Search NASA⌕ Search

SEARCH · Search NASA

Results for “model accuracy”

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 127 records · Page 7
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Coupling Remote Sensing With a Process Model for the Simulation of Rangeland Carbon Dynamics

Rangelands provide significant environmental benefits through many ecosystem services, which may include soil organic carbon (SOC) sequestration. However, quantifying SOC stocks and monitoring carbon (C) fluxes in rangelands are challenging due to the considerable spatial and temporal variability tied to rangeland C dynamics as well as limited data availability. We developed the Rangeland Carbon Tracking and Management (RCTM) system to track long-term changes in SOC and ecosystem C fluxes by leveraging remote sensing inputs and environmental variable data sets with algorithms representing terrestrial C-cycle processes. Bayesian calibration was conducted using quality-controlled C flux data sets obtained from 61 Ameriflux and NEON flux tower sites from Western and Midwestern US rangelands to parameterize the model according to dominant vegetation classes (perennial and/or annual grass, grass-shrub mixture, and grass-tree mixture). The resulting RCTM system produced higher model accuracy for estimating annual cumulative gross primary productivity (GPP) (R 2 > 0.6, RMSE <390 g C m -2 ) relative to net ecosystem exchange of CO 2 (NEE) (R 2 > 0.4, RMSE <180 g C m -2 ). Model performance in estimating rangeland C fluxes varied by season and vegetation type. The RCTM captured the spatial variability of SOC stocks with R 2 = 0.6 when validated against SOC measurements across 13 NEON sites. Model simulations indicated slightly enhanced SOC stocks for the flux tower sites during the past decade, which is mainly driven by an increase in precipitation. Future efforts to refine the RCTM system will benefit from long-term network-based monitoring of vegetation biomass, C fluxes, and SOC stocks.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2022

Analysis of Reactive Power Load Modeling Techniques for PV Impact Studies [Slides]

The increasing availability of advanced metering infrastructure (AMI) data has led to significant improvements in load modeling accuracy. However, since many AMI devices were installed to facilitate billing practices, few utilities record or store reactive power demand measurements from their AMI. When reactive power measurements are unavailable, simplifying assumptions are often applied for load modeling purposes, such as applying constant power factors to the loads. The objective of this work is to quantify the impact that reactive power load modeling practices can have on distribution system analysis, with a particular focus on evaluating the behaviors of distributed photovoltaic (PV) systems with advanced inverter capabilities. Quasi-static time-series simulations were conducted after applying a variety of reactive power load modeling approaches, and the results were compared to a baseline scenario in which real and reactive power measurements were available at all customer locations on the circuit. Overall, it was observed that applying constant power factors to loads can lead to significant errors when evaluating customer voltage profiles, but that performing per-phase time-series reactive power allocation can be utilized to reduce these errors by about 6x, on average, resulting in more accurate evaluations of advanced inverter functions.

14 SOLAR ENERGY↗
DOE OSTI2025

Comparative Analysis of TCR and TCR-pMHC Complex Structure Prediction Tools

The rapid development of computational approaches for predicting the structures of T cell receptors (TCRs) and TCR-peptide-major histocompatibility (TCR-pMHC) complexes, accelerated by AI breakthroughs such as AlphaFold, has made it feasible to calculate these structures with increasing accuracy. Although these tools show great potential, their relative accuracy and limitations remain unclear due to the lack of standardized benchmarks. Here, we systematically evaluate seven tools for predicting isolated TCR structures together with six tools for predicting TCR-pMHC complex structures. The methods include homology-based approaches, general prediction tools using AlphaFold, TCR-specific tools derived from AlphaFold2, and the newly developed tFold-TCR model. The evaluation uses a post-training data set comprising 40 αβ TCRs and 27 TCR-pMHC complexes (21 Class I and 6 Class II). Model accuracy is assessed at global, local, and interface levels using a variety of metrics. We find that each tool offers distinct advantages in various aspects of its predictions. AlphaFold2, AlphaFold3, and tFold-TCR excel in overall accuracy of TCR structure prediction, and TCRmodel2 and AlphaFold2 perform well in overall accuracy of TCR-pMHC structure prediction. However, TCR-specific tools derived from AlphaFold2 show lower accuracy in the framework region than both homology-based methods and general-purpose tools such as AlphaFold, and challenges remain for all in modeling CDR3 loops, docking orientations, TCR-peptide interfaces, and Class II MHC-peptide interfaces. Furthermore, these findings will guide researchers in selecting appropriate tools, emphasize the importance of using multiple evaluation metrics to assess model performance, and offer suggestions for improving TCR and TCR-pMHC structure prediction tools.

Chemical structure↗
DOE OSTI2024

Boundary-Aware Adversarial Learning Domain Adaption and Active Learning for Cross-Sensor Building Extraction

The use of convolutional neural networks (CNNs) for building extraction from remote sensing images has been widely studied and many public datasets have been made available for accelerating development of these CNN models. Yet adapting pretrained models at scale in real-world scenarios remains a challenging task. The main barrier is that certain new labels are still needed to compensate for domain shifting between the labeled data and new images that potentially cover new geographic locations or that are from a different sensor. In this article, we propose to add informatively labeled samples from a new image pool under the paradigm of active learning. To select the most useful samples based on model uncertainty, we first tackle the problem of uncalibrated uncertainty estimation due to distribution shifting by adapting feature extractors with boundary-based adversarial learning. Calibrated uncertainty is used as the query criterion in the active learning process, where the most uncertain samples are selected for annotation and included for model retraining. The proposed workflow was tested with three data pairs in which each workflow represents a scenario often encountered in real-world applications, including adapting pretrained models to new images collected with different sensors or to new geographic areas where appearances and types of buildings are very different. Compared to several baselines, including random sampling, temperature scaling (a well-known uncertainty calibration technique), different query strategies, and active domain adaptation methods, the proposed workflow shows that strategically querying a smaller set of samples for labeling achieves comparable or better building extraction performance. The proposed method reduces the number of labeled samples required to achieve sufficient model accuracy, thus significantly reducing hundreds of person-hours for labeled data creation. In addition, we include a few considerations when deploying this workflow in a GPU cluster that can be easily adapted to achieve operational building extraction model retraining.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

Leveraging Community and Author Context to Explain the Performance and Bias of Text-Based Deception Detection Models

Deceptive news posts shared in online communities can be detected with NLP models, and much recent research has focused on the development of such models. In this work, we use characteristics of online communities and authors --- the context of how and where content is posted --- to explain the performance of a neural network deception detection model and identify sub-populations who are disproportionately affected by model accuracy or failure. We examine who is posting the content, and where the content is posted to. We find that while author characteristics are better predictors of deceptive content than community characteristics, both characteristics are strongly correlated with model performance. Traditional performance metrics such as F1 score may fail to capture poor model performance on isolated sub-populations such as specific authors, and as such, more nuanced evaluation of deception detection models is critical.

machine learning (ML), machine learning explanatio↗
DOE OSTI2022

Model Choice Metrics to Optimize Profile-QSAR Performance

Predicting molecular activity against protein targets is difficult because of the paucity of experimental data. Approaches like multitask modeling and collaborative filtering seek to improve model accuracy by leveraging results from multiple targets, but are limited because different compounds are measured with different assays, leading to sparse data matrices. Profile-QSAR (pQSAR) 2.0 addresses this problem by fitting a series of partial least squares models for each target, using as features the predictions from single-task models on the remaining targets. Here, this method has been shown to produce better results than single task and multitask models. However, the factors determining the success of pQSAR 2.0 have as yet not been characterized. In this paper we examine the experimental conditions that lead to better pQSAR models. We limit the amount of data available to the method by retraining with decreasing amounts of data and explore the model’s ability to generalize to compounds that have never been assayed. Finally, we look at the properties of training data needed to demonstrate pQSAR improvement.

Biological and medical sciences, Computer science↗
DOE OSTI2021

Multidimensional Numerical Modeling of Combustion Dynamics in a Non-Premixed Rotating Detonation Engine With Adaptive Mesh Refinement

In the present work, a novel computational fluid dynamics (CFD) methodology was developed to simulate full-scale non-premixed rotating detonation engines (RDEs). A unique feature of the modeling approach was the incorporation of adaptive mesh refinement (AMR) to achieve a good trade-off between model accuracy and computational expense. Here, unsteady Reynolds-averaged Navier–Stokes (RANS) simulations were performed for an Air Force Research Laboratory (AFRL) non-premixed RDE configuration with hydrogen as fuel and air as the oxidizer. The finite-rate chemistry model, along with a ten-species detailed kinetic mechanism, was employed to describe the H 2 -Air combustion chemistry. Three distinct operating conditions were simulated, corresponding to the same global equivalence ratio of unity but different fuel/air mass flowrates. For all conditions, the capability of the model to capture essential detonation wave dynamics was assessed. An exhaustive verification and validation study was performed against experimental data in terms of a number of waves, wave frequency, wave height, reactant fill height, oblique shock angle, axial pressure distribution in the channel, and fuel/air plenum pressure. The CFD model was demonstrated to accurately predict the sensitivity of these wave characteristics to the operating conditions, both qualitatively and quantitatively. A comprehensive heat release analysis was also conducted to quantify detonative versus deflagrative burning for the three simulated cases. The present CFD model offers a potential capability to perform rapid design space exploration and/or performance optimization studies for realistic full-scale RDE configurations.

42 ENGINEERING↗
DOE OSTI2023

Multiphysics modeling of printed surface acoustic wave thermometers

Surface acoustic wave (SAW) transducers are a class of sensors and actuators that operate on the fundamental principle of piezoelectricity. Aerosol jet printing and other additive manufacturing techniques have resulted in the low-cost fabrication of low-power and small-footprint SAW devices that are suitable for sensing in high-temperature and radioactive environments. In this work, we developed a series of temperature-dependent finite element models for a SAW transducer consisting of printed silver interdigitated transducers (IDTs) deposited onto piezoelectric lithium niobate. Modeling accuracy was evaluated experimentally from room temperature to 200 °C using an aerosol-jet-printed SAW thermometer. A time-domain study enabled visualization of the wave propagation and successfully guided the denoising of the scattering parameter measurement. Additionally, frequency-domain models using traditional modal analysis or the unique port boundary condition feature in COMSOL Multiphysics accurately predicted the temperature-driven natural frequency drift in the SAW thermometer. The finite element models developed in this study serve to facilitate the computer-aided design of future SAW transducers for applications in harsh environments.

47 OTHER INSTRUMENTATION↗
DOE OSTI2024

Low-Cost Heliostat for High-Flux Small-Area Receivers (Final Technical Report)

This project analyzed a two-stage heliostat concept consisting of a tracking stage and a concentrating stage. The tracking stage uses mirrors mounted on a common drive that move to track the sun. The concentrating stage consists of stationary mirrors that each have a unique angle to direct rays towards a small-area, high-flux, point-focused receiver. By splitting the collection and concentrating process into two stages, multiple small, inexpensive mirrors can share a structure and be controlled by a single drive in the tracking stage. The project effort developed modeling techniques that were specifically relevant to this two-stage heliostat concept. Both field-level and unit-level models were developed. The field-level model does not explicitly consider unit-level losses which are predicted by the unit-level model and then integrated into the field-level model through a correlation referred to as an efficiency modifier. This approach is referred to as the two-model approach; the development and demonstration of this two-model approach for a multi-stage heliostat technology is a key outcome of this work. The field-level model is used to design a field that hits a specific design day power given a set of heliostat design parameters. An oversized field is simulated and then heliostat units are removed based on their annual energy production in order to generate the highest performing field. The field reduction procedure fits a smooth curve fit to annual energy production as a function of position in the field which has the effect of reducing the noise that is otherwise caused by the Monte Carlo ray tracing technique. This approach is referred to as the annual energy fit method and substantially reduces computational run time for a given field level modeling accuracy. The annual energy fit approach enables the selection of a properly sized, high-performing field using orders of magnitude fewer rays than would otherwise be possible and the development of this approach is a second key outcome of this work. These models are used within a genetic optimization algorithm in order to optimize the geometric parameters associated with a heliostat in order to achieve the lowest cost per unit of collected design day power. The cost modeling that underlies the optimization is a simple, scaling type analysis backed up by a much more detailed Design for Manufacture and Assembly (DFMA) analysis. Although the figure of merit used for optimization was not cost per mirror area, this metric is reasonable to use as a means of comparison. The optimally designed 500 kW design has a tracking mirror specific cost of $181.85/m 2 , which is significantly larger than the target value and also larger than the current state of the art. The cost of the torque-tube type linkages contributed substantially to the overall cost. Based on this observation, potentially attractive alternative design configuration utilizing a capstan type actuation system should be investigated. Finally, NREL compared the performance of the two-stage heliostat to the performance of a focused and different sized flat conventional heliostats and showed that, as expected, additional losses versus the convention heliostat caused by a worse cosine efficiency, two stages of reflection, and interstage interactions. The two-stage heliostat requires around 75% more reflective area than a flat 1x1 meter conventional heliostat (similar to a focused heliostat) and 40% more than a flat 2x2 meter conventional heliostat.

14 SOLAR ENERGY↗
DOE OSTI2023

Seismoacoustic Explosion Yield and Depth Estimation: Insights from the Large Surface Explosion Coupling Experiment

The Large Surface Explosion Coupling Experiment (LSECE) is a chemical explosion experiment conducted in Yucca Flat at the Nevada National Security Site in 2020. The experiment included two surface detonations of ~1000 kg trinitrotoluene equivalent. The main goal of this experiment was to provide the ground-truth data for seismoacoustic wave excitation by large chemical explosions near the ground surface. The seismic and acoustic energy partitioning between the surface is strongly governed by the depth or height of explosions, and either seismic or acoustic-only analysis may have inherent ambiguity in determining explosion yield and depth simultaneously. Previous studies suggested that joint seismoacoustic analysis can resolve the trade-off and reduce the uncertainty of yield and depth estimation dramatically. Here we demonstrate the capability of seismoacoustic analysis to improve the accuracy of explosion yield and depth estimation with the LSECE data. Local acoustic wave propagation in the atmosphere can be substantially affected by constantly varying weather conditions. Consisting of two detonations before dawn and in the afternoon, LSECE provides unique data to evaluate the model accuracy of acoustic wave propagation and seismoacoustic energy partitioning depending on local atmospheric conditions. We quantitatively evaluate the accuracy of yield and depth estimation depending on atmospheric variability and the improvement achieved by the joint seismoacoustic approach.

58 GEOSCIENCES↗
DOE OSTI2025

The Effects of Compounded Model Size Reductions on Adversarial Robustness

Recent advances in Edge AI and Tiny Machine Learning (TinyML) have enabled the deployment of machine learning models on resource-constrained environments. However, deploying these models on edge devices, such as micro-controllers, requires significant model footprint reduction through a variety of techniques such as quantization, pruning, and clustering. While these optimization methods offer considerable advantages, they potentially introduce AI-related security vulnerabilities, particularly concerning model robustness with respect to adversarial AI attacks. Prior research has extensively examined the impact of quantization on adversarial robustness; however, the effects of alternative reduction techniques and their combinations remain understudied. This paper investigates the impact of model size reduction techniques on adversarial robustness, when applied individually and combined. We utilized Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks to generate adversarial perturbations for both training and testing data, and then evaluated the models' accuracy under adversarial training conditions. Our findings revealed that reduction techniques generally diminished robustness; although, combining techniques was not found to make robustness any worse than when applied individually. Moreover, specific techniques can potentially enhance resistance to small size perturbations. This research provides insights into the trade-offs between model size reduction and security, establishing a foundation for future investigations into improving adversarial training techniques and methodologies for maintaining robustness while preserving memory footprint benefits.

Austria, Phillipe [ORNL] (ORCID:0000000236223973)↗
DOE OSTI2025

EcoBOT: an AI/ML enabled automated phenotyping capability for model plants

Introduction: Advances in automation and AI/ML offer new opportunities for plant science, including design, modeling, and analysis. This study aimed to develop an automated platform for researching small model plants under axenic conditions and integrate it with AI/ML tools. Methods: The EcoBOT platform was developed, which consists of sterile containers (EcoFABs) for growing plants and imaging for monitoring plant growth and health. Brachypodium distachyon was grown on the EcoBOT, and its response to nutrient limitation and copper stress was evaluated. Results: The results showed that Brachypodium distachyon grown in the EcoBOT maintained sterility and responded to nutrient limitation and copper stress. Analysis of over 6,500 root and shoot images revealed varying sensitivity and response rates to copper. Bayesian Optimization was used to improve model accuracies relating copper concentrations to plant biomass via sequential experiments, resulting in a >30% improvement. Discussion: The findings of this study demonstrate the potential of the EcoBOT platform for researching plant responses to environmental factors. Future experiments could focus on relating other chemical stresses and microbial interactions to create generalized models of plant responses.

AI image analysis↗
DOE OSTI2022

Graph-based featurization methods for classifying small molecule compounds

For over a decade, drug-induced liver injury (DILI) has posed significant drawbacks in the synthesis and development of drugs and remains a consequential concern. With finite success within the existing preclinical models, DILI is one of the main causes of drug withdrawal or termination from the market. Particularly, this withdrawal occurs during the late stages of drug development (Kullak-Ublick, 2017). Since DILI is difficult to diagnose and treat, it has become an obstacle in the drug production market that in turn affects clinicians, pharmaceutical companies, and consumers. We propose a method for learning features of DILI-positive drugs based on the graphical relationships and patterns they possess within a network of biological databases. We also train various statistical and machine learning models on these learned features in order to classify the drugs as DILI-positive or negative. Our methods include Random Forest, Neural networks, and logistic regression classification. We utilize labeled DILI-positive and DILI-negative datasets, which were developed by the FDA and the National center for toxicological research, as well as additional literature datasets (Thakkar, 2020) in order to validate our results and assess our featurization and model accuracy.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2024

Development and cross‐validation of a circumference‐based predictive equation to estimate body fat in an active population

Abstract Objective The U.S. Army uses sex‐specific circumference‐based prediction equations to estimate percent body fat (%BF) to evaluate adherence to body composition standards. The equations are periodically evaluated to ensure that they continue to accurately assess %BF in a diverse population. The objective of this study was to develop and validate alternative field expedient equations that may improve upon the current Army Regulation (AR) body fat (%BF) equations. Methods Body size and composition were evaluated in a representatively sampled cohort of 1904 active‐duty Soldiers (1261 Males, 643 Females), using dual‐energy X‐ray absorptiometry (%BF DXA ), and circumferences obtained with 3D imaging and manual measurements. Sex stratified linear prediction equations for %BF were constructed using internal cross validation with %BF DXA as the criterion measure. Prediction equations were evaluated for accuracy and precision using root mean squared error, bias, and intraclass correlations. Equations were externally validated in a convenient sample of 1073 Soldiers. Results Three new equations were developed using one to three circumference sites. The predictive values of waist, abdomen, hip circumference, weight and height were evaluated. Changing from a 3‐site model to a 1‐site model had minimal impact on measurements of model accuracy and performance. Male‐specific equations demonstrated larger gains in accuracy, whereas female‐specific equations resulted in minor improvements in accuracy compared to existing AR equations. Equations performed similarly in the second external validation cohort. Conclusions The equations developed improved upon the current AR equation while demonstrating robust and consistent results within an external population. The 1‐site waist circumference‐based equation utilized the abdominal measurement, which aligns with associated obesity related health outcomes. This could be used to identify individuals at risk for negative health outcomes for earlier intervention.

Taylor, Kathryn M.↗
DOE OSTI2023

Lightweight and effective tensor sensitivity for atomistic neural networks

Atomistic machine learning focuses on the creation of models that obey fundamental symmetries of atomistic configurations, such as permutation, translation, and rotation invariances. In many of these schemes, translation and rotation invariance are achieved by building on scalar invariants, e.g., distances between atom pairs. There is growing interest in molecular representations that work internally with higher rank rotational tensors, e.g., vector displacements between atoms, and tensor products thereof. Here, we present a framework for extending the Hierarchically Interacting Particle Neural Network (HIP-NN) with Tensor Sensitivity information (HIP-NN-TS) from each local atomic environment. Crucially, the method employs a weight tying strategy that allows direct incorporation of many-body information while adding very few model parameters. We show that HIP-NN-TS is more accurate than HIP-NN, with negligible increase in parameter count, for several datasets and network sizes. As the dataset becomes more complex, tensor sensitivities provide greater improvements to model accuracy. In particular, HIP-NN-TS achieves a record mean absolute error of 0.927 $\frac{\text{kcal}}{\text{mol}}$ for conformational energy variation on the challenging COMP6 benchmark, which includes a broad set of organic molecules. We also compare the computational performance of HIP-NN-TS to HIP-NN and other models in the literature.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2025

FedEFsz: Fair Cross-Silo Federated Learning System With Error-Bounded Lossy Compression

Cross-Silo federated learning systems have been identified as an efficient approach to scaling DNN training across geographically-distributed data silos to preserve the privacy of the training data. Communication efficiency and fairness are two major issues that need to be both satisfied when federated learning systems are deployed in practice. Simultaneously guaranteeing both of them, however, is exceptionally difficult because simply combining communication reduction and fairness optimization approaches often causes non-converged training or drastic accuracy degradation. Here, to bridge this gap, we propose FedEFsz. On the one hand, it integrates the state-of-the-art error-bounded lossy compressor SZ3 into cross-silo federated learning systems to significantly reduce communication traffic during the training. On the other hand, it achieves a high fairness (i.e., rather consistent model accuracy and performance across different clients) through a carefully designed heuristic algorithm that can tune the error-bound of SZ3 for different clients during the training. Extensive experimental results based on a GPU cluster with 65 GPU cards show that FedEFsz improves the fairness across different benchmarks by up to 60.88% and meanwhile reduces the communication traffic by up to 315×.

Cross-Silo Federated Learning Systems↗
DOE OSTI2026

ELG×LRG Distribution through Dark Matter Halo Dynamics

We investigate the clustering and halo occupation distribution (HOD) of DESI Y1 emission-line (ELGs) and luminous red (LRGs) galaxies at 0.8 < z < 1.1, including their cross-correlation (ELG×LRG), using the A BACUS S UMMIT suite and a new Halo Occupation Model (H OME ) for galaxy multitracers. This integrates intrahalo dynamics, halo exclusion, and quenching, bridging insights from hydrodynamical, HOD, abundance-matching, and semianalytic studies. Leveraging full phase-space information from the Uchuu N-body simulation, and sampling satellites from dark-matter particle positions via physically motivated prescriptions, Home reproduces the anisotropic clustering down to s = 200 h −1 kpc with unprecedented accuracy. Model parameters are inferred solely from two-point statistics using a two-level Bayesian framework, yielding high-fidelity ELG, LRG, and cross-reference catalogs. We find that satellite ELGs behave as incoherent flows within their parent halos, dominating the clustering below 4 h −1 Mpc. The HOD from the best-fit Home has the following properties: (i) 90.50% (85.91%) of ELGs (LRGs) are central galaxies without satellites, residing in halos of M vir ∼ 6.6 × 10 11 (1.2 × 10 13 ) h −1 M ⊙ ; (ii) the ELG×LRG cross-correlation is governed by central-central pairs and shaped by halo exclusion on 2–5 h −1 Mpc scales; (iii) 9.50% (14.09%) of ELGs (LRGs) are satellites, of which 1.09% (3.52%) inhabit halos with a central galaxy of the same species in a maximally conformal configuration, 7.02% (0.005%) orbit complementary hosts in a minimally conformal state, and 0.58% (10.57%) are orphans. The high sensitivity of Home precisely captures the dynamics of satellites in different host environments, opening a promising avenue for understanding systematics and the dynamical nature of dark matter, potentially distinguishing gravity models.

Favole, Ginevra [Universidad de La Laguna (Spain);↗
DOE OSTI2022

Throughput-Oriented and Accuracy-Aware DNN Training with BFloat16 on GPU

Deep Neural Networks (DNNs) have transformed the field of artificial intelligence and achieved extraordinary success in many areas. The training of DNNs is commonly compute and memory-intensive, which has resulted in several optimizations in the training phase. Among them, reduced precision is a typical and widely used technique to accelerate DNN training and reduce memory requirements. However, applying a widely adopted reduced precision format such as Float16 to all involved operations in DNN training is not optimal as the use of Float16 in some operations can hurt model accuracy. Meanwhile, additional optimizations including loss scaling and autocast techniques can mitigate the accuracy loss but lead to inherent overhead and inadequate use of reduced precision. In this work, we leverage another reduced precision format, BFloat16, and introduce a throughput-oriented and accuracy-aware approach to maximize the performance potential of DNN training. Since the high throughput provided by BFloat16 format is accompanied by low precision of the floating-point representation, this approach achieves high throughput by using BFloat16 on all DNN operations and avoids the accuracy loss through a customized accuracy-aware normalization. Results show that our approach outperforms the state-of-the-art mixed-precision training by 1.21x on an NVIDIA A100 GPU.

Xie, Zhen↗