Search NASA⌕ Search

SEARCH · Search NASA

Results for “robust optimization”

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 199 records · Page 11
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Computational Algorithms for Unit Commitment with AC Power Flows (Final Report)

Security-constrained unit commitment (SCUC) is a key component in power system operations. When AC power flow constraints are considered in the SCUC model (AC-SCUC), the problem becomes extremely difficult due to its discrete and non-convex nature, as described in “Grid Optimization Competition Challenge 3 Problem Formulation (GOCC)”. There are four main challenges: (i) Discrete decisions regarding unit online/offline status and start-up/shut-down procedures for every single unit. The number of discrete decision variables increases considerably when a system integrates multiple generators; (ii) Configuration-based combined-cycle formulations, and multi-commodity models that include ramping products, spin/non-spin products, and regulation up/down products. The combined-cycle units introduce additional discrete decision variables and auxiliary service products further complicate the model by connecting multi-commodity products’ continuous and discrete variables; (iii) SCUC models with AC power flow constraints are far more complex due to massive bilinear terms in the large-scale nonlinear power balance equations. The nonlinear power balance equations are further complicated by the discrete step control variables of shunts; (iv) N − 1 contingency analysis. The size of the model increases linearly with the number of contingencies considered, greatly increasing the size of the optimization model. Accordingly, there is an emergent need to develop a robust algorithm capable of deriving a high-quality solution in a short time and passing through contingency tests simultaneously. In this project, we explore innovative techniques to address this challenging problem by integrating advanced polyhedral theory, approximation methods, relaxation strategies, decomposition techniques, and parallel computing. Each technique approaches the problem from a different perspective, leveraging its specific strengths to tackle distinct challenges. Each individual method has demonstrated its effectiveness in the PI’s previous research. Their integration is expected to significantly reduce the computational time required to solve the proposed complex problem. Successful completion of this project has the potential to transform the industry by enhancing optimization solvers capable of handling large-scale day-ahead energy market clearing models within strict time constraints, while incorporating AC power flow constraints. This advancement will lead to reduced overall generation costs and, consequently, increased social welfare.

29 ENERGY PLANNING, POLICY, AND ECONOMY↗
DOE OSTI2026

Machine-Learning-Driven Discovery of Water Splitting BaFe 2 O 4 and Human-in-the-Loop Improvement via Al-Substitution for Increased Thermal Stability

Thermochemical hydrogen (TCH) production offers a promising method for converting thermal energy into hydrogen fuel through heat-driven redox cycles of metal oxides. Here, in this work a defect graph neural network (dGNN) was used to predict oxygen vacancy formation energies ΔH V O combined with Materials Project predictions of oxygen chemical potential stability to screen candidate oxides via high-throughput database analysis. BaFe 2 O 4 was identified as a promising material for experimental validation based on its predicted ΔH V O , oxygen chemical potential stability range, and potential for tunable substitutions to improve thermal properties. Experimental validation using thermogravimetric analysis (TGA), stagnation flow reactor (SFR), X-ray diffraction (XRD), and electron microscopy confirmed positive water-splitting behavior but also revealed limitations in thermal stability under aggressive reduction conditions. To address this, a human-in-the-loop modification strategy was employed introducing Al substitution in BaFe 2–x Al x O 4 ; this modification improves thermal stability, alters the crystal structure and enhances overall performance. These results demonstrate a combined computational and experimental workflow in which machine learning accelerates identification of promising candidates, while targeted experimental design enables optimization of functional performance. This approach advances the development of robust, cost-effective TCH materials and highlights the importance of integrating data-driven discovery with human-guided materials design in paving the way for scalable hydrogen production technologies.

organic↗
DOE OSTI2024

Waveform resampling with LMN method

In this article, resampling is a common technique applied in digital signal processing. Based on the Fast Fourier Transformation (FFT), we apply an optimization called here the LMN method to achieve fast and robust re-sampling. In addition to performance comparisons with some other popular methods, we illustrate the effectiveness of this LMN method in a particle physics experiment: re-sampling of waveforms from Liquid Argon Time Projection Chambers.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2026

Machine learning method for enforcing variable independence in background estimation with LHC data: ABCDisCoTEC

A novel solution is presented for the problem of estimating the backgrounds of a signal search using observed data while simultaneously maximizing the sensitivity of the search to the signal. The 'ABCD method' provides a reliable framework for background estimation by partitioning events into one signal-enhanced region (A) and three background-enhanced control regions (B, C, and D) via two smoothly varying, statistically independent variables. In practice, even slight correlations between the two variables can significantly undermine the method's performance. Thus, choosing appropriate variables by hand can present a formidable challenge, especially when background and signal differ only subtly. To address this issue, the ABCD with distance correlation (ABCDisCo) method was developed to construct two learned variables via a neural network trained to provide strong signal-background discrimination with small values of the distance correlation (DisCo) measure between the two learned variables. However, relying solely on minimizing the DisCo can result in learned variables that may not have distributions of background events that are smoothly varying and localized at extreme values, as necessary for the validity of the background estimation. The ABCDisCo training enhanced with closure (ABCDisCoTEC) method is introduced to solve this issue by directly minimizing the nonclosure, expressed as a dedicated differentiable loss term. This extended method is applied to a data set of proton-proton collisions at a center-of-mass energy of 13 TeV recorded by the CMS detector at the CERN Large Hadron Collider. Additionally, given the complexity of the minimization problem with constraints on multiple loss terms, the modified differential method of multipliers is applied and shown to greatly improve the stability and robustness of the ABCDisCoTEC method, compared to grid search hyperparameter optimization procedures.

Hayrapetyan, Aram [Yerevan Phys. Inst.]↗
DOE OSTI2026

Dual Representations and H ∞ -Optimal Control of Partial Differential Equations

We consider H ∞ -optimal state-feedback control of the class of linear Partial Differential Equations (PDEs) which admit a Partial Integral Equation (PIE) representation. While linear matrix inequalities are commonly used for optimal control of Ordinary Differential Equations (ODEs), the absence of a universal state-space representation and suitable dual form prevents such methods from being applied to optimal control of PDEs. Specifically, for ODEs, the controller synthesis problem is defined in state-space, and duality is used to resolve the bilinearity of that synthesis problem. Recently, the PIE representation was proposed as a universal state-space representation for linear PDE systems. In this paper, we show that any PDE system represented by a PIE admits a dual PIE with identical stability and I/O properties. This result allows us to reformulate the stabilizing and optimal state-feedback control problems as convex optimization over the cone of positive Partial Integral (PI) operators. Operator inversion formulae then allow us to construct feedback gains for the original PDE system. The results are verified through application to several canonical problems in optimal control of PDEs and indicate the resulting bounds on H ∞ norm are not conservative.

42 ENGINEERING↗
DOE OSTI2026

Phase Picking Beyond Local Distances: Where Waveform Filtering Still Matters for Deep Learning Models

Waveform filtering is a standard step in traditional seismic phase picking but often receives little attention in deep learning workflows, where models are typically trained on raw or minimally processed waveforms. Although this strategy performs well for local events, we show that performance can degrade substantially at regional distances. To address this limitation, we introduce two ways to incorporate multiband-filtered waveforms into deep learning phase pickers. The stacking approach concatenates filtered inputs along the channel dimension, while the branching approach processes each frequency band through a dedicated network branch before feature fusion. Both approaches can substantially improve performance across epicentral distances of 0° to 20°, but their effectiveness depends strongly on the selected frequency bands. Tests with multiple filter banks show that filter-bank design should be treated as part of model optimization rather than as a fixed preprocessing choice. Grad-CAM analysis of the branching model indicates that band importance varies among waveform samples and across training realizations, with only a weak overall preference for the 0.25 to 0.5 Hz band. These results show that no single filter band is consistently optimal and demonstrate that explicit feature engineering remains valuable for robust deep learning-based seismic phase picking.

58 GEOSCIENCES↗
DOE OSTI2024

Fractal Nanostructured Solar Selective Surfaces for Next Gen Concentrating Solar Power (Final Report)

This project reports a novel coating with enhanced solar absorptance and reduced thermal emittance with high efficiency at elevated temperature for next-generation concentrated solar power (CSP) plants, with targeted operating temperatures around 750°C. Highly textured single and multimetallic oxide coatings were electrodeposited onto Inconel substrate by systematically varying the composition and process parameters. The optimized coating exhibited micro-to-nano structures designed to match the wavelengths in the visible region of the solar spectrum. These structures facilitate resonant absorption of solar radiation, significantly boosting solar absorption and accommodating thermal stress during high temperature exposure. A high solar absorptance exceeding 0.985 and a low thermal emittance below 0.5, yielding a thermal efficiency near 95%, was achieved for the optimized coatings without any anti-reflective overcoat, that remained robust after 750 h of isothermal exposure to 750°C in air. The coatings are also robust to severe mechanical and environmental stressors. The innovative approach presented in this study demonstrates the potential for tailoring air-stable solar absorber coatings to achieve high absorption, low emittance, and excellent high-temperature endurance, meeting the rigorous demands of next-generation CSP systems. A technoeconomic analysis reveals the economic advantage of the coatings for Gen3 CSP installations in different geographical zones globally.

14 SOLAR ENERGY↗
DOE OSTI2025

Simulation-driven design optimization of reaction injection molding (RIM) process for polydicyclopentadiene (pDCPD): Minimizing cycle time, defects, and warpage

Replacing metal components in trucks, trailers, and buses with lightweight polymer composites is challenging due to high temperatures and complex manufacturing. The Reaction Injection Molding (RIM) process using Dicyclopentadiene (DCPD) resin offers a solution by producing robust parts with excellent stiffness, impact strength, and resistance properties. Simulations are essential for optimizing this process, predicting defects, and improving quality. However, most commercial software is tailored for thermoplastics, requiring thermoset users to generate their own datasets. In this study, a material data card for DCPD was developed to perform RIM simulations. Design of Experiments (DOE) was used to identify key factors affecting filling, curing, and warpage, aiming to minimize cycle time and defects. The simulations explored varying injection gate parameters (size, location, number) and process conditions (mold/resin temperature, injection/curing pressure). Results showed that gate design significantly impacts filling behavior and defects. A single central gate provided balanced flow with fewer defects, while two corner gates led to more defects. Additionally, lower injection pressure increased filling time, while higher mold temperature accelerated curing but led to more warpage. In conclusion, this optimization framework aims to enhance DCPD part performance and promote sustainable manufacturing by reducing waste and energy consumption.

42 ENGINEERING↗
DOE OSTI2025

From Structure to Function: Zn/Mn-Modified Maghemite as an Advanced Nanoplatform for Magnetic Hyperthermia and Radionuclide Therapy

The development of nanoplatforms capable of efficient heat generation and stable radionuclide delivery is essential for effective bimodal cancer therapy. Here, in this study, binary (Fe–M) and ternary (Fe–M–M′) metal oxide nanoparticles were synthesized via a polyol method optimized to produce flower-like γ-Fe 2 O 3 (maghemite) structures, with M and M′ representing Zn and/or Mn. Comprehensive structural and magnetic characterization was conducted to explain the relationship between composition, defect structure, and hyperthermic performance. The analyses revealed that cation substitution induced an Fe-site vacancy, primarily at octahedral positions, leading to local structural distortions, as confirmed by powder X-ray diffraction and pair distribution function analysis. The optimized composition, with Zn/Mn/Fe = 0.040:0.182:1, exhibited the highest concentration of vacancies and structural disorder. These vacancies altered the bonding environment, enhancing magnetic interactions at tetrahedral sites while weakening those at the octahedral positions. The resulting multicore nanoflowers (20–63 nm; core size 13–18 nm) displayed strong heating performance, with intrinsic loss power ranging from 0.34 to 5.77 nHm 2 kg –1 . The optimized sample achieved a temperature increase of 30 °C within 2 min and a specific absorption rate of 369 W g –1 . This composition was further coated with citrate (CA) and successfully radiolabeled with 177 Lu, achieving a radiolabeling yield of 92.7% and excellent stability, thus forming a robust nanoplatform for combined magnetic hyperthermia and radionuclide therapy. Biological evaluation of the optimized S5 composition revealed selective cytotoxicity toward HeLa and LS174 cells, while toxicity was significantly lower to A549, A375, and normal MRC-5 cells. Citrate coating of S5 nanoparticles (S5@CA) drastically reduced their cytotoxicity across all tested cell lines (IC 50 > 200 μg mL –1 ), confirming their enhanced biocompatibility for therapeutic applications. In HeLa cells subjected to magnetic hyperthermia, the viability decreased to approximately 84% after 30 min and 61% after 60 min of treatment, demonstrating the sustained hyperthermic effect at a controlled working temperature of 48 °C. These results underscore the effectiveness of cation substitution and vacancy engineering in tailoring the functional properties of maghemite-based nanomaterials for advanced multimodal cancer therapies.

36 MATERIALS SCIENCE↗
DOE OSTI2024

Solidification and crystallographic texture modeling of laser powder bed fusion Ti-6Al-4V using finite difference-monte carlo method

Laser powder bed fusion (LPBF) additive manufacturing makes near-net-shaped parts with reduced material cost and time, rising as a promising technology to fabricate Ti-6Al-4V, a widely used titanium alloy in aerospace and medical industries. However, LPBF Ti-6Al-4V parts produced with 67° rotation between layers, a scan strategy commonly used to reduce microstructure and property inhomogeneity, have varying grain morphologies and weak crystallographic textures that change depending on processing parameters. Here, this study predicts LPBF Ti-6Al-4V solidification at three energy levels using a finite difference-Monte Carlo method and validates the simulations with large-area electron backscatter diffraction (EBSD) scans. The developed model accurately shows that a <001> texture forms at low energy and a <111> texture occurs at higher energies parallel to the build direction but with a lower strength than the textures observed from EBSD. A validated and well-established method of combining spatial correlation and general spherical harmonics representation of texture is developed to calculate a difference score between simulations and experiments. The quantitative comparison enables effective fine-tuning of nucleation density (N 0 ) input, which shows a nonlinear relationship with increasing energy level. Future improvements in texture prediction code and a more comprehensive study of N 0 with different energy levels will further advance the optimization of LPBF Ti-6Al-4V components. These developments contribute a novel understanding of crystallographic texture formation in LPBF Ti-6Al-4V, the development of robust model validation and calibration pipeline methodologies, and provide a platform for mechanical property prediction and process parameter optimization.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2024

Research Highlight - Dispatch Optimization, System Design and Cost Benefit Analysis of a Nuclear Reactor with Molten Salt Thermal Storage

We highlight our work from our previous publication titled "Dispatch Optimization, System Design, and Cost Benefit Analysis of a Nuclear Reactor with Molten Salt Thermal Storage" published in MDPI Energies 2022 (doi.org/10.3390/en15103599). Variable renewable energy availability has increased the volatility in energy prices in most markets. Nuclear power plants, with a large ratio of capital to variable costs, have historically operated as base load energy suppliers but the need for more flexible operation is increasing. We simulate the techno-economic performance of a 950 MWt nuclear power plant, based on the Westinghouse lead-cooled fast reactor, coupled with molten salt thermal storage as a method for flexible energy dispatch. We use the System Advisor Model to model the nuclear reactor thermal power input and power cycle operating modes. We combine this robust engineering model with a mixed-integer linear program model for optimized dispatch scheduling. We then simulate the coupled nuclear and thermal storage system under different market scenarios with varying price volatility. We find that the coupled plant outperforms the base plant under markets where energy price peaks fluctuate by a factor of two or more about the mean price. We show that a calculated power purchase agreement price for the plant improves by up to 10% when operating under California energy market conditions. Sensitivity analysis on the thermal storage cost shows that the optimal design remains unchanged even when doubling costs.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2026

DESI DR1 Ly α 1D power spectrum: Validation of estimators

The Data Release 1 (DR1) of the Dark Energy Spectroscopic Instrument (DESI) is the largest sample to date for small-scale Lyα forest cosmology, accessed through its one-dimensional power spectrum (P 1D ). The Lyα forest P 1D is extracted from quasar spectra that are highly inhomogeneous (both in wavelength and between quasars) in noise properties due to intrinsic properties of the quasar, atmospheric and astrophysical contamination, and also sensitive to low-level details of the spectral extraction pipeline. We employ two estimators in DR1 analysis to measure P 1D : the optimal estimator and the fast Fourier transform (FFT) estimator. To ensure robustness of our DR1 measurements, we validate these two power spectrum and covariance matrix estimation methodologies against the challenging aspects of the data. First, using a set of 20 synthetic 1D realizations of DR1, we derive the masking bias corrections needed for the FFT estimator and the continuum fitting bias needed for both estimators. We demonstrate that both estimators, including their covariances, are unbiased with these corrections using the Kolmogorov-Smirnov test. Second, we substantially extend our previous suite of CCD image simulations to include 675,000 quasars, allowing us to accurately quantify the pipeline's performance. This set of simulations reveals biases at the highest k values, corresponding to a resolution error of a few percent. We base the resolution systematics error budget of DR1 P 1D on these values, but do not derive corrections from them since the simulation fidelity is insufficient for precise corrections.

Lyman alpha forest↗
DOE OSTI2026

Time-Evolved Hetero-Alkali Interphases Enable Long-Life Sulfide-Based Anode-Free Solid-State Batteries

Sulfide-based anode-free solid-state batteries (AFSSBs) offer compelling advantages in terms of energy density and safety, yet their practical implementation is severely hindered by undesirable interfacial reactions between sulfide solid electrolytes (SEs) and freshly plated lithium (Li), as well as non-uniform Li plating/stripping behavior. Herein, an effective interfacial stabilization strategy by incorporating sodium bis(fluorosulfonyl)imide (NaFSI) additive into the Li5.4PS4.4Cl1.6 (LPSC) is investigated. Unlike conventional Li-based additives that form static passivation layers, NaFSI introduces a transient hetero-alkali chemistry that kinetically governs interphase evolution during fresh Li plating. NaFSI induces a timesequenced interphase evolution: an initial NaF/LiF-rich layer that suppresses early sulfide reduction, followed by a LiF/Li3N-rich layer that optimizes Li⁺ transport during repeated anode-free cycling. This evolved robust and fast ion conducting layer mitigates interfacial impedance growth, enhances Li + transport kinetics, and suppresses localized Li growth and filamentary shorting. As a result, the anode-free full cell with NaFSI modified LPSC as the interlayer exhibits an excellent cycling stability over 500 cycles at 0.2 C with a capacity retention of 77.6%, whereas the cell with bare LPSC suffers from rapid capacity decay after 100 cycles, retaining only 32.1% of its initial capacity. This work establishes dynamic heteroalkali additive chemistry as a general strategy to kinetically program solid-solid interphases, guiding the interface design in anode-free solid-state batteries.

25 ENERGY STORAGE↗
DOE OSTI2026

There and Back Again: Reimagining Cryogenic Cooling for Scalable Arrays of Dilution Refrigerators for future Quantum Datacenters

While pulse tube cryocoolers enabled the rapid expansion of dilution refrigerator technology over the past two decades, the transition to large-scale quantum systems is now driving a reassessment of the DR’s higher-temperature-stage cooling strategies and how these systems can be effectively scaled in a modular way. Quasi-wet architectures based on centralized cryoplants and forced-flow helium distribution offer compelling advantages in energy efficiency, operational cost, and scalability. With appropriate redundancy, standardized interfaces, and optimized distribution system designs, these architectures will provide a practical and robust path forward for the next generation of quantum computing infrastructure.

Hansen, B. [Fermilab]↗
DOE OSTI2026

A 3D-Printed Millimeter-Wave Inline Waveguide-to-Coplanar-Waveguide Transition to Enable Dense Spectrometer Arrays for Intensity Mapping Surveys

We present a 3D-printed millimeter-wave, octave-bandwidth, in-line waveguide-to-coplanar-waveguide transition designed to enable focal planes with dense arrays of on-chip spectrometers. These arrays will enable compelling surveys of the large-scale structure of the universe through millimeter-wave intensity mapping. The transition consists of a four-step ridge-waveguide transformer that couples light from a rectangular waveguide onto a coplanar waveguide via an electrical connection made with indium bump bonds. We develop a tolerance-aware optimization approach to identify high-performance transition geometries that are robust to manufacturing variations; the same formulation can be applied to other tolerance-sensitive design problems. We also describe the implementation of a custom apparatus and procedure for bump-bonding a silicon chip to a metallized 3D-printed component. We detail the fabrication of the coplanar waveguide chip and three-dimensional waveguide structure, simulations and metrology of a test device, and room temperature reflectance measurements of this device. The room temperature metrology and reflection measurements are consistent with a model that predicts a coupling efficiency of $\mathord{\sim} 95\%$ at cryogenic temperatures in the 85-170 GHz frequency range.

Stover, Austin [Chicago U.; Chicago U., KICP] (ORC↗
DOE OSTI2026

Optimizing Heat Recovery with Storage: Control Validation and Sensitivity Analysis of the Time-Independent Energy Recovery Plant Using Modelica

Heat recovery in large building central plants saves energy but traditionally requires simultaneous heating and cooling. The Time-Independent Energy Recovery (TIER) plant shifts this paradigm by integrating thermal energy storage (TES) to enable heat recovery regardless of concurrent demand, offering a highly efficient, space-saving solution to achieve California’s energy goals. However, its integration of heat recovery chillers, cooling-only chillers, cooling towers, and trim air-source heat pumps (ASHPs) creates growing control and sizing complexity. To overcome this, this study employs high-fidelity Modelica dynamic simulation to validate TIER control sequences and optimize equipment sizing. We translated the written Sequences of Operation into executable Control Description Language (CDL) to test logic against sub-hourly loads. This verification workflow successfully identified and resolved critical vulnerabilities, such as thermal storage freezing and equipment short-cycling, in a virtual environment prior to physical deployment. Then, the study analyzes TIER plant performance across three simulated building types in three locations, and a real building load profile, ensuring variety in heating and cooling loads, and simultaneity factors and explores sizing rules for the TES and ASHP capacity. The analysis shows that the TIER plant operates equipment efficiently leading to a plant SCOP of around 7.5 across all scenarios, higher than a traditional ASHP plant, and a viable pathway to de-risk complex system design and control through simulation to identify optimal designs that maximize energy efficiency, minimize operational costs, and ensure robust operation in varied environmental conditions, thereby facilitating the broader adoption of such a solution for large buildings.

Zanetti, Ettore↗
DOE OSTI2025

Data-Driven Tailoring Optimization of Thermoset Polymers Using Ultrasonics and Machine Learning

Thermoset polymers are highly demanded for their structural robustness, thermal stability, and chemical resistance. Tailoring the properties of these polymers for high-performance applications is often preferred to designing brand-new polymers. However, the traditional destructive techniques used to characterize their properties as a function of manufacturing parameters are expensive and time-consuming. A novel non-destructive, data-driven method leveraging ultrasonics and machine learning techniques to tailor the properties of thermosets as a function of the manufacturing parameters is demonstrated. Thermoset epoxy samples with varying curing temperatures (15–40 °C) and curing agent amounts (±40%) were manufactured and tested. Their curing kinetics were monitored by determining the sound speed in the material in real time, while the longitudinal modulus of the samples was determined post-cure. Machine learning models were developed using a k-nearest neighbors algorithm. These models were implemented to predict the curing and final elastic properties using the manufacturing parameters, i.e., stoichiometry and curing temperature, and vice versa. Understanding and modeling how these parameters affect the cure kinetics and final properties will allow for efficient and reliable optimization of thermoset tailoring and manufacturing.

36 MATERIALS SCIENCE↗
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↗