Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “CaP”

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 415 records · Page 23
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Sequential Infiltration Synthesis of Cadmium Sulfide Discrete Atom Clusters

Exposure of soft material templates to alternating volatile chemical precursors can produce inorganic deposition within the permeable template (e.g. a polymer thin film) in a process akin to atomic layer deposition (ALD). While such sequential infiltration synthesis (SIS) processes have now been demonstrated for many metal oxides, we report an SIS process for a transition metal sulfide – CdS. Gas phase dimethyl cadmium and hydrogen sulfide precursors infiltrated into poly(4-vinylpyridine) thin films result in the 3D-nucleation of clusters consistent with a cubane-type Cd 4 S 4 core that are variably terminated with methyl, thiol and hydroxy capping ligands. First principles models and simulation of few-atom Cd-based clusters are consistent with electronic and vibrational spectroscopy and grazing-incidence total X-ray scattering measurements of 3D-cluster-arrays synthesized at 80 °C. The direct synthesis of few-atom transition metal sulfide clusters within polymer thin films will provide a versatile new route to precision architectures for light-absorbing materials including solar energy harvesting and conversion applications.

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

Surface Protected Organozirconium Catalyzes C─H Alumination of Saturated Hydrocarbons

Surface grafted organozirconium catalyzes C─H/Et─Al exchange reactions, involving saturated hydrocarbons and AlEt 3 , to afford organoaluminum compounds and ethane. The Zr(O t Bu) 3 @SiO 2 -Al 2 O 3 –700 (1) catalyst contains monopodal ≡SiO─Zr(O t Bu) 3 and only a few residual silanols (<5%). Nonetheless, these silanols are the Achille's heel of 1, providing a pathway for surface and catalyst degradation during catalysis, limiting the alkylaluminum yield and catalyst turnover. Support degradation, involving the cleavage of Si─O bonds by activated surface organometallics, is inhibited by capping silanols with ─SiMe 3 . Residual silanols in 1 react with allyltrimethylsilane, as determined by solid-state 13 C and 29 Si nuclear magnetic resonance (NMR) spectroscopy, infrared (IR) spectroscopy, and reaction stoichiometry, to form Zr(O t Bu) 3 /SiMe 3 @SiO 2 -Al 2 O 3–700 (2), which is resistant to degradation by AlEt 3 . C─H alumination of dodecane catalyzed by 2 produces higher yields of the 1-dodecylaluminum product in comparison to 1, and in >95% selectivity. Additionally, methane undergoes 2-catalyzed C─H alumination, providing a route to AlMe 3 .

Banerjee, Sayak [Iowa State Univ., Ames, IA (Unite↗
DOE OSTI2020

Structure and activity of supported bimetallic NiPd nanoparticles: influence of preparation method on CO 2 reduction

Here, bimetallic Ni-Pd and monometallic reference catalysts were prepared by decomposing organometallic precursors, Ni(cod) 2 and Pd 2 (dba) 3 , leading to nanoparticles with sizes ranging from 3 to 6 nm. Two different synthesis procedures were followed: i) solution synthesis using capping ligand (hexadecylamine) followed by impregnation of pre-formed nanoparticles on SiO 2 , called Sol-immobilization (SI); and 2) direct precursor decomposition onto SiO 2 , without stabilizer, called Direct Decomposition (DD). Samples prepared by SI procedure are alloyed bimetallic nanoparticles, whereas samples obtained by DD one show phase segregation. Interestingly, DD samples show better activity for CO 2 hydrogenation into CO (reverse water-gas shift reaction - RWGS) than SI ones. The best compromise between activity for CO 2 activation (at lower temperature) and CO selectivity was achieved with Ni DD and NiPd DD catalysts. Moreover, the addition of palladium increased the concentration of surface undercoordinated sites, which chemisorb CO weakly, thus improving activity and selectivity, in opposition to other samples that chemisorb CO strongly, in multiband configuration. In the presence of Pd, different decomposition rates drive the formation of smaller and more active Ni clusters. The knowledge acquired here on the influence of synthesis conditions on the catalytic properties of Ni-Pd catalysts should guide us to better catalysts for CO 2 transformations into valuable products.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2022

Decomposition Kinetics of H 2 O 2 on Pd Nanocrystals with Different Shapes and Surface Strains

Direct synthesis of hydrogen peroxide (H 2 O 2 ) from H 2 and O 2 on a Pd-based catalyst has emerged as a promising route to replace the energy-consuming, highly inefficient anthraquinone process. However, Pd is also a good catalyst for the decomposition of H 2 O 2 , thereby compromising the selectivity toward the desired product. The coupling between the formation and decomposition reactions makes it difficult to single out the most important parameter that controls the selectivity toward direct synthesis of H 2 O 2 . Herein, support-free monometallic Pd nanocrystals with different shapes and surface strains are used to investigate their impacts on the decomposition kinetics of H 2 O 2 . The kinetics are analyzed by tracking the concentration of the remaining H 2 O 2 using infrared spectroscopy. The data indicates that both surface structure and strain affect the decomposition kinetics of H 2 O 2 , but their impacts are inferior to that caused by Br – , a surface capping agent for the Pd{100} facets. The experimental results are consistent with the trend obtained through density functional theory calculations. Furthermore, this work helps shed light on the development of Pd-based catalysts for the direct synthesis of H 2 O 2 by offering strategies to mitigate the decomposition of the desired product.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2020

Mononuclear to Polynuclear U(IV) Structural Units: Effects of Reaction Conditions on U-Furoate Phase Formation

In this work, Uranium(IV) complexation by 2-furoic acid was examined to better understand the effects of ligand identity and reaction conditions on species formation and stability. Five compounds were isolated: [UCl 2 (2-FA) 2 (H 2 O) 2 ] n (1), [U 4 Cl 10 O 2 (THF) 6 (2-FA) 2 ]•2(THF) (2), [U 6 O 4 (OH) 4 (H 2 O) 3 (2-FA) 12 ]•7(THF)•(H 2 O) (3), [U 6 O 4 (OH) 4 (H 2 O) 2 (2-FA) 12 ]•8.76(H 2 O) (4), and [U 38 Cl 42 O 54 (OH) 2 (H 2 O) 20 ]•m(H 2 O)•n(THF) (5). The structures were determined by single crystal X-ray diffraction and further characterized by Raman, IR, and optical absorption spectroscopy. The thermal stability and magnetic behavior of the compounds were also examined. Variations in the synthetic conditions led to notable differences in the structural units observed in the solid state. At low H 2 O/THF ratios, a tetranuclear oxo bridged [U 4 O 2 ] core was isolated. Aging of this solution resulted in the formation a U 38 oxo cluster capped by chloro and water ligands. However, at increasing water concentrations only hexanuclear units were observed. In all cases, at temperatures of 100-120 °C, UO 2 nanoparticles formed.

38 RADIATION CHEMISTRY, RADIOCHEMISTRY, AND NUCLEA↗
DOE OSTI2021

Effect of Ligand Polarity on the Internal Dipoles and Ferroelectric Distortion in BaTiO 3 Nanocubes

Here, surface adsorbates and surrounding matrix species have been demonstrated to affect the properties of nanoscale ferroelectrics and nanoscale ferroelectric composites; potentially counteracting performance losses that can occur in small particle sizes. In this work, the effects of nonpolar oleic acid (OA) and polar tetrafluoroborate (BF 4 - ) ligand capping on the surface of various sizes of BaTiO 3 nanocubes have been investigated with combined neutron diffraction and neutron pair distribution function (PDF), density functional theory (DFT), and ab initio molecular dynamics (AIMD) methods. The low real space PDF region provides an unobstructed view of rhombohedral (split short and long) Ti-O distances in BaTiO 3 nanocubes, mimicking the well-established order-disorder local structure found in bulk BaTiO 3 . Interestingly, the intermediate-range order in nanocubes is found to be orthorhombic, rather than tetragonal. It is concluded that polar ligands adsorbed at BaTiO 3 surfaces stabilize the correlation length scale of local rhombohedral distortions in ferroelectric nanoparticles relative to nonpolar ligands.

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

PCET‐Driven Reactivity of Neptunyl(VI) Yields Oxo‐Bridged Np(V) and Np(IV) Species

Two unconventional polynuclear complexes of neptunium (Np) featuring mono-mathematical equation -oxo motifs have been accessed by proton-coupled electron transfer (PCET) reactivity involving the dissolution of neptunyl(VI) diacetate dihydrate (NpO 2 (OAc) 2 (H 2 O) 2 ∙ HOAc) in methanol followed by addition of a pentadentate Schiff-base ligand. One complex is a mixed-valent [Np V ,Np IV , Np V ] trimer with two bridging mathematical equation μ 2 -oxos and the other is a [Np V , Np V ] dimer featuring a single mathematical equation μ 2 -oxo. In both complexes the outer Np centers are capped with terminal oxo ligands. Spectroscopic and spectrokinetic studies aimed at elucidating mechanistic details of complex formation in this system show that intermediate multinuclear [Np V O 2 (OAc)] n species form prior to metal chelation by the ligand; electrolysis experiments demonstrate that production of Np(V) gives rise to asynchronous proton transfer that does not occur otherwise (in the Np(VI) state) as well as condensation with loss of H 2 O and formation of the polynuclear complexes. We attribute the oxo-deficient nature of these products, with respect to conventional actinyl ([AnO 2 ] m+ ) species, to the reduction/condensation reaction sequence of PCET.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2020

Ligand-Guided Growth of Alloyed Shells on Intermetallic Seeds as a Route toward Multimetallic Nanocatalysts with Shape-Control

In this work, shape-controlled core@shell PdCu@PtCu nanoparticles (NPs) were synthesized by seed-mediated co-reduction. Specifically, cubic-shaped NPs were achieved by selecting tetraoctylammonium bromide (TOAB) and triphenylphosphine (TPP) as capping ligands. Their roles were investigated by incorporating independently each ligand into the reaction and analyzing the products by transmission electron microscopy (TEM). This analysis revealed that the quasi-spherical PdCu seeds acquired cubic shapes during the synthesis and directed shell deposition. This process was mediated by TPP, which was central to achieving monodisperse NPs with shape-control. The synthesis conditions were modified to tune both the thickness and composition of the shells. Evaluation of the NPs as catalysts for the electrooxidation of formic acid found that the NPs with the thinnest Pt-rich shells gave the highest specific activity, with the nanocubic shape also enhancing performance with respect to the spherical counterpart. These results highlight the benefits of integrating compositional, architectural, and shape-control all in one NP construct.

36 MATERIALS SCIENCE↗
DOE OSTI2022

Silver-Mediated Squaric Acid Reduction as a Facile, Ambient-temperature and Seedless Route to Tunable Bimetallic Au/Ag Nanostars and Nanosnowflakes

In this work, bimetallic Au/Ag nanoparticles were successfully synthesized using a facile and reproducible method involving rapid, room temperature mixing of aqueous solutions of HAuCl 4 and AgNO 3 in the presence of squaric acid as both the reducing agent and surface capping ligand. Variation in the initial Au:Ag molar ratio yielded a tunable anisotropic shape and affiliated surface plasmon resonance. The nanoparticles displayed excellent stability in water for at least 10 weeks and showed robust catalytic activity in the borohydride-assisted reduction of the model substrate 4-nitrophenol.

36 MATERIALS SCIENCE↗
DOE OSTI2024

Morphology‐controlled synthesis of multi‐metal‐based spinel oxide nanocatalysts and their performance for oxygen reduction

Abstract We present a one‐pot colloidal synthesis method for producing monodisperse multi‐metal (Co, Mn, and Fe) spinel nanocrystals (NCs), including nanocubes, nano‐octahedra, and concave nanocubes. This study explores the mechanism of morphology control, showcasing the pivotal roles of metal precursors and capping ligands in determining the exposed crystal planes on the NC surface. The cubic spinel NCs, terminated with exclusive {100}‐facets, demonstrate superior electrocatalytic activity for the oxygen reduction reaction (ORR) in alkaline media compared to their octahedral and concave cubic counterparts. Specifically, at 0.85 V, (CoMn)Fe 2 O 4 spinel oxide nanocubes achieve a high mass activity of 23.9 A/g and exhibit excellent stability, highlighting the promising ORR performance associated with {100}‐facets of multi‐metal spinel oxides over other low‐index and high‐index facets. Motivated by exploring the correlation between ORR performance and surface atom arrangement (active sites), surface element composition, as well as other factors, this study introduces a prospective approach for shape‐controlled synthesis of advanced spinel oxide NCs. It underscores the significance of catalyst shape control and suggests potential applications as nonprecious metal ORR electrocatalysts.

Li, Can↗
DOE OSTI2021

Risk-based area of review estimation in overpressured reservoirs to support injection well storage facility permit requirements for CO 2 storage projects

This paper by the Energy & Environmental Research Center presents a workflow and modeling approach for delineating a risk-based area of review (AOR) to support a U.S. Environmental Protection Agency (EPA) Class VI permit for a carbon dioxide (CO 2 ) storage project. The approach combines semianalytical solutions for estimating formation fluid leakage through a hypothetical leaky wellbore with the results of numerical reservoir simulations to define the AOR. The modeling utilizes 1) semianalytical solutions from the peer-reviewed literature for formation fluid leakage through abandoned wellbores by Raven (1990) and Avci (1994), 2) a FORTRAN model compiled and described in Cihan et al. (2011, 2012) called ASLMA (Analytical Solution for Leakage in Multilayered Aquifers), and 3) a computational framework for estimating a risk-based AOR first proposed by Oldenburg et al. (2014, 2016). Therefore, the approach builds upon well-established research and underlying hydrogeological principles that have been upheld for nearly three decades. Moreover, the ASLMA model has been broadly applied to an array of storage projects. The work presented herein extends these earlier works using a custom wrapper written in the software environment, R (R Core Team, 2020), which was developed to perform multiple runs of the ASLMA model using given ranges for one or more input parameters. In addition, the current work simulates the pressure buildup within the storage reservoir in response to CO 2 injection using a compositional simulator to better accommodate the temporospatial evolution of pressure buildup within the storage reservoir that is more accurately modeled using a heterogeneous geologic model and a compositional simulator that accounts for the multiphase interactions. The workflow is demonstrated using a case study for a 180,000-metric-ton-per-year storage project located in the PCOR (Plains CO 2 Reduction) Partnership region. For the storage project evaluated here, under the scenario where the leaky wellbore is open to a saline aquifer (thief zone) between the overlying seal (cap rock) and the underground sources of drinking water (USDW), the risk-based AOR essentially collapses to the areal extent of the CO 2 plume in the storage reservoir because the pressure buildup in the storage reservoir beyond the CO 2 plume is insufficient to drive formation fluids up a hypothetical leaky wellbore into the USDW. However, even under the conservative assumption that the leaky wellbore is not open to a thief zone, beyond the areal extent of the CO 2 plume, the incremental leakage is less than 400 m 3 over 20 years, which represents ~0.0001% or less of the total volume of water contained within the USDW rock volume. As discussed in the text, the threshold criterion for defining the risk-based AOR is site-specific and should be informed by the results of the sensitivity analysis and available site characterization data. The approach outlined in this paper is designed to be protective of USDWs and, therefore, comply with the Safe Drinking Water Act requirements and provisions for the U.S. EPA Class VI Underground Injection Control (UIC) Program (Class VI Rule) and North Dakota Administrative Code Chapter 43-05-01.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2024

Effect of Iron Contamination and Polysilicon Gettering on the Performance of Polysilicon‐Based Passivating Contact Solar Cells

Over the past decade, silicon solar cells with carrier-selective passivating contacts based on polysilicon capping an ultra-thin silicon oxide (commonly known as TOPCon or POLO) have demonstrated promising efficiency potentials and are regarded as an evolutionary upgrade to the PERC (passivated emitter and rear contact) cells in manufacturing. The polysilicon-based passivating contacts also exhibit excellent gettering effects that relax the wafer and cleanroom requirements to some extent. Here, in this work, we experimentally explore the impact of bulk iron contamination and polysilicon gettering on the passivation quality of the polysilicon/oxide structure and the resulting solar cells performance. Results show that both n- and p-type polysilicon/oxide passivating contacts are not affected by iron gettering, demonstrating robust and stable passivation quality. However, for a very high bulk iron contamination (1 × 10 13 cm −3 ), the accumulated iron in the p-type lightly boron-doped emitter in crystalline silicon would degrade the emitter saturation current density. This can cause a reduction in both open-circuit voltage and short-circuit current. Meanwhile, this very high iron content (1 × 10 13 cm −3 ) can further degrade the fill factor and temperature coefficient of the cells. On the other hand, for an initial iron content of 2 × 10 12 cm −3 , which should be well above the iron level in the current industrial Czochralski silicon wafers, the resulting cells demonstrate similar performance as the control group with no intentional iron contamination. This work brings attention to both the benefits of polysilicon gettering effects as well as the potential degradation due to the accumulation of metal impurities in the p-type emitter region.

14 SOLAR ENERGY↗
DOE OSTI2022

Pinhole‐Free Ultrathin Silicon Oxide Layer by Ozone‐Dissolved Deionized Water

The pinhole‐free silicon oxide (SiO x ) layer grown using dissolved ozone in deionized water (DI‐O 3 ) is demonstrated. An ultrathin SiO x layer of thickness 1.53 nm ± 1.5% is grown on silicon wafer using 7 ppm DI‐O 3 . A four‐step methodology, including 25 wt% aq. tetramethylammonium hydroxide (TMAH) solution at 80 °C, is employed to magnify the pinhole signatures in the underlying silicon and visualized with dark‐field optical microscopy (DFOM) and scanning electron microscopy (SEM). The DFOM images show spots on the surface of wafer that originate from airborne particles, contamination, etc. SEM images reveal the absence of etch pits in silicon even after etching of SiO x /silicon in TMAH for 300 s. The minority carrier lifetime and interface states density in AlO x ‐capped SiO x /Si structure are ≥2 ms at carrier density of 1 × 10 15 cm −3 and less than 2 × 10 11 cm −2 eV −1 , respectively.

42 ENGINEERING↗
DOE OSTI2023

Synthesis of Free-Standing Pd-Ni-P Metallic Glass Nanoparticles with Durable Medium-Range Ordered Structure for Enhanced Electrocatalytic Properties

Topologically disordered metallic glass nanoparticles (MGNPs) with highly active and tailorable surface chemistries have immense potential for functional uses. The synthesis of free-standing MGNPs is crucial and intensively pursued because their activity strongly depends on their exposed surfaces. Herein, a novel laser-evaporated inert-gas condensation method is designed and successfully developed for synthesizing free-standing MGNPs without substrates or capping agents, which is implemented via pulse laser-induced atomic vapor deposition under an inert helium atmosphere. In this way, the metallic atoms vaporized from the targets collide with helium atoms and then condense into short-range-order (SRO) clusters, which mutually assemble to form the MGNPs. Using this method, free-standing Pd 40 Ni 40 P 20 MGNPs with a spherical morphology are synthesized, which demonstrates satisfactory electrocatalytic activity and durability in oxygen reduction reactions. Moreover, local structure investigations using synchrotron pair distribution function techniques reveal the transformation of SRO cluster connection motifs of the MGNPs from face-sharing to edge-sharing modes during cyclic voltammetry cycles, which enhances the electrochemical stability by blocking crystallization. Further, this approach provides a general strategy for preparing free-standing MGNPs with high surface activities, which may have widespread functional applications

36 MATERIALS SCIENCE↗
DOE OSTI2025

Monolayer TiS 2 Nanosheets on Au(111)–Structural Characterization and Effect of Edge Stability for Shape Control

Transition metal dichalcogenides are promising alternatives to noble metal catalysts, e.g., for (photo-)activation of greenhouse gases or hydrogenations. Herein, a direct synthetic route for 2D TiS 2 nanosheets on Au(111) by titanium deposition in the presence of a mild, organic, non-oxidizing sulfur source is presented. High-resolution scanning tunneling microscopy (STM) is used to gain atomic-level insights into the TiS 2 nanosheet morphology. In contrast to the literature, this protocol gains mostly hexagonal and truncated triangular nanosheets with an increased edge contrast in STM, analog to metallic edge states in MoS 2 . Synchrotron-based photoelectron spectroscopy allows insights into compositional details, specifically to distinguish different S sites on the TiS 2 sheets and other S species on the sample. Further, a minimum size is identified (9 S atoms side length), which underlines the importance of moiré reconstructions for stress relief. The TiS 2 sheets coexist with [Au]Ti 1 S 3 clusters, in which a single gold atom is alloyed into the surface and capped by three S atoms. Together with the finding of a critical sheet size, this points toward on-surface Ostwald ripening as a relevant process in the sheet formation. Ab-initio calculations (density functional theory) underscore that the chemical potential of S is an essential descriptor to maintain shape control.

2D nanomaterials↗
DOE OSTI2024

Generalized Synthesis of Highly-Dispersed, Ultrafine Transition Metal Nanoparticles on Silica Spheres for Enhanced Optical Absorption

Robust synthesis of ultrafine metal nanoparticles (ufMNPs) below 5 nm with clean surfaces and strong optical absorption in the visible spectral range is challenging due to their instability originating from large surface-to-volume ratios. Here, this work reports a general strategy involving two sequential steps: i) loading metal precursor ions onto the surface of silica nanospheres (SiO x NSs) by forming a uniform coating of metal oxyhydroxide [MO y (OH) z ] through preferred surface acid–base reactions and ii) thermally reducing MO y (OH) z in forming gas at elevated temperatures to form ufMNPs evenly dispersed on the surface of SiO x NSs. The capability of this synthesis strategy is verified by loading ufMNPs of various transition metals and bimetallic combinations onto the SiO x NSs. The ufMNPs exhibit strong optical absorption enhanced by the optical scattering resonances in the SiO x NSs, which generate intense electric fields near the surface of the SiO x NSs. The SiO x NSs also support stabilizing the ufMNPs, which do not need additional organic capping reagents. The successful synthesis of SiO x -NS-supported ufMNPs with clean surfaces and enhanced optical absorption is promising for exploring the photocatalytic properties of ufMNPs.

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

Nanopinhole-Enabled, Hole-Selective Poly-Si/SioxNy Passivating Contacts on Textured c-Si for Si Solar Cells

The next-generation silicon photovoltaics will be based on passivating electron- and hole-selective contacts with both very low interface recombination and contact resistivities. While the emerging mainstream TOPCon technology has developed excellent electron-selective poly-Si/tunneling SiOx contacts, hole-selective contacts, especially on textured surfaces, have remained a significant challenge. This contribution introduces novel high-performance hole selective poly-Si contacts on pyramid-textured Si, enabled by electrochemically produced hole transport nanopinholes in a 10 nm oxynitride passivating dielectric stack capped by p+ poly-Si. The highly passivating oxynitride layer is produced via atomic intermixing of O and N atoms in the initial SiOx/SiNy layer stack upon thermal annealing. Carrier transport is governed by nanopinhole density and size are tuned by Ag nanoparticle electrodeposition and surface attachment chemistries. This results in passivating hole contact resistivities in the m..omega..-cm2 range, while preserving interface recombination current prefactor around 5 fA/cm2.

14 SOLAR ENERGY↗
DOE OSTI2025

Robustness of Deep Learning Classification to Adversarial Input on GPUs: Asynchronous Parallel Accumulation Is a Source of Vulnerability

The ability of machine learning (ML) classification models to resist small, targeted input perturbations—known as adversarial attacks—is a key measure of their safety and reliability. We show that floating-point non associativity (FPNA) coupled with asynchronous parallel programming on GPUs is sufficient to result in misclassification, without any perturbation to the input. Additionally, we show that this misclassification is particularly significant for inputs close to the decision boundary and that standard adversarial robustness results may be overestimated up to 4.6 when not considering machine-level details. We first study a linear classifier, before focusing on standard Graph Neural Network (GNN) architectures and datasets used in robustness assessments. We develop a novel black-box attack using Bayesian optimization to discover external workloads that can change the instruction scheduling which bias the output of reductions on GPUs and reliably lead to misclassification. Motivated by these results, we present a new learnable permutation (LP) gradient-based approach to learning floating-point operation orderings that lead to misclassifications. The LP approach provides a worst-case estimate in a computationally efficient manner, avoiding the need to run identical experiments tens of thousands of times over a potentially large set of possible GPU states or architectures. Finally, using instrumentation-based testing, we investigate parallel reduction ordering across different GPU architectures under external background workloads, when utilizing multi-GPU virtualization, and when applying power capping. Our results demonstrate that parallel reduction ordering varies significantly across architectures under the first two conditions, substantially increasing the search space required to fully test the effects of this parallel scheduler-based vulnerability. These results and the methods developed here can help to include machine-level considerations into adversarial robustness assessments, which can make a difference in safety and mission critical applications.

Shanmugavelu, Sanjif [Maxeler Technologies, a Groq↗