SEARCH · Search NASA

Results for “Neural Network Potentials”

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.

SourceResearch connectionFrom yearThrough year

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

At least 379 records · Page 21

All sources NASA NTRS DOE OSTI

DOE OSTI2025

Machine learning enabled discovery of superhard and ultrahard carbon polymorphs

The demand for multifunctional materials has motivated the move from near-equilibrium materials to metastable i.e. out-of-equilibrium phases that can meet several desired target properties. The search for such metastable phases with exotic properties is non-trivial and often serendipitous. Inverse design approaches based on evolutionary search have been powerful tools, but such traditional searches have focused on identifying primarily stable and metastable materials with the lowest enthalpy. The inverse design of materials, with a focus on a desired property such as, for example, hardness is a challenging task because of the expensive computational cost involved in sampling multiple structures. The recent advances in machine learning have brought new powerful AI techniques to the forefront which can potentially revolutionize the inverse design and discovery of materials, especially metastable phases capable of meeting multifunctionality. Here, in this work, we develop and apply an automated reinforcement learning workflow for inverse design that integrates first principles physics and atomistic simulations with machine learning (ML), and high-performance computing to allow rapid exploration of the superhard and ultrahard metastable phases of Carbon. We demonstrate an automatic machine learning based inverse design workflow to map new undiscovered metastable states ranging from near equilibrium to those far-from-equilibrium that satisfy multiple property objectives, specifically bulk moduli, shear moduli and hardness. We create a comprehensive library of carbon stable and metastable phases with varying hardness and subsequently shortlist 10 top performing candidate carbon structures, including two newly reported phases, based on their hardness and characterize their temperature dependent mechanical properties. A neural network model is built using featurization of allotropes of carbon to predict the quasi-harmonic Gibbs free energies. The Gibbs free energies of the top performing phases are analyzed to get an estimate of the experimental synthesizability of these superhard and ultrahard carbon phases. In general, we show using machine learning based inverse design approaches how hitherto inaccessible metastable states can be identified and potentially synthesized to meet the demand for multifunctional materials.

Balasubramanian, Karthik [Univ. of Illinois, Chica↗

DOE OSTI2024

Development of Whole System Digital Twins for Advanced Reactors: Leveraging Graph Neural Networks and SAM Simulations

Here, in this work, we introduce a novel method to develop whole system digital twins (DTs) for advanced nuclear reactors. This method treats a complex reactor system as a heterogeneous graph: with the system components as different types of graph nodes and their physical interconnections as edges. Based on the heterogeneous graph, a graph neural network combining graph convolution and temporal node attention is developed as the DT, facilitating a comprehensive understanding of the system's dynamic behavior. By utilizing the System Analysis Module (SAM) code for simulating various operational transients, we develop a graph-based database that trains the DT. This DT is characterized by two primary functions: It can infer the entire system's status using sparse node information, and it can predict the progress of transients based on current and historical system information. Our approach is validated through case studies on the Experimental Breeder Reactor II (EBR-II) system and a generic Fluoride-salt-cooled High-temperature Reactor (gFHR), demonstrating the DT's accuracy in forecasting operational transients. The DT's rapid computation capabilities enhance its potential for supporting advanced reactor operations, offering benefits in intelligent simulation, autonomous control, and anomaly detection, paving the way for improved safety analysis and intelligent component health management for advanced reactor systems and reducing their operations and maintenance cost.

EBR-II↗

DOE OSTI2025

A Machine Learning Framework for Modeling Ensemble Properties of Atomically Disordered Materials

Atomic disorder can strongly influence material properties such as charge transport, optical response, and catalytic activity. However, efficiently modeling these disorder effects remains challenging for first-principles methods due to the cost of sampling large configurational spaces and computing complex physical quantities. Recent advances of machine learning techniques, particularly graph neural networks (GNNs), has enabled the efficient and accurate predictions of complex material properties, offering promising tools for studying disordered systems. In this work, we present a general machine-learning-assisted computational framework that integrates equivariant GNNs with Monte Carlo simulations to compute the thermodynamic and ensemble-averaged functional properties of disordered materials. Using the surface-termination-disordered MXene monolayer Ti 3 C 2 T 2–x as a representative system, we find that electrical conductivity exhibits an emergent peak near the order–disorder phase transition temperature due to the interplay between electron scattering and doping. In contrast, optical conductivity remains largely insensitive to local atomic disorder and reflects the global surface chemical composition. These results highlight the role of atomic disorder in affecting material properties and demonstrate the potential of our approach for statistically modeling disorder effects in a wide range of materials such as high-entropy alloys and spin liquids.

MXene↗

DOE OSTI2025

Machine Learning for Anomaly Detection in Neural Network Security and SRF Cavities

This dissertation explores the development and deployment of machine learning approaches to address critical challenges in anomaly detection across two distinct domains: neural network security in federated learning settings and cavity behavior analysis in particle accelerator operations at Jefferson Lab in Newport News, Virginia. Anomaly detection identifies deviations from expected patterns, safeguarding systems in cybersecurity, industry, and research against malicious activities and failures. This dissertation demonstrates how our machine learning approaches enhance detection accuracy and efficiency in both neural network security and industrial applications. First, we investigate vulnerabilities in deep neural networks deployed in federated learning. Although federated learning preserves user privacy by training models locally, it remains vulnerable to backdoor attacks, in which malicious participants embed hidden triggers that induce targeted misbehavior. We propose a self-supervised contrastive learning framework to detect and mitigate such backdoor attacks. In our experiments, this method achieves higher detection accuracy and lower false positive rates than existing defenses, while operating without access to local model updates or original training data and thus preserving the privacy guarantees of the federated setting. Second, we address the operational reliability of superconducting radio-frequency (SRF) cavities at the Continuous Electron Beam Accelerator Facility (CEBAF). Our research leverages an unsupervised learning approach, combined with Principal Component Analysis (PCA) and k-means clustering, to identify anomalous behaviors in SRF cavities. Our method detects subtle anomalous behavior by analyzing SRF signal data. This knowledge allows for the early detection and resolution of potential faults, significantly improving the efficiency and reliability of operations. Third, we extend these insights to time-series anomaly detection more broadly. We design a contrastive-learning based model tailored to increasingly dynamic environments and academic research. This model improves detection accuracy in settings that require real-time monitoring and predictive maintenance. Our research underscores the broader applicability and impact of advanced machine learning techniques in anomaly detection. By extracting meaningful patterns from complex data, machine learning can significantly enhance security in distributed neural networks and improve the efficiency of particle accelerator operations. This dissertation serves as a stepping stone for future investigations into the vast possibilities of anomaly detection, inspiring further exploration and development of machine learning techniques in this field.

Ferguson, Hal [Old Dominion University]↗

DOE OSTI2024

Human limits in machine learning: prediction of potato yield and disease using soil microbiome data

Abstract Background The preservation of soil health is a critical challenge in the 21st century due to its significant impact on agriculture, human health, and biodiversity. We provide one of the first comprehensive investigations into the predictive potential of machine learning models for understanding the connections between soil and biological phenotypes. We investigate an integrative framework performing accurate machine learning-based prediction of plant performance from biological, chemical, and physical properties of the soil via two models: random forest and Bayesian neural network. Results Prediction improves when we add environmental features, such as soil properties and microbial density, along with microbiome data. Different preprocessing strategies show that human decisions significantly impact predictive performance. We show that the naive total sum scaling normalization that is commonly used in microbiome research is one of the optimal strategies to maximize predictive power. Also, we find that accurately defined labels are more important than normalization, taxonomic level, or model characteristics. ML performance is limited when humans can’t classify samples accurately. Lastly, we provide domain scientists via a full model selection decision tree to identify the human choices that optimize model prediction power. Conclusions Our study highlights the importance of incorporating diverse environmental features and careful data preprocessing in enhancing the predictive power of machine learning models for soil and biological phenotype connections. This approach can significantly contribute to advancing agricultural practices and soil health management.

Aghdam, Rosa↗

NASA NTRS2024

Presound: UAV Diagnostic System Enabled by Vibration-Based Machine Learning

A low-weight, inexpensive small unmanned aerial system (sUAS) that takes off, performs a mission, lands, and safely stows and recharges itself has myriad future applications ranging from agricultural imaging to last-mile package delivery. Likewise, Urban Air Mobility (UAM) systems will enable people to take air taxis from point to point in cities, rapidly moving commuters long distances without concern for road traffic and congestion. Fully electric aviation systems will be cleaner and quieter than ground transport. Cities could eliminate cars and buses, and convert roads to higher capacity bike and pedestrian throughways. Yet, for sUAS as well as UAM, system reliability and assurance is a limiting factor to deploying affordable autonomous flight systems. For this bright future of aviation to be realized, aircraft must be able to autonomously and accurately self-diagnose health issues both before takeoff and during flight. The GreenSight PreSound system is designed to identify defects on aircraft through intelligent analysis of vibration. It accomplishes this by measuring structural vibrations induced by the vehicle’s own propellers, and analyzing that data using a machine learning model that determines whether a defect is present. The PreSound system is designed to require no human oversight, and to operate across a wide array of vehicles through re-training of the model for each target aircraft. PreSound has been developed and seen limited early success using data collected from the GreenSight Dreamer sUAS, a 5lb quadrotor vehicle designed for aerial imaging applications. The final detection model, trained on data with props spinning at 50% throttle, achieves excellent performance with over 99% average accuracy in detecting blade damage using a single FFT vector input. It demonstrates the ability to generalize to new types of blade damage, correctly classifying a different type of blade damage with 98% accuracy. Full test pulses were classified with 100% accuracy, and in live testing, all sets of data during blade movement were classified accurately with over 95% confidence. When trained on in-flight data, the same model achieves an average accuracy of 85% in distinguishing between undamaged and blade-damaged states in flight. The authors believe that these accuracies show significant potential of this approach to expand unmanned flight safety, with significant potential benefits in accelerating Advanced Aerial Mobility (AAM) and UAM aviation applications.

UAS↗

NASA NTRSDate not supplied

Inverse Text Normalization of Air Traffic Control System Command Center Planning Telecon Transcriptions

We present a hybrid neural network and rule-based Inverse Text Normalization (ITN) method for domains containing unique technical phraseology, specifically Air Traffic Control System Command Center (ATCSCC) planning telecon audio transcriptions. The ATCSCC hosts bi-hourly planning telephone conferences (or planning telecons) to ensure smooth operations within the National Airspace (NAS). Access to both live and post meeting transcripts of this speech audio would enable quick review of meetings. Provided speech transcripts, ITN is the process of converting unformatted raw Automated Speaker Recognition (ASR) model transcripts into a human (expert) readable written form. Our hybrid ITN framework utilizes a fine-tuned Bidirectional Encoder Representations from Transformers neural network to format conversational English, and rule-based methods to format domain-specific aviation text. With an overall Punctuation Error Rate (PER) of 25.56 and Word Error Rate with Punctuation and Capitalization (WER PC) of 5.47, we show that this method has vast potential in being applied to ATCSCC planning telecon audio and other audio/text based data available in ATM.

ATM↗

NASA NTRSDate not supplied

Inverse Text Normalization of Air Traffic Control System Command Center Planning Telecon Transcriptions

ATM↗

NASA NTRS2004

Intelligent Control for the BEES Flyer

This paper describes the effort to provide a preliminary capability analysis and a neural network based adaptive flight control system for the JPL-led BEES aircraft project. The BEES flyer was envisioned to be a small, autonomous platform with sensing and control systems mimicking those of biological systems for the purpose of scientific exploration on the surface of Mars. The platform is physically tightly constrained by the necessity of efficient packing within rockets for the trip to Mars. Given the physical constraints, the system is not an ideal configuration for aerodynamics or stability and control. The objectives of this effort are to evaluate the aerodynamics characteristics of the existing design, to make recommendaaons as to potential improvements and to provide a control system that stabilizes the existing aircraft for nominal flight and damaged conditions. Towards this several questions are raised and analyses are presented to arrive at answers to some of the questions raised. CART3D, a high-fidelity inviscid analysis package for conceptual and preliminary aerodynamic design, was used to compute a parametric set of solutions over the expected flight domain. Stability and control derivatives were extracted from the database and integrated with the neural flight control system. The Integrated Vehicle Modeling Environment (IVME) was also used for estimating aircraft geometric, inertial, and aerodynamic characteristics. A generic neural flight control system is used to provide adaptive control without the requirement for extensive gain scheduling or explicit system identification. The neural flight control system uses reference models to specify desired handling qualities in the roll, pitch, and yaw axes, and incorporates both pre-trained and on-line learning neural networks in the inverse model portion of the controller. Results are presented for the BEES aircraft in the subsonic regime for terrestrial and Martian environments.

Krishnakumar, K.↗

NASA NTRS2014

The Effects of Long Duration Head Down Tilt Bed Rest on Neurocognitive Performance: The Effects of Exercise Interventions

We are conducting ongoing experiments in which we are performing structural and functional magnetic resonance brain imaging to identify the relationships between changes in neurocognitive function and neural structural alterations following a six month International Space Station mission and following 70 days exposure to a spaceflight analog, head down tilt bedrest. Our central hypothesis is that measures of brain structure, function, and network integrity will change from pre to post intervention (spaceflight, bedrest). Moreover, we predict that these changes will correlate with indices of cognitive, sensory, and motor function in a neuroanatomically selective fashion. Our interdisciplinary approach utilizes cutting edge neuroimaging techniques and a broad ranging battery of sensory, motor, and cognitive assessments that will be conducted pre flight, during flight, and post flight to investigate potential neuroplastic and maladaptive brain changes in crewmembers following long-duration spaceflight. Success in this endeavor would 1) result in identification of the underlying neural mechanisms and operational risks of spaceflight-induced changes in behavior, and 2) identify whether a return to normative behavioral function following re-adaptation to Earth's gravitational environment is associated with a restitution of brain structure and function or instead is supported by substitution with compensatory brain processes. Our ongoing bed rest participants are also engaging in exercise studies directed by Dr. Lori Ploutz Snyder. In this presentation, I will briefly highlight the existing literature linking exercise and fitness to brain and behavioral functions. I will also overview the metrics from my study that could be investigated in relation to the exercise and control subgroups.

Seidler, R. D.↗

NASA NTRS2014

The Effects of Spaceflight and a Spaceflight Analog on Neurocognitive Perfonnance: Extent, Longevity, and Neural Bases

We are conducting ongoing experiments in which we are performing structural and functional magnetic resonance brain imaging to identify the relationships between changes in neurocognitive function and neural structural alterations following a six month International Space Station mission and following 70 days exposure to a spaceflight analog, head down tilt bedrest. Our central hypothesis is that measures of brain structure, function, and network integrity will change from pre to post intervention (spaceflight, bedrest). Moreover, we predict that these changes will correlate with indices of cognitive, sensory, and motor function in a neuroanatomically selective fashion. Our interdisciplinary approach utilizes cutting edge neuroimaging techniques and a broad ranging battery of sensory, motor, and cognitive assessments that will be conducted pre flight, during flight, and post flight to investigate potential neuroplastic and maladaptive brain changes in crewmembers following long-duration spaceflight. Success in this endeavor would 1) result in identification of the underlying neural mechanisms and operational risks of spaceflight-induced changes in behavior, and 2) identify whether a return to normative behavioral function following re-adaptation to Earth's gravitational environment is associated with a restitution of brain structure and function or instead is supported by substitution with compensatory brain processes. With the bedrest study, we will be able to determine the neural and neurocognitive effects of extended duration unloading, reduced sensory inputs, and increased cephalic fluid distribution. This will enable us to parse out the multiple mechanisms contributing to any spaceflight-induced neural structural and behavioral changes that we observe in the flight study. In this presentation I will discuss preliminary results from six participants who have undergone the bed rest protocol. These individuals show decrements in balance and functional mobility, and alterations in brain structure and function, in association with extended bed rest.

Seidler, R. D.↗

DOE OSTI2025

A deep generative model for deciphering cellular dynamics and in silico drug discovery in complex diseases

Human diseases are characterized by intricate cellular dynamics. Single-cell transcriptomics provides critical insights, yet a persistent gap remains in computational tools for detailed disease progression analysis and targeted in silico drug interventions. Here we introduce UNAGI, a deep generative neural network tailored to analyse time-series single-cell transcriptomic data. This tool captures the complex cellular dynamics underlying disease progression, enhancing drug perturbation modelling and screening. When applied to a dataset from patients with idiopathic pulmonary fibrosis, UNAGI learns disease-informed cell embeddings that sharpen our understanding of disease progression, leading to the identification of potential therapeutic drug candidates. Validation using proteomics reveals the accuracy of UNAGI’s cellular dynamics analysis, and the use of the fibrotic cocktail-treated human precision-cut lung slices confirms UNAGI’s predictions that nifedipine, an antihypertensive drug, may have anti-fibrotic effects on human tissues. UNAGI’s versatility extends to other diseases, including COVID, demonstrating adaptability and confirming its broader applicability in decoding complex cellular dynamics beyond idiopathic pulmonary fibrosis, amplifying its use in the quest for therapeutic solutions across diverse pathological landscapes.

Neural Network↗

DOE OSTI2024

Analyzing inference workloads for spatiotemporal modeling

Ensuring power grid resiliency, forecasting climate conditions, and optimization of transportation infrastructure are some of the many application areas where data is collected in both space and time. Spatiotemporal modeling is about modeling those patterns for forecasting future trends and carrying out critical decision-making by leveraging machine learning/deep learning. Once trained offline, field deployment of trained models for near real-time inference could be challenging because performance can vary significantly depending on the environment, available compute resources and tolerance to ambiguity in results. Users deploying spatiotemporal models for solving complex problems can benefit from analytical studies considering a plethora of system adaptations to understand the associated performance-quality trade-offs. To facilitate the co-design of next-generation hardware architectures for field deployment of trained models, it is critical to characterize the workloads of these deep learning (DL) applications during inference and assess their computational patterns at different levels of the execution stack. In this paper, we develop several variants of deep learning applications that use spatiotemporal data from dynamical systems. We study the associated computational patterns for inference workloads at different levels, considering relevant models (Long short-term Memory, Convolutional Neural Network and Spatio-Temporal Graph Convolution Network), DL frameworks (Tensorflow and PyTorch), precision (FP16, FP32, AMP, INT16 and INT8), inference runtime (ONNX and AI Template), post-training quantization (TensorRT) and platforms (Nvidia DGX A100 and Sambanova SN10 RDU). Overall, our findings indicate that although there is potential in mixed-precision models and post-training quantization for spatiotemporal modeling, extracting efficiency from contemporary GPU systems might be challenging. Instead, co-designing custom accelerators by leveraging optimized High Level Synthesis frameworks (such as SODA High-Level Synthesizer for customized FPGA/ASIC targets) can make workload-specific adjustments to enhance the efficiency.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2025

Solving key challenges in collider physics with foundation models

Foundation models are neural networks that are capable of simultaneously solving many problems. Large language foundation models like ChatGPT have revolutionized many aspects of daily life, but their impact for science is not yet clear. In this paper, we use a new foundation model for hadronic jets to solve three key challenges in collider physics. In particular, we show how experiments can (1) save significant computing power when developing reconstruction algorithms, (2) perform a complete uncertainty quantification for high-dimensional measurements, and (3) search for new physics with model agnostic methods using low-level inputs. In each case, there are significant computational or methodological challenges with current methods that limit the science potential of deep learning algorithms. By solving each problem, we take jet foundation models beyond proof-of-principle studies and into the toolkit of practitioners.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

DOE OSTI2025

Enhancing Drinking Water Quality Modeling: Leveraging Physics Informed Neural Networks for Learning with Imperfect Reaction Models and Partial Data

Chemical kinetics models, typically formulated as systems of ordinary or partial differential equations, are valuable tools for simulating drinking water quality. However, these models often face inaccuracies due to discrepancies between the laboratory and the real-world conditions, as well as limitations in experimental analytical methods, hindering the accurate representation of the true underlying chemical mechanisms. In this study, we propose a Physics Informed Neural Network (PINN), using the eXtreme Theory of Functional Connections, to improve the prediction of chemical concentrations over time. The PINN method accounts for imperfect chemical models and incorporates partial data to improve predictions. Focusing on reactions describing water disinfection residual and disinfectant byproduct formation, which are crucial for public health and regulatory compliance, we demonstrate that the PINN model is able to accurately predict the concentrations of chemical species across various pH values. Notably, the model extends its accuracy to predict concentrations of chemical species not originally included in its training data. The developed method can be extended to a variety of chemical systems, offering a wide array of potential applications.

13 HYDRO ENERGY↗

DOE OSTI2024

Deep operator network surrogate for phase-field modeling of metal grain growth during solidification

A deep operator network (DeepONet) has been constructed that generates accurate representations of phase-field model simulations for evolving two dimensional metal grain morphology growing from melt. These representations serve as lower resolution, computationally efficient stand-ins for quick parameter space exploration of solutions to the the Allen-Cahn equations that dictate the phase-field model simulations. The experimental target for the phase-field model is a uranium casting system cooling a 434 g uranium charge from a maximum temperature of 1400° C at an average rate of 30° C / min , traversing the crystallographic phases of the pure metal. Experimental parameters inform the phase-field model, whose higher resolution computational model solutions are used to train the DeepONet in a given parameter space with the aim of developing a faster, more efficient method for predicting the solidifying metal's microstructure at different potential experimental values. The final DeepONet generates high accuracy, lower resolution predictions with cumulative relative approximation error over all timesteps of less than 0.5%, while ensuring solutions remain within physically feasible ranges. Further, these relative error values are comparable with other state-of-the-art DeepONet models for microstructure evolution, while significantly reducing the amount of training data required. Training a convolutional neural network simultaneously with the DeepONet, enforcing realistic values at the complex metal grain boundaries, and mathematically encoding boundary conditions into the structure of the DeepONet improved prediction accuracy and computational efficiency over a standard DeepONet model.

36 MATERIALS SCIENCE↗

NASA NTRS2023

Use of Machine Learning and Principal Component Analysis to Retrieve Nitrogen Dioxide (NO 2 ) With Hyperspectral Imagers and Reduce Noise in Spectral Fitting

Nitrogen dioxide (NO 2 ) is an important trace-gas pollutant and climate agent whose presence also leads to spectral interference in ocean color retrievals. NO 2 column densities have been retrieved with satellite UV–Vis spectrometers such as the Ozone Monitoring Instrument (OMI) and the Tropospheric Monitoring Instrument (TROPOMI) that typically have spectral resolutions of the order of 0.5 nm or better and spatial footprints as small as 3.6 km × 5.6 km. These NO 2 observations are used to estimate emissions, monitor pollution trends, and study effects on human health. Here, we investigate whether it is possible to retrieve NO 2 amounts with lower-spectral-resolution hyperspectral imagers such as the Ocean Color Instrument (OCI) that will fly on the Plankton, Aerosol, Cloud, ocean Ecosystem (PACE) satellite set for launch in early 2024. OCI will have a spectral resolution of 5 nm and a spatial resolution of ∼ 1 km with global coverage in 1–2 d. At this spectral resolution, small-scale spectral structure from NO 2 absorption is still present. We use real spectra from the OMI to simulate OCI spectra that are in turn used to estimate NO 2 slant column densities (SCDs) with an artificial neural network (NN) trained on target OMI retrievals. While we obtain good results with no noise added to the OCI simulated spectra, we find that the expected instrumental noise substantially degrades the OCI NO 2 retrievals. Nevertheless, the NO 2 information from OCI may be of value for ocean color retrievals. OCI retrievals can also be temporally averaged over timescales of the order of months to reduce noise and provide higher-spatial-resolution maps that may be useful for downscaling lower-spatial-resolution data provided by instruments such as OMI and TROPOMI; this downscaling could potentially enable higher-resolution emissions estimates and be useful for other applications. In addition, we show that NNs that use coefficients of leading modes of a principal component analysis of radiance spectra as inputs appear to enable noise reduction in NO 2 retrievals. Once trained, NNs can also substantially speed up NO 2 spectral fitting algorithms as applied to OMI, TROPOMI, and similar instruments that are flying or will soon fly in geostationary orbit.

NO2↗

DOE OSTI2025

wa-hls4ml and lui-gnn: A benchmark and GNN-based surrogate model for hls4ml resource and latency estimation

As machine learning (ML) increasingly serves as a tool for addressing real-time challenges in scientific applications, the development of advanced tooling has significantly reduced the time required to iterate on various designs. These advancements have solved major obstacles, but also exposed new challenges. For example, processes that were not previously considered bottlenecks, such as model synthesis, are now becoming limiting factors in the rapid iteration of designs. To reduce these emerging constraints, multiple efforts are being launched toward designing an ML-based surrogate model that estimates resource usage of synthesized accelerator architectures. This model would reduce the design iteration time, especially when designing within a set of given hardware constraints. This approach shows considerable potential, but as it stands, the effort is early and would benefit from coordination and standardization to assist future work as it emerges. We introduce wa-hls4ml, a benchmark for ML accelerator resource and latency estimation, and its corresponding initial dataset of more than 100,000 fully connected neural networks, all synthesized using hls4ml and targeting Xilinx FPGAs. In addition to the resource utilization and latency data provided, the dataset includes generated artifacts and log files for many of the synthesized neural networks, in order to support future research in ML-based code generation. The benchmark evaluates the performance of resource and latency predictors against several common ML model architectures, primarily originating from scientific domains, as exemplar models, as well as the average performance across a subset of the dataset. We measure the performance of a given predictor model through multiple metrics, including $R^2$ score and SMAPE on regression tasks, as well as inference time to further characterize the estimator under test. Additionally, we introduce the latency/utilization inference graph neural network (lui-gnn), a surrogate model that uses a graph neural network to represent input architectures in the form of a directed graph. This graph representation allows for a diverse set of model architectures to all be effectively handled by a surrogate model. We present the architecture and performance of the model, as evaluated by the new proposed benchmark, including SMAPE, $R^2$ score, and inference times, and find that lui-gnn generally predicts latency and utilization for the 75\% quantile within several percent of the synthesized resources on the synthetic test dataset, indicating that this approach of estimating resource and latency via a surrogate models has promise and warrants further research.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗