SEARCH · Engineering Papers

Results for “robust statistics”

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 127 records · Page 7

All sources NASA NTRS DOE OSTI

DOE OSTI2025

The DECADE cosmic shear project IV: cosmological constraints from 107 million galaxies across 5,400 deg$^2$ of the sky

We present cosmological constraints from the Dark Energy Camera All Data Everywhere (DECADE) cosmic shear analysis. This work uses shape measurements for 107 million galaxies measured through Dark Energy Camera (DECam) imaging of $5,\!412$ deg$^2$ of sky that is outside the Dark Energy Survey (DES) footprint. We derive constraints on the cosmological parameters $S_8 = 0.791^{+0.027}_{-0.032}$ and $Ω_{\rm m} =0.269^{+0.034}_{-0.050}$ for the $Λ$CDM model, which are consistent with those from other weak lensing surveys and from the cosmic microwave background. We combine our results with cosmic shear results from DES Y3 at the likelihood level, since the two datasets span independent areas on the sky. The combined measurements, which cover $\approx\! 10,\!000$ deg$^2$, prefer $S_8 = 0.791 \pm 0.023$ and $Ω_{\rm m} = 0.277^{+0.034}_{-0.046}$ under the $Λ$CDM model. These results are the culmination of a series of rigorous studies that characterize and validate the DECADE dataset and the associated analysis methodologies (Anbajagane et. al 2025a,b,c). Overall, the DECADE project demonstrates that the cosmic shear analysis methods employed in Stage-III weak lensing surveys can provide robust cosmological constraints for fairly inhomogeneous datasets. This opens the possibility of using data that have been previously categorized as ``unusable'' for cosmic shear analyses, thereby increasing the statistical power of upcoming weak lensing surveys.

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2024

Evaluating downscaled products with expected hydroclimatic co-variances

Abstract. There has been widespread adoption of downscaled products amongst practitioners and stakeholders to ascertain risk from climate hazards at the local scale (e.g., ∼ 5 km resolution). Such products must nevertheless be consistent with physical laws to be credible and of value to users. Here we evaluate statistically and dynamically downscaled products by examining local co-evolution of downscaled temperature and precipitation during convective and frontal precipitation events (two mechanisms testable with just temperature and precipitation). We find that two widely used statistical downscaling techniques (Localized Constructed Analogs version 2, LOCA2, and Seasonal Trends and Analysis of Residuals Empirical Statistical Downscaling Model, STAR-ESDM) generally preserve expected co-variances during convective precipitation events over the historical and future projected intervals as compared to European Centre for Medium-Range Weather Forecasts Reanalysis v5 (ERA5) and two observation-based data products (Livneh and nClimGrid-Daily). However, both techniques dampen future intensification of frontal precipitation that is otherwise robustly captured in global climate models (i.e., prior to downscaling) and with process-based dynamical downscaling across five different regional climate models. In the case of LOCA2, this leads to appreciable underestimation of future frontal precipitation event intensity. This study is one of the first to quantify a likely ramification of the stationarity assumption underlying statistical downscaling methods and identify a phenomenon where projections of future change diverge depending on data production method employed. Finally, our work proposes expected co-variances during convective and frontal precipitation as useful evaluation diagnostics that can be universally applied to a wide range of statistically downscaled products.

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2026

Machine learning of factors for improving oyster hatchery production

Oyster aquaculture and restoration in the Chesapeake Bay are vital, yet hatcheries frequently struggle with inconsistent larval growth and sudden mass mortality events. Unpredictable disruptions in larval production cause large economic losses, represent a perceived risk to growers, and impede industry expansion. To better understand associations between production yield and its potential predictors, we applied machine learning (random forest, and neural network) and statistical (generalized additive model) models to a comprehensive dataset of environmental, water quality, and operational parameters from a Maryland oyster hatchery, aiming to identify key yield predictors and develop a robust forecasting tool. We used recursive Boruta algorithm for variable selection, pinpointing critical predictors, and employed cross-validation to fine-tune model settings. Shapley value analysis offered crucial insights into model interpretations, highlighting week number, Normalized Difference Vegetation Index, salinity, turbidity, and fecundity as primary drivers of yield variability. For low-yield cases, salinity-related variables were particularly important. Our findings provide an early warning system for potential production downturns, empowering hatchery operators to make data-driven decisions for optimizing water conditions, feeding schedules, and broodstock management. By boosting predictability and efficiency, this research directly supports economic stability of the oyster industry and ecological health of the Chesapeake Bay.

Vishwakarma, Srishti [Oak Ridge National Laborator↗

DOE OSTI2025

Robust Dark Energy Constraints with the Dark Energy Spectroscopic Survey (Final Technical Report)

This project developed and applied advanced theoretical, computational, and data-analysis methodologies to extract robust and precise cosmological constraints from the Dark Energy Spectroscopic Instrument (DESI). The work focused on maximizing the scientific return of DESI through optimized survey strategy, novel higher-order clustering statistics, improved modeling of small-scale structure, and rigorous mitigation of observational systematics. Over the award period, the project made substantial contributions to DESI science planning, produced new methods for bispectrum and three-point correlation function analyses, advanced constraints on primordial non-Gaussianity, and delivered widely used software tools. The project also played a major role in training graduate students and a postdoctoral researcher who contributed directly to DESI key projects. The results have significantly enhanced the cosmological reach of DESI and provide a strong foundation for future surveys such as DESI-II and Stage-V experiments.

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2025

CMB low multipole alignments across WMAP and Planck data releases

ABSTRACT The first observations of the cosmic microwave background (CMB) from NASA's Wilkinson Microwave Anisotropy Probe (WMAP) led to finding ‘alignment’ anomalies not expected from fluctuations in the isotropic cosmological model. We study the data of all 8 full-sky public releases since then to test for anomalous alignments and shapes of the first 60 multipoles, i.e. over the range $2\le l \le 61$. We use rotationally invariant and covariant statistics to test isotropy of all subsequent WMAP data releases, along with those from the ESA’s Planck mission. Anomalous alignments among the multipoles $l=1, 2, 3$ are very consistent and robust. More alignments are detected, some of them new, while significance is diluted by the large range of the search. Power entropy, a measure of the randomness of the multipoles, is consistently anomalous at about $2\sigma$ level or better across all data releases. It appears that the CMB is not as random as the cosmological principle predicts on large angular scales.

Patel, Sanjeet Kumar↗

DOE OSTI2025

PV Fleet Performance Data Initiative Final Technical Report (FTR)

Improved analysis and reporting of photovoltaic (PV) field performance increases the certainty of owners and financiers that systems will perform as expected. Advanced module technologies (e.g., PERC, HJT, and bifacial) introduce new degradation mechanisms and performance characteristics. This project will leverage data from the ever-increasing PV fleet to develop models and understanding of the field performance of existing and new technologies. Please see our list of public reports at https://www.nrel.gov/pv/fleet-performance-data-initiative.html. Objective 1: Support the global PV industry with scalable, robust data analysis tools that reduce the uncertainty of PV system performance and loss calculation. Objective 2: Reduce perceived risk arising from degradation rate, soiling loss, and system availability by publishing detailed statistics on U.S. fleet performance. Objective 3: Highlight factors leading to system underperformance including module type, climate, mounting configuration, etc. Objective 4: Enable continued high system performance in modern PV systems, as turnover and advances in technology bring new suppliers and high-efficiency modules into the market.

14 SOLAR ENERGY↗

DOE OSTI2026

Data driven investigation to understand the influence of total solids on biological biogas upgrading

In situ biogas upgrading achieves CO 2 conversion to CH 4 via hydrogenotrophic methanogenesis; however, gas-liquid mass transfer constraints limit the upgrading performance. Recognizing that optimization studies often underrepresent the effects of total solids (TS) and organic loading rate (OLR), this study undertook a holistic, statistics driven assessment of operating conditions for in situ H 2 assisted biogas upgrading, centering the analysis on TS and OLR. A dataset of 31 studies was compiled and comprised 99 observations. A rigorous analytical framework was employed, combining data standardization, fixed- and random-effects (REML) weighted regressions with cluster-robust errors, stratified analyses, and machine learning. Mixed-effects meta regression indicated that TS was the main factor explaining differences of methane fraction (CH 4 %) when considering the between studies heterogeneity. Focusing on a near-stoichiometric subset (H 2 /CO 2 ≈ 4:1), TS remained significant. Stratified results showed a stronger negative relationship between TS and CH 4 % in UASB reactors than in CSTRs, with a negative effect under mesophilic conditions and no significant effect under thermophilic conditions. A Random Forest model corroborated the statistical findings, consistently ranking H 2 /CO 2 ratio, OLR, TS, and hydrogen injection rate (HIR) as the most influential predictors. These findings delineate trends across increasing TS levels, particularly between 1% and 10%, and provide preliminary insights for TS above 15% in in situ biogas upgrading. They further provide insights for the influence of TS by reactor type and temperature, thereby advancing the evidence base for implementing biological CO 2 conversion to CH 4 in practice.

In situ biogas upgrading↗

DOE OSTI2025

X-ray thermal diffuse scattering as a texture-robust temperature diagnostic for dynamically compressed solids

We present a model of x-ray thermal diffuse scattering (TDS) from a cubic polycrystal with an arbitrary crystallographic texture, based on the classic approach of Warren [B. E. Warren, Acta Crystallogr. 6, 803 (1953)]. We compare the predictions of our model with femtosecond x-ray diffraction patterns gathered from ambient and dynamically compressed rolled copper foils obtained at the High Energy Density instrument of the European X-Ray Free-Electron Laser facility and find that the texture-aware TDS model yields more accurate results than does the conventional powder model owed to Warren. Nevertheless, we further show: with sufficient angular detector coverage, the TDS signal is largely unchanged by sample orientation and in all cases strongly resembles the signal from a perfectly random powder; shot-to-shot fluctuations in the TDS signal resulting from grain-sampling statistics are at the percent level, in stark contrast to the fluctuations in the Bragg-peak intensities (which are over an order of magnitude greater); and TDS is largely unchanged even following texture evolution caused by compression-induced plastic deformation. We conclude that TDS is robust against texture variation, making it a flexible temperature diagnostic applicable just as well to off-the-shelf commercial foils as to ideal powders.

Crystal lattices↗

DOE OSTI2025

Efficient screening of rare large pit anomalies on polished surfaces using a minimalist sampling scheme

Lawrence Livermore National Laboratory (LLNL) has made significant strides in generating clean energy through its inertial confinement fusion (ICF) experiments. These experiments rely on high-density carbon (HDC) coated shells to encapsulate the fusion fuel. The success of these experiments is heavily dependent on the surface quality of these shells, as even minor imperfections, such as deep pits, can negatively impact fusion yield. Ensuring the required smoothness involves an extensive surface-finishing process that spans approximately 20 stages, making it both time-intensive and resource-demanding. A critical challenge in this process is the need for high-resolution scans to detect rare deep pits, which can be costly and impractical if performed on every shell. This highlights the necessity of developing more efficient scanning methods to optimize time and cost without compromising accuracy. To address these challenges, we introduce a novel approach that employs the multivariate Dvoretzky–Kiefer–Wolfowitz (DKW) inequality to provide a probabilistic upper bound on the error in estimating pit distribution characteristics via a Kernel Density Estimator (KDE). This error bound enables efficient and reliable estimation of pit distribution characteristics at a specified statistical confidence level using a minimal number of surface scans. The integrated DKW-KDE approach was validated through surface-finishing experiments across two batches of HDC-coated shells, demonstrating consistent and robust performance across multiple stages of the surface-finishing experiments. The validation studies suggest that the integrated DKW-KDE approach achieves comparable accuracy in estimating the risk of deleterious large pits with six scans, thus conserving time and resources. Further evaluations show that performance remains consistent across batches and over multiple polishing stages. In conclusion, based on these findings, one can leverage the minimal-scan insights to strategically improve the bottleneck inspection process, thus enhancing the productivity and quality of shell polishing and similar challenging manufacturing processes.

Inertial confinement fusion↗

DOE OSTI2025

Braiding for the win: Harnessing braiding statistics in topological states to play quantum games

Nonlocal quantum games provide proof of principle that quantum resources can confer an advantage at certain tasks. They also provide a compelling way to explore the computational utility of phases of matter on quantum hardware. In a recent paper [O. Hart et al., Phys. Rev. Lett. 134, 130602 (2025)], we demonstrated that a toric code resource state conferred advantage at a certain nonlocal game, which remained robust to small deformations of the resource state. In this paper we demonstrate that this robust advantage is a generic property of resource states drawn from topological or fracton ordered phases of quantum matter. To this end, we illustrate how several other states from paradigmatic topological and fracton ordered phases can function as resources for suitably defined nonlocal games, notably the three-dimensional toric-code phase, the X-cube fracton phase, and the double-semion phase. The key in every case is to design a nonlocal game that harnesses the characteristic braiding processes of a quantum phase as a source of contextuality. We unify the strategies that take advantage of mutual statistics by relating the operators to be measured to order and disorder parameters of an underlying generalized symmetry-breaking phase transition. Additionally, by connecting the win probability to twist products, we show that success at the game serves as a many-body entanglement witness. Namely, if the players implement a perfect quantum strategy on large length scales, the quantum state they share cannot be connected to a trivial product state via a constant-depth local unitary circuit. Lastly, we massively generalize the family of games that admit perfect strategies when codewords of homological quantum error-correcting codes are used as resources.

Fractons↗

DOE OSTI2025

Simulating quantum-classical interfaces via the Lindblad master equation

In hybrid quantum systems, the interface between quantum and classical domains is essential for the generation, control, and measurement of quantum states. Quantum-classical interfaces (QCIs) are ubiquitous in devices such as optical modulators, quantum sensors, and signal processors, where classical signals influence quantum dynamics. In this paper, we employ the Lindblad master equation to simulate the evolution of a quantum system interacting with a classical control system. Our model captures both linear and nonlinear interactions by incorporating first- and second-order susceptibilities, and it quantifies the influence of externally applied control parameters on decoherence and state evolution. As an illustrative example, we analyze an optical modulator and demonstrate how variations in material response and drive conditions affect photon statistics, coherence, and phase-space distributions. In conclusion, the findings offer a path to an all-encompassing model for understanding and optimizing QCIs, with wide-ranging implications for the performance, design, and robustness of next-generation quantum devices.

Quantum engineering↗

DOE OSTI2026

Measuring labor productivity dynamics in U.S. industrial and electric power sectors: a case study (2014–2023)

This study proposes a subsystem methodology for measuring labor productivity in the U.S. industrial and electric power sectors by leveraging public data available between 2014 and 2023. Building on Pasinetti’s framework and subsequent developments, the approach employs Vertically Integrated Sectors (VIS) to account for both direct and indirect productivity effects. The novelty of this work is twofold. First, it enables the estimation of productivity trends over time, providing a robust foundation for empirical analysis. Second, it applies the methodology to a case study of the electric generation sector, highlighting its practical relevance. Using data from the Bureau of Economic Analysis, the Bureau of Labor Statistics, and the Impact Analysis for Planning (IMPLAN) tool, the study reveals significant discrepancies between conventional productivity measures and those derived from the VIS approach. Furthermore, the proposed method aligns with the principles of Integrated Energy Systems by capturing the interrelations among generation, distribution, storage, and consumption. This alignment underscores its utility and applications for energy-related policy and planning. Overall, the findings contribute to more precise labor productivity assessments, supporting informed decision-making and future research. Additionally, the method highlights the importance of considering the whole supply chain, providing interrelated metrics for labor productivity which includes both, direct and indirect effects on the final labor productivity metric. By incorporating intersectoral dependencies, this method offers a more comprehensive and accurate measure of labor productivity compared to traditional metrics.

11 - NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗

DOE OSTI2025

Neutrino-nucleus cross section impacts on neutrino oscillation measurements

The challenges in neutrino-nucleus cross section modeling and its impact on neutrino oscillation experiments are widely recognized. However, a comprehensive and theoretically robust estimation of cross section uncertainties has been lacking, and few studies have quantitatively examined their impact on oscillation measurements. In this work, we evaluate the effect of cross section uncertainties on oscillation parameters using setups inspired by NOvA and DUNE. To characterize these uncertainties, we adopt multiple neutrino-nucleus event generators and simulate a realistic experimental procedure that incorporates near-detector data and near-to-far-detector extrapolation. Our results confirm that cross section uncertainties cannot significantly bias oscillation results in current statistics-dominated experiments like NOvA. However, they could lead to substantial bias for future systematics-dominated experiments like DUNE, even when near-detector data are employed to mitigate uncertainties. These findings underscore the need for further studies on the quantitative impacts of cross section modeling, improved strategies to utilize near-detector data and the PRISM concept, and more robust cross section models to optimize the success of future experiments.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

DOE OSTI2024

Nonparametric Multiparticle Set Methods for Interpreting Environmental Samples

Collection and analysis of environmental samples is commonly used by a range of stakeholders in nuclear safeguards and security contexts. While the ubiquity of samples and their transport in the environment allow regular collection, developing and demonstrating methods for analyzing these samples is difficult. In this work, an environmental sample consists of a set of one or more individual particles. Recent advances in reactor simulation have allowed us to generate data that are more representative of real-world environmental samples, enabling statistically defensible method development and testing. The most notable of these advances is a drastic increase in the number of material depletion regions, which allows our simulations to capture the variation in isotopic composition seen at length scales consistent with environmental samples. Traditional approaches for handling multiparticle samples treat each particle in the sample individually, estimating the quantity of interest (e.g., core-average burnup) resulting from measurement and analysis of signatures (e.g., nuclide assays) from each individual particle. Individual estimates are then averaged to generate a single estimate of the quantity of interest over the entire sample. In this presentation, we introduce two novel approaches for interpreting environmental samples that comprise of multiple particles: (1) the Quantile-Quantile Comparator, which uses a multivariate generalization of quantile-quantile plots for comparing unknown statistical distributions, and (2) the Set Transformer, an attention-based neural network module designed to model interactions among elements (particles) in the input set (sample). Statistically representative sampling cannot be guaranteed as samples are passively collected and are beholden to what particles are available in the environment. These new analysis methods for set-input problems are expected to be more robust than traditional approaches to issues of sampling bias where particles are not uniformly distributed throughout regions of interest, as well as generally outperform traditional approaches by jointly considering all elements in the set. We will present results comparing the performance of traditional single particle approaches and the novel Quantile-Quantile Comparator and Set Transformer for interpretation of simulated environmental samples.

Phathanapirom, Birdy↗

DOE OSTI2024

Generalization error guaranteed auto-encoder-based nonlinear model reduction for operator learning

Many physical processes in science and engineering are naturally represented by operators between infinite-dimensional function spaces. The problem of operator learning, in this context, seeks to extract these physical processes from empirical data, which is challenging due to the infinite or high dimensionality of data. An integral component in addressing this challenge is model reduction, which reduces both the data dimensionality and problem size. In this paper, we utilize low-dimensional nonlinear structures in model reduction by investigating Auto-Encoder-based Neural Network (AENet). AENet first learns the latent variables of the input data and then learns the transformation from these latent variables to corresponding output data. Our numerical experiments validate the ability of AENet to accurately learn the solution operator of nonlinear partial differential equations. Furthermore, we establish a mathematical and statistical estimation theory that analyzes the generalization error of AENet. Finally, our theoretical framework shows that the sample complexity of training AENet is intricately tied to the intrinsic dimension of the modeled process, while also demonstrating the robustness of AENet to noise.

Auto-encoder↗

DOE OSTI2026

Measurements of three-flavor neutrino oscillations from a PISCES two-detector fit to the NOvA Experiment data

NOvA is a long-baseline neutrino oscillation experiment with two functionally identical detectors: a Near Detector (ND) at Fermilab, placed 1 km from the neutrino source, and a Far Detector (FD) located 810 km away from the ND in Minnesota. NOvA s primary physics goals are to measure the neutrino oscillation parameters $\theta_{23}$ and $\Delta m^2_{32}$ with high precision, determine the neutrino mass hierarchy, and constrain the value of $\delta_{CP}$, primarily via the study of muon neutrino to electron neutrino oscillation. Extracting values for oscillation parameters from fits to data usually relies on treating systematic uncertainties as nuisance parameters, a strategy that suffers from poor scalability as the number of uncertainties becomes larger. This work introduces PISCES (Parameter Inference with Systematic Covariance and Exact Statistics), a novel method that circumvents this scalability problem by encoding systematic uncertainties into a covariance matrix. PISCES utilizes a nested minimization in which optimal systematic pulls are first computed using the covariance matrix in an inner minimization step, then the oscillation parameters are profiled over in the outer minimization. PISCES also uses a Poisson Likelihood term, making it ideal for the inclusion of low-statistic samples in the fits. PISCES is a flexible framework that also supports complex fits, such as a joint Near and Far detector fit. In the standard NOvA analysis, oscillation parameters are extracted using an extrapolation technique in which the ND data indirectly constrain the FD prediction via a ratio method. PISCES, on the other hand, enables a simultaneous ND+FD fit, allowing the high-statistics ND data to directly constrain systematic uncertainties across all samples. This thesis presents the full PISCES joint ND+FD fit for the NOvA three-flavor analysis, details its implementation, and evaluates its performance through extensive robustness tests and fake data studies. It also provides a comparison between the PISCES joint ND+FD results and the standard NOvA extrapolation method using the full NOvA 10-year data set. The results demonstrate that PISCES can successfully fit NOvA data while incorporating the constraints from the ND detectors consistently, using physically motivated systematic uncertainties to account for data/MC discrepancies.

Rajaoalisoa, Miriama [Cincinnati U.]↗

DOE OSTI2025

Leveraging transfer learning and leaf spectroscopy for leaf trait prediction with broad spatial, species, and temporal applicability

Accurate and reliable prediction of leaf traits is crucial for understanding plant adaptations to environmental variation, monitoring terrestrial ecosystems, and enhancing comprehension of functional diversity and ecosystem functioning. Currently, various approaches (e.g., statistical, physical models) have been developed to estimate leaf traits through hyperspectral remote sensing and leaf spectroscopy. However, the absence of high-performing, transferable, and stable models across various domains of space, plant functional types (PFTs) and seasons hinder our ability to quantify and comprehend spatiotemporal variations in leaf traits. This study proposes robust and highly transferable models for better predicting leaf traits with hyperspectral reflectance. Initially, three datasets were assembled, pairing common leaf traits — chlorophyll (Chla+b), carotenoids (Ccar), leaf mass per area (LAM), equivalent water thickness (EWT) — with leaf spectra measurements collected across diverse geographic locations in the U.S. and Europe, PFTs, and seasons. Measurements were acquired using spectroradiometers (e.g., ASD FieldSpec 3/4/Pro and SVC HR-1024i) with integrating spheres, leaf clips, and contact probes. Here, we then developed transfer learning-based hybrid models that incorporated the domain knowledge of radiative transfer models (RTMs) through pretraining processes and were well-constrained by fine-tuning with field measurements. Through comparison with other state-of-the-art statistical models, including partial-least squares regression (PLSR) and Gaussian Process Regression (GPR), as well as pure physical models, we found that the proposed transfer learning models achieved better predictive performance and higher transferability. Specifically, compared to other statistical models and pure RTMs, the transfer learning model exhibited higher coefficient of determination (R 2 ) values with range of 0.01 to 0.79, lower normalized root mean square error (NRMSE) with range of 0.06 % to 33.25 % in model performance. Additionally, the models exhibited improved transferability, with higher R 2 values range from 0.04 to 0.32, lower NRMSE range from 0.08 % to 30.81 %. The findings underscore that transfer learning models through integrating domain knowledge from RTMs and limited observations, can harness the advantages of both RTMs and statistical models and serve as a promising approach for effectively predicting leaf traits.

59 BASIC BIOLOGICAL SCIENCES↗

DOE OSTI2024

Impacts of floating-point non-associativity on reproducibility for HPC and deep learning applications

Run to run variability in parallel programs caused by floating-point non-associativity has been known to significantly affect reproducibility in iterative algorithms, due to accumulating errors. Non-reproducibility can critically affect the efficiency and effectiveness of correctness testing for stochastic programs. Recently, the sensitivity of deep learning training and inference pipelines to floating-point non-associativity has been found to sometimes be extreme. It can prevent certification for commercial applications, accurate assessment of robustness and sensitivity, and bug detection. New approaches in scientific computing applications have coupled deep learning models with high-performance computing, leading to an aggravation of debugging and testing challenges. Here we perform an investigation of the statistical properties of floating-point non-associativity within modern parallel programming models, and analyze performance and productivity impacts of replacing atomic operations with deterministic alternatives on GPUs. We examine the recently-added deterministic options in PyTorch within the context of GPU deployment for deep learning, uncovering and quantifying the impacts of input parameters triggering run to run variability and reporting on the reliability and completeness of the documentation. Finally, we evaluate the strategy of exploiting automatic determinism that could be provided by deterministic hardware, using the Groq LPUTM accelerator for inference portions of the deep learning pipeline. We demonstrate the benefits that a hardware-based strategy can provide within reproducibility and correctness efforts.

Shanmugavelu, Sanjif↗