodal load forecasting. Principal Component Analysis (PCA) is used to capture the features of historical loads in low-dimensional space compared to the original high-dimensional load space where features are barely possible to depict. Additionally, Bayesian Ridge Regression (BRR) is utilized to decide the parameters of the prediction model from a statistics perspective. Tests based on modified PJM load data demonstrate the effectiveness of the proposed STNLF method compared to the state-of-the-art General Regression Neural Network (GRNN) method. Moreover, the reliability of the day-ahead Unit Commitment (UC) solution is shown to have been improved, based on the forecasted load data using the proposed STNLF method.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2022

Convolutional neural networks for intra-hour solar forecasting based on sky image sequences

Accurate and timely solar forecasts play an increasingly critical role in power systems. Compared to longer forecasting timescales, very short-term solar forecasting has lagged behind in both research and practice. In this paper, we propose deep convolutional neural networks (CNNs) to provide operational intra-hour (10-minute-ahead to 60-minute-ahead) solar forecasts. We develop two CNN structures inspired by a widely-used CNN architecture. The CNNs are tailored to our solar forecasting regression tasks and rely solely on sky image sequences. Case studies based on six years of data (over 150,000 data points) demonstrate that the best CNN model has forecast skill scores of 20%-39% over the naive persistence of cloudiness benchmark, even at these very short timescales. The CNNs also have consistently superior performance when compared to shallow machine learning models with meteorological predictors, where the improvement averages around 7%. The sensitivity analyses show that the sky image length, resolution, and weather conditions have impacts on the deep learning model accuracy. In our intra-hour problem with specific setups, two sky images with a 10-minute 128 x 128 resolution yield the most accurate forecasts. Current limitations, future work, and deployment challenges and solutions are also discussed.

14 SOLAR ENERGY↗
DOE OSTI2022

Deep Neural Network Algorithm for CMC Microstructure Characterization and Variability Quantification

Microstructure characterization and variability quantification are crucial for understanding ceramic matrix composites (CMCs) mechanical behavior and deformation mechanisms across length scales. Traditionally, analyses of the micrographs obtained from microscopy are labor-intensive. However, with the vast improvement in computer vision (CV) and deep learning (DL), an automated algorithm can be designed to extract essential microstructure variability from micrographs which can then be used to construct a statistically representative volume element (SRVE). The DL-based algorithm spans the taxonomy of microstructure analyses, including semantic segmentation of microstructure constituents, secondary phases, matrix/fiber interface, and defects, and quantifying the microstructure variability in terms of probability distributions. In this work, C/SiNC and SiC/SiNC CMCs microstructures are semantically segmented through a deep convolutional neural network, followed by variability quantification through the implementation of a fully connected regression layer, hence forming a deep regression network. The deep regression network operates in a feedforward regime, in which the neuron output signal traverses through the network in a unidirectional manner. The weight tensor associated with each layer is updated through a backpropagation stochastic gradient descent approach. The input gray-scale image obtained through in-house scanning electron microscope and confocal microscope micrographs is augmented through affine transformations to increase the training set size, which is then processed through four strided convolutional layers. This compresses the image resolution by half at each layer while increasing the image depth by applying different filters (image encoding). The class activation maps (CAMs) corresponding to the applied filters highlight the key architectural features and assist with the semantic segmentation of the microstructure.

Hamza, Mohamed H.↗
DOE OSTI2025

Even Higher-Level Synthesis: An Exploration of AI Hardware Accelerators using HLS4ML

With the rise of artificial intelligence, the popularization of deep learning, and a constantly evolving industry, the demand for flexible and efficient tools has never been greater. As algorithms grow more complex, their runtime and energy consumption increase exponentially. Customized hardware accelerators, long used for specific mathematical operations, remain essential for managing modern applications' computational and power demands. Hardware accelerators can speed up complex computations by orders of magnitude, but their manual design and verification processes are often challenging and time-consuming. High-Level Synthesis (HLS) provides a solution by transforming high-level algorithm descriptions, typically written in C++ or SystemC, into synthesizable RTL suitable for hardware implementation. This approach reduces development time for RTL engineers while offering flexibility beyond what traditional handwritten RTL can provide. We extended this capability to the machine-learning domain with the open-source framework hls4ml, which allows neural networks trained in Python frameworks like Tensorflow or PyTorch to be synthesized into efficient hardware representations for the traditional FPGA and ASIC flows. This breakthrough addresses the growing need for reduced design turnaround and easy verification of ML hardware accelerators with low latency and power efficiency constraints. During this tutorial, we will demonstrate how Python complements HLS by simplifying the ML design process, bridging the gap between software and hardware development. Attendees will explore how we translate neural networks modeled in Python into fixed-point C++ models suitable for HLS workflows. We will dive into strategies like Value-Range Analysis and Quantization-Aware Training, which optimize these designs for deployment and evaluate their accuracy, power consumption, and energy efficiency. To exemplify these concepts, experts from Fermilab will share their experiences applying this technology to high-energy physics experiments, where real-time, low-latency processing is critical. Over the years, Fermilab engineers have demonstrated how deep neural networks, optimized for hardware using hls4ml, can meet the stringent requirements of trigger systems at the CERN Large Hadron Collider. These systems rely on rapid decision-making to process immense data volumes while retaining only the most relevant events for further analysis. The application of hls4ml has also been extended to innovative technologies like smart pixel arrays. These smart pixels integrate ML inference capabilities directly into sensor devices, enabling localized data processing at the pixel level. This approach drastically reduces the need to transmit raw data to external processing units, significantly decreasing power consumption and latency. By embedding neural networks within the pixel architecture, the smart pixels can identify and prioritize relevant data in real time, providing a highly efficient solution for edge computing in scenarios such as particle detectors and imaging systems. Fermilab's work highlights the potential of hardware-accelerated ML in scenarios where both speed and power efficiency are mission-critical. Through this tutorial, attendees will gain valuable insights into the challenges and solutions of deploying ML in hardware. Understanding how HLS and hls4ml streamline the development of neural network-based hardware accelerators is fundamental for the industry's future. Participants will learn how these technologies are shaping the future of AI and scientific computing.

Di Guglielmo, Giuseppe [Fermilab]↗
DOE OSTI2025

Improving the Transportability of a Deep Learning Denoising Model Using Transfer Learning Techniques

The adoption of machine learning techniques in the seismology community has led to great performance improvements in several areas, including signal processing. Specifically, the development of deep learning–based seismic waveform denoising models has the potential to yield improvements in signal detection capabilities for networks operating in particularly noisy environments. Recent advancements in the design of these deep learning denoising models have included the incorporation of continuous and discrete wavelet transform functions into the network architecture to improve the learning capabilities and efficiency of said models. These wavelet transform–based seismic denoising models have shown improved denoising capabilities in regions where there is good agreement between the data features present in the training and evaluation datasets. However, questions remain about the overall transportability of these models to other monitoring regions. Here, in this study, we will determine the baseline transportability of a newly developed multilevel wavelet‐transform convolutional neural network (MWCNN) seismic denoising model. We accomplish this by taking a version of the MWCNN denoising model trained on data collected from the Utah region and evaluating its denoising performance on datasets collected from the neighboring Nevada region, which differ with regard to monitoring sensor types and event histories. We find that there is a notable variability in denoising performance related to the degree of similarity between the initial and new target datasets. The most notable difference in denoising performance is the ability of the denoising model to preserve accurate amplitude information associated with the signal energy present in the waveform data. Finally, we evaluate the ability of transfer learning techniques to improve the transportability of the MWCNN denoising model. We find that although there is still a performance gap present in the denoising results of the MWCNN model, transfer learning did yield improved results.

Quinones, Louis [Sandia National Laboratories (SNL↗
DOE OSTI2021

Machine learning methods for probabilistic locked-mode predictors in tokamak plasmas

A rotating tokamak plasma can interact resonantly with the external helical magnetic perturbations, also known as error fields. This can lead to locking and then to disruptions. We leverage machine learning (ML) methods to predict the locking events. We use a coupled third-order nonlinear ordinary differential equation model to represent the interaction of the magnetic perturbation and the plasma rotation with the error field. This model is sufficient to describe qualitatively the locking and unlocking bifurcations. Here, we explore using ML algorithms with the simulation data and experimental data, focusing on the methods that can be used with sparse datasets. These methods lead to the possibility of the avoidance of locking in real-time operations. We describe the operational space in terms of two control parameters: the magnitude of the error field and the rotation frequency associated with the momentum source that maintains the plasma rotation. The outcomes are quan- tified by order parameters that completely characterize the state, whether locked or unlocked. We use unsupervised ML methods to classify locked/unlocked states and note the usefulness of a certain normalization of the order parameters. Three supervised ML classifiers are used in suite to estimate the probability of locking in the region of control parameter space with hysteresis, i.e., the set of control parameters for which both locked and unlocked states can exist. The results show that a neural network gives the best estimate of the locking probability. An analogy of the present locking model with the van der Waals equation of state is also provided.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2022

Fault Injection for TensorFlow Applications

As machine learning (ML) has seen increasing adoption in safety-critical domains (e.g., autonomous vehicles), the reliability of ML systems has also grown in importance. While prior studies have proposed techniques to enable efficient error-resilience (e.g., selective instruction duplication), a fundamental requirement for realizing these techniques is a detailed understanding of the application’s resilience. In this work, we present TensorFI 1 and TensorFI 2, high-level fault injection (FI) frameworks for TensorFlow-based applications. TensorFI 1 and 2 are able to inject both hardware and software faults in any general TensorFlow 1 and 2 program respectively. Both are configurable FI tools that are flexible, easy to use, and portable. They can be integrated into existing TensorFlow programs to assess their resilience for different fault types (e.g., bit-flips in particular operations or layers). We use the TensorFI 1 and TensorFI 2 to evaluate the resilience of 12 and 10 ML programs written in TensorFlow, including DNNs used in the autonomous vehicle domain. The results give us insights into why some of the models are more resilient. We also measure the performance overheads of the two injectors, and present 4 case studies, two for each tool, to demonstrate their utility.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Inexact iterative numerical linear algebra for neural network-based spectral estimation and rare-event prediction

Understanding dynamics in complex systems is challenging because there are many degrees of freedom, and those that are most important for describing events of interest are often not obvious. The leading eigenfunctions of the transition operator are useful for visualization, and they can provide an efficient basis for computing statistics, such as the likelihood and average time of events (predictions). Here, we develop inexact iterative linear algebra methods for computing these eigenfunctions (spectral estimation) and making predictions from a dataset of short trajectories sampled at finite intervals. We demonstrate the methods on a low-dimensional model that facilitates visualization and a high-dimensional model of a biomolecular system. Implications for the prediction problem in reinforcement learning are discussed.

Chemistry↗
DOE OSTI2024

Device Feasibility Analysis of Multi-level FeFETs for Neuromorphic Computing

As an emerging non-volatile memory device technology, Ferroelectric Field-Effect Transistors (FeFETs) can enable low-power, adaptive intelligent system design. However, device dimension and operating voltage dependent reliability issues of scaled FeFETs can ultimately lead to degraded performance in solving machine learning tasks. In this article, detailed experimental characterization of FeFET devices of different dimensions have been carried out to explicitly evaluate the non-ideal behavior in device conductance programming properties like number of programming states, cycle-to-cycle (C2C) variations, device-to-device (D2D) variations, and state retention. A hardware-aware software simulation approach has been adopted to capture the adversarial effects of the non-idealities on recognition accuracy through algorithm-level performance assessment by including them in NeuroSim, a popular neural network hardware simulator, to execute a neural network model considering all other hardware constraints. With the added non-idealities, significant accuracy degradation has been observed compared to the ideal scenarios where D2D variations play the most critical role. Thereafter, feasibility of a variation-aware training method has been evaluated to tackle the accuracy drop.

42 ENGINEERING↗
NASA NTRS1994

From neural-based object recognition toward microelectronic eyes

Engineering neural network systems are best known for their abilities to adapt to the changing characteristics of the surrounding environment by adjusting system parameter values during the learning process. Rapid advances in analog current-mode design techniques have made possible the implementation of major neural network functions in custom VLSI chips. An electrically programmable analog synapse cell with large dynamic range can be realized in a compact silicon area. New designs of the synapse cells, neurons, and analog processor are presented. A synapse cell based on Gilbert multiplier structure can perform the linear multiplication for back-propagation networks. A double differential-pair synapse cell can perform the Gaussian function for radial-basis network. The synapse cells can be biased in the strong inversion region for high-speed operation or biased in the subthreshold region for low-power operation. The voltage gain of the sigmoid-function neurons is externally adjustable which greatly facilitates the search of optimal solutions in certain networks. Various building blocks can be intelligently connected to form useful industrial applications. Efficient data communication is a key system-level design issue for large-scale networks. We also present analog neural processors based on perceptron architecture and Hopfield network for communication applications. Biologically inspired neural networks have played an important role towards the creation of powerful intelligent machines. Accuracy, limitations, and prospects of analog current-mode design of the biologically inspired vision processing chips and cellular neural network chips are key design issues.

Sheu, Bing J.↗
DOE OSTI2020

nPINNs: nonlocal Physics-Informed Neural Networks for a parametrized nonlocal universal Laplacian operator. Algorithms and Applications

Physics-informed neural networks (PINNs) are effective in solving inverse problems based on differential and integro-differential equations with sparse, noisy, unstructured, and multifidelity data. PINNs incorporate all available information, including governing equations (reflecting physical laws), initial-boundary conditions, and observations of quantities of interest, into a loss function to be minimized, thus recasting the original problem into an optimization problem. In this paper, we extend PINNs to parameter and function inference for integral equations such as nonlocal Poisson and nonlocal turbulence models, and we refer to them as nonlocal PINNs (nPINNs). The contribution of the paper is three-fold. First, we propose a unified nonlocal Laplace operator, which converges to the classical Laplacian as one of the operator parameters, the nonlocal interaction radius δ goes to zero, and to the fractional Laplacian as δ goes to infinity. This universal operator forms a super-set of classical Laplacian and fractional Laplacian operators and, thus, has the potential to fit a broad spectrum of data sets. We provide theoretical convergence rates with respect to δ and verify them via numerical experiments. Second, we use nPINNs to estimate the two parameters, δ and α, characterizing the kernel of the unified operator. The strong non-convexity of the loss function yielding multiple (good) local minima reveals the occurrence of the operator mimicking phenomenon, that is, different pairs of estimated parameters could produce multiple solutions of comparable accuracy. Third, we propose another nonlocal operator with spatially variable order α(γ), which is more suitable for modeling turbulent Couette flow. Our results show that nPINNs can jointly infer this function as well as δ. More importantly, these parameters exhibit a universal behavior with respect to the Reynolds number, a finding that contributes to our understanding of nonlocal interactions in wall-bounded turbulence.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗
DOE OSTI2024

Deep Learning-Based Failure Prognostic Model for PV Inverter Using Field Measurements

Here, this study presents a novel approach for the precise monitoring and prognosis of photovoltaic (PV) inverter status, which is crucial for the proactive maintenance of PV systems. It addresses the gaps in traditional model-based methods, which tend to neglect the overall reliability of inverters, and the limitations of data-driven approaches that largely depend on simulated data. This research presents a robust solution applicable to real-world scenarios. The proposed data-driven model for PV inverter failure prognosis employs actual inverter measurements, integrating various operational and weather-related factors based on domain knowledge. This approach effectively represents inverter stressors and operational status. Utilizing an Enhanced Siamese Convolutional Neural Network (ESCNN), the model merges operational data with domain knowledge features, redefining the prognosis challenge as a classification task. Furthermore, the paper discusses an ESCNN-based real-time inverter failure monitoring method developed on the well-trained model. The proposed models are rigorously trained and tested with real inverter data and a novel filtering method is included to address accidental failures in practical scenarios. The results validate the model's efficacy, and the directions for future research are also outlined.

42 ENGINEERING↗
DOE OSTI2025

Non-destructive structural characterization of graphite components using mechanical resonance and deep learning

As compared to conventional nuclear reactors, microreactors have the potential to significantly reduce construction timelines and capital costs, decreasing the barriers for advanced nuclear reactor technologies. However, the lower power output of these microreactors (typically < 20 MWe) creates challenging economics if operation and maintenance costs cannot be sufficiently reduced. The compact size of these designs presents an opportunity for comprehensive in-situ structural health monitoring to provide real-time feedback in order to reduce operational costs associated with maintenance and downtime. Many microreactor concepts use graphite for both in-core neutron moderation and as a structural material, which has typically required some form of periodic and laborious inspection. This report provides a description and assessment of recent work with graphite to couple acoustic-based experimental measurements and characterization with machine learning models to mature structural health monitoring capabilities and generate benefits for the nuclear microreactor industry. With resilient embedded sensors in development in other programs funded by the US Department of Energy’s Office of Nuclear Energy and elsewhere, the work described herein builds upon previously funded efforts to mature non-destructive testing technology that relates measured vibrational signatures to structural changes, using a combination of new experimental measurements and machine learning processing. Building on past successful demonstrations of predictive workflows to identify structural changes in a hexagonal stainless steel test article with excellent acoustic propagation, we first performed baseline characterization on graphite samples with canonical geometries to ensure compatibility and confidence in the applied techniques for a material with distinctly different mechanical properties. In contrast to efforts in previous years, we worked exclusively with unidirectional vibration data that is more comparable to those expected from the existing embedded sensor technologies which are suitable for deployment in a reactor setting. Established acoustic and modern machine-learning-based characterization approaches were applied to the resulting datasets from these simple geometries. Both approaches were found to be highly capable of detecting even small geometric irregularities amongst nominally identical samples. As such, we then moved to testing these approaches for detection of artificial local stress perturbations introduced into a more complex geometry: a hexagonal block with drilled holes. A main outcome of this work is that a generalizable ML workflow can be used to detect and predict the characteristics of small artificial anomalies in a graphite component with a relevant geometry. While this work was performed using surficial vibration data, we expect the approach to be flexible and viable for other monitoring scenarios, such as those with different arrangements or types of sensor arrays. As compared to previously funded efforts, an existing ML workflow based on neural networks was enhanced through the addition of recently developed Fourier neural operators. As applied to previously collected and new vibration datasets, prediction accuracies of anomaly characterizations were greatly improved with minimal added computational cost. As trained on small durations of vibration data (tens of seconds) collected over a realistic number of locations, the model was able to reliably determine the presence of a subtle stress anomaly and begin to provide location estimates. Such an approach is likely to be viable for more relevant reactor damage scenarios for graphite components, such as progressive crack growth or creep.

36 MATERIALS SCIENCE↗
DOE OSTI2024

E(n)-Equivariant cartesian tensor message passing interatomic potential

Machine learning potential (MLP) has been a popular topic in recent years for its capability to replace expensive first-principles calculations in some large systems. Meanwhile, message passing networks have gained significant attention due to their remarkable accuracy, and a wave of message passing networks based on Cartesian coordinates has emerged. However, the information of the node in these models is usually limited to scalars, and vectors. In this work, we propose High-order Tensor message Passing interatomic Potential (HotPP), an E(n) equivariant message passing neural network that extends the node embedding and message to an arbitrary order tensor. By performing some basic equivariant operations, high order tensors can be coupled very simply and thus the model can make direct predictions of high-order tensors such as dipole moments and polarizabilities without any modifications. The tests in several datasets show that HotPP not only achieves high accuracy in predicting target properties, but also successfully performs tasks such as calculating phonon spectra, infrared spectra, and Raman spectra, demonstrating its potential as a tool for future research.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2022

A Length Adaptive Algorithm-Hardware Co-design of Transformer on FPGA Through Sparse Attention and Dynamic Pipelining

Transformers are considered one of the most important deep learning models since 2018, in part because it establishes state-of-the-art (SOTA) records and could potentially replace existing Deep Neural Networks (DNNs). Despite the remarkable triumphs, the prolonged turnaround time of Transformer models is a widely recognized roadblock. The variety of sequence lengths imposes additional computing overhead where inputs need to be zero-padded to the maximum sentence length in the batch to accommodate the parallel computing platforms. This paper targets the field-programmable gate array (FPGA) and proposes a coherent sequence length adaptive algorithm–hardware co-design for Transformer acceleration. Particularly, we develop a hardware-friendly sparse attention operator and a length-aware hardware resource scheduling algorithm. The proposed sparse attention operator brings the complexity of attention-based models down to linear complexity and alleviates the off-chip memory traffic. The proposed length-aware resource hardware scheduling algorithm dynamically allocates the hardware resources to fill up the pipeline slots and eliminates bubbles for NLP tasks. Experiments show that our design has very small accuracy loss and has 80.2 × and 2.6 × speedup compared to CPU and GPU implementation, and 4 × higher energy efficiency than state-of-the-art GPU accelerator optimized via CUBLAS GEMM.

Peng, Hongwu↗