Search NASA⌕ Search

SEARCH · Search NASA

Results for “classification algorithms”

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

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

At least 1,243 records · Page 69
All sourcesNASA NTRSDOE OSTI
NASA NTRS1985

Discrete Random Media Techniques for Microwave Modeling of Vegetated Terrain

Microwave remote sensing of vegetated terrain has been studied. Vegetation is modeled so that backscattered radar signals can be used to infer parameters which characterize the vegetation and underlying ground. The vegetation is modeled by discrete lossy dielectric scatterers with prescribed characteristics. The goal of the modeling effort is to remotely sense vegetation type (classification), growth stage, and plant/ground moisture. This information can then be used as input into agricultural, forestry and global circulation models. The microwave frequency spectrum, particularly L and C bands, are especially appropriate for this purpose since the wavelength is comparable to plant leaf and stem size. The resulting resonant interaction leads to backscattered data highly depend on plant shape and orientation. In addition, the transparent nature of the atmosphere in this frequency regime allows for algorithm development which requires no atmospheric correction.

Lang, R. H.↗
DOE OSTI2025

The System for Classification of Low-Pressure Systems (SyCLoPS): An All-In-One Objective Framework for Large-Scale Data Sets

We propose the first unified objective framework (SyCLoPS) for detecting and classifying all types of low-pressure systems (LPSs) in a given data set. We use the state-of-the-art automated feature tracking software TempestExtremes (TE) to detect and track LPS features globally in ERA5 and compute 16 parameters from commonly found atmospheric variables for classification. A Python classifier is implemented to classify all LPSs at once. The framework assigns 16 different labels (classes) to each LPS data point and designates four different types of high-impact LPS tracks, including tracks of tropical cyclone (TC), monsoonal system, subtropical storm and polar low. The classification process involves disentangling high-altitude and drier LPSs, differentiating tropical and non-tropical LPSs using novel criteria, and optimizing for the detection of the four types of high-impact LPS. A comparison of our labels with those in the International Best Track Archive for Climate Stewardship (IBTrACS) revealed an overall accuracy of 95% in distinguishing between tropical systems, extratropical cyclones, and disturbances. SyCLoPS produces a better TC detection skill compared to the previous algorithms, highlighted by an approximately 6% reduction in the false alarm rate compared to the previous TE algorithm. The vertical cross section composite of the four types of high-impact LPS we detect each shows distinct structural characteristics. Finally, we demonstrate that SyCLoPS is valuable for investigating various aspects of LPSs in climate data, such as the evolution of a single LPS track, patterns of LPS frequencies, and precipitation or wind influence associated with a particular LPS class.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2024

Digital Analytics, Causal Knowledge Acquisition and Reasoning for Technical Language Processing

Complex engineering systems such as nuclear power plants (NPPs) generate and collect large amounts of equipment reliability (ER) data elements that contain information on the status of components, assets, and systems. Some of this information is textual in form and can be found in documents such as incident reports (IRs) and work orders (WOs). Analyses of textual data in current NPPs-using natural language processing (NLP) methods-have been expanded over the last decade, and it is only recently that the true potential of such analyses has emerged. So far, applications of NLP methods have mostly been limited to classification and prediction, the goal being to identify the nature of the textual element (e.g., safety or non-safety related). Here, we target a more complex problem: automatically extracting knowledge from a textual element in order to assist system engineers in conducting system health assessments. Knowledge extraction is a very broad concept, and its definition may vary depending on the application context. Our methods are a blend of both rule-based and machine learning (ML) algorithms. For our purposes, knowledge extraction means identifying the systems or assets mentioned in a given textual element, as well as the type of event described (e.g., component failure or maintenance activity). In addition, we want to capture details such as measured quantities and the temporal/cause-effect relations between events. In this tool, we also demonstrate how textual data elements are preprocessed in order to handle typos, acronyms, and abbreviations. One main feature of these methods is that they are not based solely on data, but are in fact model-based. In other words, they also rely on MBSE models that are designed to capture-from a functional point of view-the architecture of the systems/assets under consideration. The main purpose of such models is to digitally emulate system engineers' knowledge of system and asset architecture and to identify dependencies among systems, assets, and components. Provided these models, analyses of textual and numeric ER data can be performed by first identifying the OPM model elements to which the ER data elements are referring. The relationships between ER data elements are then identified by checking for any temporal or logical dependencies.

Mandelli, Diego [Idaho National Laboratory (INL), ↗
DOE OSTI2024

Emulation and detection of physical faults and cyber-attacks on building energy systems through real-time hardware-in-the-loop experiments

The increasing use of remote or mobile access, integrated wearable technologies, data exchange, and cloud-based data analytics in modern smart buildings is steering the building industry towards open communication technologies. The increased connectivity and accessibility could lead to more cyber-attacks in smart buildings. On the other hand, physical faults (e.g., HVAC -heating, ventilation, and air-conditioning faults) may have similar adverse impacts as those from the cyber-attacks on building energy systems, such as occupant discomfort, energy wastage, and equipment downtime. However, current physical behavior-based anomaly detection methods fail to differentiate between cyber-attacks and physical faults in building energy systems. Moreover, the challenge in collecting real-world threat data with ground truth has led researchers to rely on numerical models with user-defined assumptions, which may not accurately reflect real-world conditions due to the lack of in-situ experimental datasets. To address these challenges and gaps, this paper presents a flexible hardware-in-the-loop (HIL) testbed for generating cyber-attack and physical fault datasets and demonstrating threat detection algorithms in a real building automation system (BAS) environment. This testbed combines hardware (i.e., real BAS with local HVAC controllers and a physical network) with software (i.e., high-fidelity models to represent behaviors of building envelope and HVAC energy systems), enabling emulations of realistic threats. Five HIL experiments, including one baseline without any threats, two with physical faults, and two with cyber-attacks, were conducted to generate datasets containing detailed network traffic and system states. A joint classification framework, incorporating a network analyzer and a physical HVAC fault detector, was proposed to automatically detect cyber-physical abnormalities on BAS at both the network and the physical HVAC levels. The network analyzer comprises a conditional random fields (CRF) based command validator and a statistics-based detection strategy. The fault detector employs a weather and schedule-based pattern matching and feature-based principal component analysis (WPM-FPCA) method. Evaluation of the classification using four metrics from the multi-class confusion matrix revealed an average accuracy of 90.2%, recall of 89.7%, precision of 88.5% and F1-score of 89.2%. Finally, these results demonstrate that the proposed joint classification framework can effectively differentiate between specific types of cyber-attacks (e.g., device reinitialization attack, network Denial-of-Service attack) and physical faults (e.g., air handling unit operational fault, cooling coil valve stuck) in real time for improved building energy management.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2026

Wasserstein normalized autoencoder for anomaly detection

A novel anomaly detection algorithm is presented. The Wasserstein normalized autoencoder (WNAE) is a normalized probabilistic model that minimizes the Wasserstein distance between the learned probability distribution—a Boltzmann distribution where the energy is the reconstruction error of the autoencoder (AE)—and the distribution of the training data. This algorithm has been developed and applied to the identification of semivisible jets—conical sprays of visible standard model (SM) particles and invisible dark matter states—with the CMS experiment at the CERN LHC. Trained on jets of particles from simulated SM processes, the WNAE is shown to learn the probability distribution of the input data in a fully unsupervised fashion, such that it effectively identifies new physics jets as anomalies. The model exhibits stable, convergent training and recovers strong classification performance for a wide range of signals against the selected background process, for which a standard AE fails because of outlier reconstruction. In addition, the model improves upon standard normalized autoencoders while remaining fully agnostic to the signal. The WNAE directly tackles the problem of outlier reconstruction, a common failure mode of autoencoders in anomaly detection tasks.

Hayrapetyan, Aram [Yerevan Phys. Inst.]↗
DOE OSTI2024

PAIReD jet: A multi-pronged resonance tagging strategy across all Lorentz boosts

We propose a new approach of jet-based event reconstruction that aims to optimally exploit correlations between the products of a hadronic multi-pronged decay across all Lorentz boost regimes. The new approach utilizes clustered small-radius jets as seeds to define unconventional jets, referred to as PAIReD jets. The constituents of these jets are subsequently used as inputs to machine learning-based algorithms to identify the flavor content of the jet. We demonstrate that this approach achieves higher efficiencies in the reconstruction of signal events containing heavy-flavor jets compared to other event reconstruction strategies at all Lorentz boost regimes. Classifiers trained on PAIReD jets also have significantly better background rejections compared to those based on traditional event reconstruction approaches using small-radius jets at low Lorentz boost regimes. The combined effect of a higher signal reconstruction efficiency and better classification performance results in a two to four times stronger rejection of light-flavor jets compared to conventional strategies at low Lorentz-boosts, and rejection rates similar to classifiers based on large-radius multi-pronged jets at high Lorentz-boost regimes.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2025

A Novel Protection Scheme for Unbalanced Faults in Inverter Dominated Networks: A Computationally Efficient Algorithm for Entry-Level Relays

Microgrids are now a common practice in distribution systems to increase resilience and reliability. However, microgrid protection remains a critical challenge, considering its requirement to operate in both grid connected and islanded, and the variability in fault characteristics under each mode of operation. This paper presents unbalanced power (S unb ) based fault detection algorithm, which considers local voltage and current unbalances to determine faults in the system. S unb is a computationally efficient fault detection algorithm that is suitable for implementation in the programmable logic of entry level protective relays. In addition, the difference in current and voltage unbalance (D n ) is used to determine the fault type. The proposed method demonstrates high sensitivity and selectivity for line-to-ground (LG), line-to-line (LL), and double line-to-ground (LLG) faults, representing the most common faults in distribution systems. It also allows relay coordination with upstream and downstream protection devices in both island and grid connected operation, while preserving grading margins. The same pickup and time multiplier settings of a particular relay for both modes of operation eliminates the need for adaptive settings, which rely on communication networks. Validation was performed with a hardware-in-the-loop (HIL) setup using Typhoon HIL real time simulator interfaced with three entry-level, SEL 751 relays. Results confirmed the algorithm’s ability to discriminate fault conditions, and determine the fault type under both operating modes, maintain fast detection times, and ensure proper protection coordination.

fault classification↗
DOE OSTI2024

Energy Efficient Streaming Time Series Classification with Attentive Power Iteration

Efficiently processing time series data streams in real-time on resource-constrained devices offers significant advantages in terms of enhanced computational energy efficiency and reduced time-related risks. We introduce an innovative streaming time series classification network that utilizes attentive power iteration, enabling real-time processing on resource-constrained devices. Our model continuously updates a compact representation of the entire time series, enhancing classification accuracy while conserving energy and processing time. Notably, it excels in streaming scenarios without requiring complete time series access, enabling swift decisions. Experimental results show that our approach excels in classification accuracy and energy efficiency, with over 70% less consumption and threefold faster task completion than benchmarks. This work advances real-time responsiveness, energy conservation, and operational effectiveness for constrained devices, contributing to optimizing various applications.

97 MATHEMATICS AND COMPUTING↗
NASA NTRS2002

Wire Detection Algorithms for Navigation

In this research we addressed the problem of obstacle detection for low altitude rotorcraft flight. In particular, the problem of detecting thin wires in the presence of image clutter and noise was studied. Wires present a serious hazard to rotorcrafts. Since they are very thin, their detection early enough so that the pilot has enough time to take evasive action is difficult, as their images can be less than one or two pixels wide. Two approaches were explored for this purpose. The first approach involved a technique for sub-pixel edge detection and subsequent post processing, in order to reduce the false alarms. After reviewing the line detection literature, an algorithm for sub-pixel edge detection proposed by Steger was identified as having good potential to solve the considered task. The algorithm was tested using a set of images synthetically generated by combining real outdoor images with computer generated wire images. The performance of the algorithm was evaluated both, at the pixel and the wire levels. It was observed that the algorithm performs well, provided that the wires are not too thin (or distant) and that some post processing is performed to remove false alarms due to clutter. The second approach involved the use of an example-based learning scheme namely, Support Vector Machines. The purpose of this approach was to explore the feasibility of an example-based learning based approach for the task of detecting wires from their images. Support Vector Machines (SVMs) have emerged as a promising pattern classification tool and have been used in various applications. It was found that this approach is not suitable for very thin wires and of course, not suitable at all for sub-pixel thick wires. High dimensionality of the data as such does not present a major problem for SVMs. However it is desirable to have a large number of training examples especially for high dimensional data. The main difficulty in using SVMs (or any other example-based learning method) is the need for a very good set of positive and negative examples since the performance depends on the quality of the training set.

Kasturi, Rangachar↗
DOE OSTI2018

Parallel k-means Clustering of Geospatial Data Sets Using Manycore CPU Architectures

The increasing availability of high-resolution geospatiotemporal data sets from sources such as observatory networks, remote sensing platforms, and computational Earth system models has opened new possibilities for knowledge discovery and mining of weather, climate, ecological, and other geoscientific data sets fused from disparate sources. Many of the standard tools used on individual workstations are impractical for the analysis and synthesis of data sets of this size; however, new algorithmic approaches that can effectively utilize the complex memory hierarchies and the extremely high levels of parallelism available in state-of-the-art high-performance computing platforms can enable such analysis. Here, we describe pKluster, an open-source tool we have developed for accelerated k-means clustering of geospatial and geospatiotemporal data, and discuss algorithmic modifications and code optimizations we have made to enable it to effectively use parallel machines based on novel CPU architectures—such as the Intel “Knights Landing” Xeon Phi and Skylake Xeon processors—with many cores and hardware threads, and employing significant single instruction, multiple data (SIMD) parallelism. We outline some applications of the code in ecology and climate science contexts and present a detailed discussion of the performance of the code for one such application, LiDAR-derived vertical vegetation structure classification.

Mills, Richard T↗
NASA NTRSDate not supplied

Towards an Aviation Large Language Model by Fine-tuning and Evaluating Transformers

In the aviation domain, there are many applications for machine learning and artificial intelligence tools that utilize natural language. For example, there is a desire to know the commonalities in written safety reports such as voluntary post incidents reports or aerial wildfire operations reports to better understand the risks present. Another use-case is the possibility of extracting airspace procedures and constraints currently written in documents such as Letters of Agreement. These applications can benefit from the use of state-of-the-art natural language processing techniques when adapted to the language/phraseology specific to the aviation domain. This paper evaluates the viability of adaptation of NLP tools to the aviation domain by fine-tuning transformer based models using aviation data sets. In 2018, a novel language model based on neural units (also called transformers) was created and became known as “Bidirectional Encoder Representations from Transformers” or BERT. This architecture combined with large amounts of English training data and innovative semi-supervised training tasks set the standard for what would later emerge as Large Language Models. The performance of these models was further improved by hyperparameter tuning and refinement of the semi-supervised training task and resulted in “Robustly Optimized BERT Pre-training Approach through hyperparameter tuning” or RoBERTa models. These pre-trained Large Language Models proved to be useful for a wide variety of natural language processing tasks such as text classification and question answering through a process called fine-tuning. The transformer architecture with pre-trained weights served as the basis with the last few layers replaced with layers fine-tuned to perform a new task e.g., a layer that provides a label for the entire input text. This process of fine-tuning can also be used to adapt the models to new domains; e.g., BioBERT started with the pre-trained BERT model and was completed by additional fine-tuning and training on biomedical documents. Transformer-based architectures can also be used to create rich representations of text called embeddings which can serve as the input to other machine learning models. This allows simpler algorithms such as logistic regression to use context-rich representations of the text while still remaining quick to train and evaluate. In the world of aviation, there is a growing demand for natural language processing and understanding but the domain presents unique challenges. Due to the technical content (and specialized language) of most aviation documents, fine-tuning pre-trained Large Language Models to specific tasks has not met the benchmark on natural language processing tasks set by simpler models trained from scratch on the data. To address this deficiency, this paper evaluates the improvements from fine-tuning a Large Language Model on a large set of aviation documents using the original semi-supervised training tasks before performing specific natural language tasks. In fine-tuning, a domain-specific dataset is used on the original training task but with the pre-trained Large Language Model instead of starting from a random initialization. This approach allows the model to be adapted to the specific domain language without discarding the information gained from training on general English data. This paper utilized two major dataset types to train and assess the RoBERTa fine-tuning performance. The first are 7,057 Letters of Agreement which are Federal Aviation Administration (FAA) documents that formalize airspace operations across the national airspace system. They contain many examples of ‘aviation English’ using domain specific terminology and phrasing which serves as a representative basis to perform the semi-supervised fine-tuning. The second type is the 494 document classification labels to be used for evaluation. This down-stream evaluation aims to show the performance of the fine-tuned model, better understand how much data is needed for an effective fine-tuning, and how fine-tuning can be adapted for different applications in-the domain. After semi-supervised training, evaluation begins by encoding the documents for classification using the fine-tuned RoBERTa model. Then a logistic regression classifier is trained to label the document type and compared against our ground truth labels. This currently leads to a 82.8% accuracy on 10-fold cross validation showing improvement over baseline RoBERTa which achieved 81.0%. We plan to measure the improvements on additional tasks and it is expected that these improvements will lead to more robust models that can tackle the natural language processing challenges present in aviation datasets.

ATM↗
DOE OSTI2023

PCAfold 2.0—Novel tools and algorithms for low-dimensional manifold assessment and optimization

We describe an update to our open-source Python package, PCAfold, designed to help researchers generate, analyze and improve low-dimensional data manifolds. In the current version, PCAfold 2.0, we introduce novel tools and algorithms for assessing and optimizing low-dimensional manifolds. This includes a method that generates a “map” of local feature sizes that can help pinpoint researchers to problematic regions on a manifold. We introduce a novel cost function that characterizes the quality of a manifold topology with a single number. We develop two algorithms for feature selection based on principal component analysis (PCA) that use the cost function as an objective function to minimize. We introduce a quantity of interest (QoI)-aware dimensionality reduction strategy where data projections are computed using an artificial neural network and are directly optimized towards representing various projection-independent and projection-dependent QoIs. We also introduce an implementation of partition of unity networks (POUnets) for efficient reconstruction of QoIs from low-dimensional manifolds based on combining neural network classification with localized polynomial regression. Our software can be broadly applicable in all domains of science and engineering that aim to reduce data dimensionality, as well as in the fundamental research on representation learning.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2020

Evolving Ensembles of Spiking Neural Networks for Neuromorphic Systems

Evolutionary algorithms have been proposed as a solution to overcome many of the challenges associated with training spiking neural networks. While evolutionary optimization for spiking neural networks is very flexible, its performance has difficulty scaling to complex tasks and correspondingly complex network structures. Here we propose a method for evolving ensembles of spiking neural networks. By using ensemble learning, the flexibility of evolutionary optimization is fully preserved while scaling to more challenging tasks. We test the performance of the proposed method using handwritten digit classification. We investigate multiple strategies for constructing ensembles of spiking neural networks, and demonstrate that evolving ensembles of SNNs offers significant performance advantages over evolutionary optimization.

Elbrecht, Daniel↗
DOE OSTI2022

Deep Learning for In-Situ Layer Quality Monitoring during Laser-Based Directed Energy Deposition (LB-DED) Additive Manufacturing Process

Defects are a leading issue for the rejection of parts manufactured through the Directed Energy Deposition (DED) Additive Manufacturing (AM) process. In an attempt to illuminate and advance in situ quality monitoring and control of workpieces, we present an innovative data-driven method that synchronously collects sensing data and AM process parameters with a low sampling rate during the DED process. The proposed data-driven technique determines the important influences that individual printing parameters and sensing features have on prediction at the inter-layer qualification to perform feature selection. Three Machine Learning (ML) algorithms including Random Forest (RF), Support Vector Machine (SVM), and Convolutional Neural Network (CNN) are used. During post-production, a threshold is applied to detect low-density occurrences such as porosity sizes and quantities from CT scans that render individual layers acceptable or unacceptable. This information is fed to the ML models for training. Training/testing are completed offline on samples deemed “high-quality” and “low-quality”, utilizing only features recorded from the build process. CNN results show that the classification of acceptable/unacceptable layers can reach between 90% accuracy while training/testing on a “high-quality” sample and dip to 65% accuracy when trained/tested on “low-quality”/“high-quality” (respectively), indicating over-fitting but showing CNN as a promising inter-layer classifier.

36 MATERIALS SCIENCE↗
NASA NTRS2022

Science Autonomy for Ocean Worlds Astrobiology: A Perspective

Astrobiology missions to ocean worlds in our solar system must overcome both scientific and technological challenges due to extreme temperature and radiation conditions, long communication times, and limited bandwidth. While such tools could not replace ground-based analysis by science and engineering teams, machine learning algorithms could enhance the science return of these missions through development of autonomous science capabilities. Examples of science autonomy include onboard data analysis and subsequent instrument optimization, data prioritization (for transmission), and real-time decision-making based on data analysis. Similar advances could be made to develop streamlined data processing software for rapid ground-based analyses. Here we discuss several ways machine learning and autonomy could be used for astrobiology missions, including landing site selection, prioritization and targeting of samples, classification of “features” (e.g., proposed biosignatures) and novelties (uncharacterized, “new” features, which may be of most interest to agnostic astrobiological investigations), and data transmission.

ocean worlds↗
DOE OSTI2024

Statistical properties of filaments in the cosmic web

ABSTRACT In the context of the cosmological and constrained Exploring the Local Universe with the reConstructed Initial Density field (ELUCID) simulation, this study explores the statistical characteristics of filaments within the cosmic web, focussing on aspects such as the distribution of filament lengths and their radial density profiles. Using the classification of the cosmic web environment through the Hessian matrix of the density field, our primary focus is on how cosmic structures react to the two variables $R_{\rm s}$ and $\lambda _{\rm th}$. The findings show that the volume fractions of knots, filaments, sheets, and voids are highly influenced by the threshold parameter $\lambda _{\rm th}$, with only a slight influence from the smoothing length $R_{\rm s}$. The central axis of the cylindrical filament is pinpointed using the medial-axis thinning algorithm of the COsmic Web Skeleton (COWS) method. It is observed that median filament lengths tend to increase as the smoothing lengths increase. Analysis of filament length functions at different values of $R_{\rm s}$ indicates a reduction in shorter filaments and an increase in longer filaments as $R_{\rm s}$ increases, peaking around $2.5R_{\rm s}$. The study also shows that the radial density profiles of filaments are markedly affected by the parameters $R_{\rm s}$ and $\lambda _{\rm th}$, showing a valley at approximately $2R_{\rm s}$, with increases in the threshold leading to higher amplitudes of the density profile. Moreover, shorter filaments tend to have denser profiles than their longer counterparts.

Zhang, Youcai (ORCID:0000000319674091)↗
DOE OSTI2019

FPGAs as a Service to Accelerate Machine Learning Inference [PowerPoint]

Large-scale particle physics experiments face challenging demands for high-throughput computing resources both now and in the future. New heterogeneous computing paradigms on dedicated hardware with increased parallelization, such as Field Programmable Gate Arrays (FPGAs), offer exciting solutions with large potential gains. The growing applications of machine learning algorithms in particle physics for simulation, reconstruction, and analysis are naturally deployed on such platforms. We demonstrate that the acceleration of machine learning inference as a web service represents a heterogeneous computing solution for particle physics experiments that requires minimal modification to the current computing model. As examples, we retrain the ResNet-50 convolutional neural network to demonstrate state-of-the-art performance for top quark jet tagging at the LHC, and apply a ResNet-50 model with transfer learning for neutrino event classification. Using Project Brainwave b y Microsoft to accelerate the ResNet-50 image classification model, we achieve average inference times of 60 (10) milliseconds with our experimental physics software framework using Brainwave as a cloud (edge or on-premises) service, representing an improvement by a factor of approximately 30 (175) in model inference latency over traditional CPU inference in current experimental hardware. A single FPGA service accessed by many CPUs achieves a throughput of 600--700 inferences per second using an image batch of one, comparable to large batch-size GPU throughput and significantly better than small batch-size GPU throughput. Deployed as an edge or cloud service for the particle physics computing model, coprocessor accelerators can have a higher duty cycle and are potentially much more cost-effective.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2019

FPGA-Accelerated Machine Learning Inference as a Service for Particle Physics Computing

Large-scale particle physics experiments face challenging demands for high-throughput computing resources both now and in the future. New heterogeneous computing paradigms on dedicated hardware with increased parallelization, such as Field Programmable Gate Arrays (FPGAs), offer exciting solutions with large potential gains. The growing applications of machine learning algorithms in particle physics for simulation, reconstruction, and analysis are naturally deployed on such platforms. We demonstrate that the acceleration of machine learning inference as a web service represents a heterogeneous computing solution for particle physics experiments that potentially requires minimal modification to the current computing model. As examples, we retrain the ResNet-50 convolutional neural network to demonstrate state-of-the-art performance for top quark jet tagging at the LHC and apply a ResNet-50 model with transfer learning for neutrino event classification. Using Project Brainwave by Microsoft to accelerate the ResNet-50 image classification model, we achieve average inference times of 60 (10) ms with our experimental physics software framework using Brainwave as a cloud (edge or on-premises) service, representing an improvement by a factor of approximately 30 (175) in model inference latency over traditional CPU inference in current experimental hardware. A single FPGA service accessed by many CPUs achieves a throughput of 600–700 inferences per second using an image batch of one, comparable to large batch-size GPU throughput and significantly better than small batch-size GPU throughput. Deployed as an edge or cloud service for the particle physics computing model, coprocessor accelerators can have a higher duty cycle and are potentially much more cost-effective.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗