Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “hierarchical models”

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 289 records · Page 16
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Multiscale Flow for robust and optimal cosmological analysis

We propose Multiscale Flow, a generative Normalizing Flow that creates samples and models the field-level likelihood of two-dimensional cosmological data such as weak lensing. Multiscale Flow uses hierarchical decomposition of cosmological fields via a wavelet basis and then models different wavelet components separately as Normalizing Flows. The log-likelihood of the original cosmological field can be recovered by summing over the log-likelihood of each wavelet term. This decomposition allows us to separate the information from different scales and identify distribution shifts in the data such as unknown scale-dependent systematics. The resulting likelihood analysis can not only identify these types of systematics, but can also be made optimal, in the sense that the Multiscale Flow can learn the full likelihood at the field without any dimensionality reduction. We apply Multiscale Flow to weak lensing mock datasets for cosmological inference and show that it significantly outperforms traditional summary statistics such as power spectrum and peak counts, as well as machine learning–based summary statistics such as scattering transform and convolutional neural networks. We further show that Multiscale Flow is able to identify distribution shifts not in the training data such as baryonic effects. Finally, we demonstrate that Multiscale Flow can be used to generate realistic samples of weak lensing data.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2026

A bespoke model of Arctic river basins based on hillslope delineation: Model Archive

This dataset is a model archive of the paper A bespoke model of Arctic river basins based on hillslope delineation (in prep), which introduces a watershed decomposition and parameterization method for large scale permafrost hydrology simulation. With this dataset, this study aims to address the research question: whether a computationally efficient hillslope-based modeling framework can reliably simulate discharge at Arctic river-basin scales. This dataset contains model input and output data for five modeling scenarios at a study site located in the Sagavanirktok River basin. The five modeling scenarios include three modeling cases under temperate conditions using full 3D, decomposed 3D, and decomposed 2D modeling strategies; and two modeling cases under actual Arctic conditions with permafrost using full 3D and decomposed 2D modeling strategies. Simulations were performed using the Advanced Terrestrial Simulator (ATS, v1.6 for three temperate scenarios and v1.5 for two Arctic scenarios), a physics-rich integrated surface–subsurface hydrologic model with cryo-hydrology features. For the three temperate models, simulations were conducted for the period of 10/01/1993 - 09/30/2002; and for the two Arctic models, simulations were conducted for the period of 01/01/1994 - 12/31/2002. To facilitate reproducibility of simulations, all datasets are organized hierarchically. The dataset contains: (1) Mesh files (.exo) for full 3D model, decomposed 3D models, and decomposed 2D models, located in huc/190604020802_gauge15906000/mesh/. Mesh files can be visualized through Paraview or read by Python. (2) Climate forcings (.h5) for full 3D model and decomposed 3D/2D models are located in huc/190604020802_gauge15906000/daymet_onePiece/, and huc/190604020802_gauge15906000/vp_pr_revised_daymet_1980_2006_with_wind/ separately. Accessible by Python. (3) Raw measured gage discharge (.csv) from USGS, located in huc/190604020802_gauge15906000/gaged_basin15906000_discharge_usgs/. Accessible by Python. (4) Delineated subdomain raster (.tif) and shape files (.shp), and the final parameterized results (.npy) for decomposed models, located in huc/190604020802_gauge15906000/data_preprocessed-meshing. Accessible by Python. (5) Temperate models are located in nonpermaf_huc190604020802_gauge15906000/, which includes three cases: decomposed 2D models (inside model_0*-hillslope_*), decomposed 3D models (inside model_1*-subcatchment_*), and full 3D model (inside model_2*-onepiece_*). Two step spin-up results (checkpoint_final.h5) are located in model_*1-*_spinup_steadystate and model_*2-*_spinup_cycle, separately, which are used to initialize real transient models. The input files (.xml) and output results (.dat) of the real transient models are located in model_*3-*_transient/. Especially, for two example hillslope models (ID=-11 and 11), additional h5py files are included in model_03-hillslope_transient/hillslope-11/, model_03-hillslope_transient/hillslope11, model_13-subcatchment_transient/subcatchment-11/, model_13-subcatchment_transient/subcatchment/11, respectively, which are used to plot the saturation figure (Figure 5) in the manuscript. Accessible by Python. (6) Arctic models are located in huc190604020802_gauge15906000/, which includes two cases: decomposed 2D models (inside model_04-hillslope_transient), and full 3D model (inside model_05-onepiece_transient_mannp1_ra). Three step spin-up results (checkpoint_final.h5) are located in model_01-column_freezeup/, model_02-column_spinup/, model_03-hillslope_spinup/, respectively, which are used to initialize real 2D transient hillslope models. The input files (.xml) and output results (.dat) of transient 2D hillslope models are located in model_04-hillslope_transient/. The input files (.xml) and output results (.dat) of the full 3D transient model is located in model_05-onepiece_transient_mannp1_ra/. The full 3D transient model is initialized by model_02-column_spinup/. Accessible by Python. (7) The MOSART routed discharge results (.csv) under Arctic conditions is located in huc190604020802_gauge15906000/MOSART/. Accessible by Python. (8) All Python codes (.py) used to parameterize full 3D model to decomposed 2D models are located in script/. These codes fit with watershed workflow (a watershed delineation tool) v1.4 under the branch gaob/v1.4 from https://github.com/gaobhub/watershed-workflow.git.

EARTH SCIENCE > CRYOSPHERE↗
DOE OSTI2024

Metabolic complexity drives divergence in microbial communities

Microbial communities are shaped by environmental metabolites, but the principles that govern whether different communities will converge or diverge in any given condition remain unknown, posing fundamental questions about the feasibility of microbiome engineering. Here, in this work, we studied the longitudinal assembly dynamics of a set of natural microbial communities grown in laboratory conditions of increasing metabolic complexity. We found that different microbial communities tend to become similar to each other when grown in metabolically simple conditions, but they diverge in composition as the metabolic complexity of the environment increases, a phenomenon we refer to as the divergence-complexity effect. A comparative analysis of these communities revealed that this divergence is driven by community diversity and by the assortment of specialist taxa capable of degrading complex metabolites. An ecological model of community dynamics indicates that the hierarchical structure of metabolism itself, where complex molecules are enzymatically degraded into progressively simpler ones that then participate in cross-feeding between community members, is necessary and sufficient to recapitulate our experimental observations. In addition to helping understand the role of the environment in community assembly, the divergence-complexity effect can provide insight into which environments support multiple community states, enabling the search for desired ecosystem functions towards microbiome engineering applications.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2025

CSGL: chemical synthesis graph learning for molecule representation

Abstract Motivation Molecule representation learning (MRL) translates molecules into a real vector space, serving as input to downstream tasks in biology, chemistry, and computer science. This article introduces a chemical synthesis graph learning (CSGL) framework, which enhances MRL by considering both the atomic structures of molecules and their roles in chemical reactions through a hierarchical graph representation. Specifically, molecules are first modeled based on their molecular graphs, which capture atomic-level structural information. They are then further refined using a chemical synthesis graph, where nodes represent reactant and product molecule sets, and edges encode chemical transformations between reactants and products (e.g. changes in molecular structures). CSGL optimizes molecular embeddings of reactant and product nodes in a fashion that ensures the embeddings conform to a chemical balance constraint. Results Experimental results show that our method CSGL achieves strong performance on a variety of tasks, including product prediction, reaction classification, and molecular property prediction. Availability and implementation https://github.com/li-2023/CSGL.

Biochemistry & Molecular Biology↗
DOE OSTI2023

Integrated Design of Ultradurable, Low CO 2 Alternative Binder Systems via Machine Learning

This ARPA-E project developed a machine learning tool to use in formulation design of cementitious binders for concrete having 50% less embodied CO 2 and possessing twice the durability compared to concrete based on ordinary portland cement (OPC) binders. The technical focus was on limestone/calcined clay cement (LC3), the leading replacement for OPC. Here, hierarchical machine learning (HML) was used to model the flowability, set time, strength, and durability of LC3 concrete. This methodology identifies latent variables derived from domain knowledge and empirical models that develop an accurate model for a response surface from small datasets. For the flowability metric, particle packing was a dominant factor, while strength and durability were both strongly determined by the fraction of metakaolin and the water:solids ratio. Under constraints of water:binder ratio, material performance metrics, embodied CO 2 , and cost per tonne of OPC, multi-objective optimization was used to design binders parameterized by the mineral composition replacing OPC, particle size distributions, and water:solids ratio. The trained algorithm was able to predict multiple mixes met these performance criteria, and experimental testing validated the predictions. The model demonstrated here is relevant for North America, where pure kaolin deposits are found broadly. The approach is being taken forward into commercial application by Ansatz AI, a materials informatics company founded by PI Washburn and co-PI Poczos. Through collaborations with the cement and concrete industry, and funding from SBIR programs, a commercial software will be developed in future research.

36 MATERIALS SCIENCE↗
DOE OSTI2020

Using AI Simulations to Dynamically Model Multi-agent Multi-team Energy Systems

The complexity of energy systems is well known as they are complex and intricate systems. As a result, many extant studies have used many simplifications or generalizations that do not accurately reflect the nature of this complex system. In particular, most HVAC systems are modeled as a single unit, or several large units, rather than as a hierarchical composite (e.g., as a floor rather than as a collection of disparate rooms). The net result of this is that the simulations are too generic to perform meaningful analysis, machine learning, or integrated simulation. We propose using a multi-agent multi-team strategic simulations framework called SiMAMT to better define, model, simulate, and learn the HVAC environment. SiMAMT allows us to create distinct models for each type of room, hierarchically aggregate them into units (like floors, or sections), and then into larger sets (like buildings or a campus), and then perform a simulation that interacts with each sub-element individually, the teams of sub-elements collectively, and the entire set in aggregation. Further, and most importantly, we additionally model another ‘team’ within the simulation framework - the users of the systems. Again, each individual is modeled distinctly, aggregated into sub-sets, then collected into large sets. Each user, or agent, is performing on their own but with respect to the larger team goals. This provides a simulation that has a much higher model fidelity and more applicable results that match the real-world.

Franklin, D. Michael↗
DOE OSTI2021

Polymer Coatings on Virus-like Particle Nanoreactors at Low Ionic Strength—Charge Reversal and Substrate Access

Virus-like particles (VLPs) are a class of biomaterials which serve as platforms for achieving the desired functionality through interior and exterior modifications. Through ionic strength-mediated electrostatic interactions, VLPs have been assembled into hierarchically ordered materials. This work builds on predictive models to prepare polymer-coated VLP dusters at very low ionic strength. Zeta potential measurements showed that the clusters carried a strongly positive charge, a complete charge reversal from the VLP building block. SAXS analysis confirmed polymer adsorption onto the VLP exterior. Furthermore, we then studied the activity of an encapsulated enzyme toward small molecular and macromolecular substrates to determine the effect of each component of the hierarchically assembled material. We found that while encapsulation and polymer coating did not have a large effect on access to the enzyme by its native, small molecular substrate, substrate modification with a macromolecule caused the polymer coating and encapsulation to affect the access to the enzyme.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2022

Physically rigorous reduced-order flow models of fractured subsurface environments without explosive computational cost

Fractured media models comprise discontinuities of multiple lengths (e.g. fracture lengths and apertures, wellbore area) that fall into the relatively insignificant length scales spanning millimeter-scale fractures to centimeter-scale wellbores in comparison to the extensions of the field of interest, and challenge the conventional discretization methods imposing highly-fine meshing and formidably large numerical cost. By utilizing the recent developments in the finite element analysis of electromagnetics that allow to represent material properties on a hierarchical geometry, this project develops computational capabilities to model fluid flow, heat conduction, transport and induced polarization in large-scale geologic environments that possess geometrically-complex fractures and man-made infrastructures without explosive computational cost. The computational efficiency and robustness of this multi-physics modeling tool are demonstrated by considering various highly-realistic complex geologic environments that are common in many energy and national security related engineering problems.

42 ENGINEERING↗
DOE OSTI2022

Viscoelastic analysis of mussel threads reveals energy dissipative mechanisms

Mussels use byssal threads to secure themselves to rocks and as shock absorbers during cyclic loading from wave motion. Byssal threads combine high strength and toughness with extensibility of nearly 200%. Researchers attribute tensile properties of byssal threads to their elaborate multi-domain collagenous protein cores. Because the elastic properties have been previously scrutinized, we instead examined byssal thread viscoelastic behaviour, which is crucial for withstanding cyclic loading. By targeting protein domains in the collagenous core via chemical treatments, stress relaxation experiments provided insights on domain contributions and were coupled with in situ small-angle X-ray scattering to investigate relaxation-specific molecular reorganizations. Results show that when silk-like domains in the core were disrupted, the stress relaxation of the threads decreased by nearly 50% and lateral molecular spacing also decreased, suggesting that these domains are essential for energy dissipation and assume a compressed molecular rearrangement when disrupted. A generalized Maxwell model was developed to describe the stress relaxation response. The model predicts that maximal damping (energy dissipation) occurs at around 0.1 Hz which closely resembles the wave frequency along the California coast and implies that these materials may be well adapted to the cyclic loading of the ambient conditions.

36 MATERIALS SCIENCE↗
DOE OSTI2021

Calibration of the Diffusivity Predictions of Centipede Using Approximate Bayesian Computation and Applications in Nyx (Engineering Scale) and Xolotl-MARMOT (Meso-Scale) Simulations

Fission gas evolution and release in UO 2 nuclear fuel are important fuel performance metrics and occur in several distinct stages: 1) nucleation, growth and resolution of intra-granular bubbles, 2) diffusion to grain boundaries and 3) nucleation and growth of bubbles at grain boundaries, which eventually form a connected network (percolation) enabling release of gas from grain boundaries through connections to triple junctions, grain edges or free surfaces. The NE-SciDAC project is developing several computational tools to model this problem, which are connected in a hierarchical multi-scale framework. The information transfer in the multi-scale framework is a critical step that, in addition to best-estimates, should include uncertainty quantification. Despite taking a first-principles multi-scale approach, there is a need to perform parameter calibration to ensure consistency with available experimental data. In the present study, uncertainty quantification (UQ) and parameter calibration is demonstrated for one of the lower length scale codes in the multi-scale framework (Centipede) and then the results, including instances of the propagated uncertainties, are used in other codes within the framework, specifically Nyx and Xolotl-MARMOT. We calibrated the model parameters in Centipede, a computer code used to predict diffusivities of uranium (U) and xenon (Xe) in the context of the simulation of fission gas in uranium oxide (UO 2 ) nuclear fuel. The Centipede code depends on 183 parameters, all of which are subject to uncertainty. The three data sets used in our calibration effort are taken from the literature. This data is available as a set of measurements, including measurement errors. Our goal is to calibrate a statistical model that predicts both the value of the measurement and the uncertainty associated with the measurement. We perform a Bayesian calibration of the model parameters using a dedicated approximate Bayesian computation (ABC) likelihood function. To avoid excessive computational costs, we replace the expensive Centipede simulation code by a higher-order surrogate model, constructed using only the 9 most important parameters. These important parameters are identified by a preliminary global sensitivity analysis (GSA) study. Among the important parameters are T0 (the temperature at which UO 2 is perfectly stoichiometric) and Hf_pO2 (the temperature dependence of the oxygen (O) partial pressure) that should be considered as operating conditions to be estimated along with the other parameters. We consider two different cases: one where we define one set of these operating conditions for all data sets, and one where we define distinct operating condition parameters for each data set. The Xe diffusivities predicted by the latter case show distinct features that could not be observed in the former. Next, we use the diffusivity predictions by Centipede as input to Nyx, a reduced order fuel performance code focused on gas behavior alone, in order to estimate quantities associated with inter-granular bubble formation at conditions specified by the experiments. Finally, the diffusivities obtained from the calibrated Centipede runs were used in coupled Xolotl-MARMOT simulations of intra- and inter-granular gas evolution. The results are compared to simulations using the baseline diffusivities from Turnbull et al.

11 NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗
DOE OSTI2020

A Step Towards Machine Learning-based Coherent Generator Grouping for Emergency Control Applications in Modern Power Grid

A dynamic neural network (NN) based multi-class classifier is proposed for improving online prediction of coherent generator groups (CGGs), following the occurrences of various contingencies in the power grid. This is motivated by the increasing availability of the measurements from phasor measurement units (PMUs) and the number of grouping schemes is limited. The proposed method consists of three steps. First, by performing offline simulations, a library of system dynamic responses characterized by post-contingency rotor angles and speeds of individual generators is obtained. To generate sufficient data, up to N-2 contingencies and the uncertain parameters associated with the power grid including type and location of disturbance and fault clearing times are modeled. Secondly, the training data-set is produced by generating labels for individual contingencies using a hierarchical clustering method based on rotor angle and speed data. Lastly, the dynamic NN models are trained for online applications such as emergency controls and controlled islanding. The proposed method is tested on the standard 16-generator 68-bus system to demonstrate its performance. Furthermore, the impact of the sample data lengths on the CGG numbers is evaluated. It is interesting to observe that the time domain stability behaviors can be determined by examining the changes in the CGG numbers.

29 ENERGY PLANNING, POLICY, AND ECONOMY↗
DOE OSTI2022

Hierarchical Division and Clustering of Group Structures

When using multigroup neutron transport models, it is important to choose a suitable group structure due to the impact on runtime and accuracy. Group structures at LANL have been largely pared down to a handful of commonly used group structures, such as the LANL-30 and LANL-70 group structures. By identifying the key boundaries for a given problem, more insight can be gained into why the commonly used group structures are accurate and when they are inaccurate. This work focuses on using hierarchical division and agglomeration methods to identify the most critical boundaries in these group structures, using the ICSBEP HEU-MET-FAST-001-001 benchmark as a case study.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗
DOE OSTI2023

Adaptive, Active Learning, and Multifidelity Monte Carlo Methods in the MOOSE Stochastic Tools Module

MOOSE is an open-source computational platform for constructing multi-physics models and executing them in a massively parallel fashion. It has a stochastic tools module (STM) for forward/inverse uncertainty quantification (UQ) and surrogate modeling. This presentation details some recent developments to the STM with respect to the implementation of adaptive, active learning, and multifidelity Monte Carlo methods for forward UQ of computational models. Specifically, the adaptive Monte Carlo methods include Markov Chain Monte Carlo (MCMC)-driven algorithms like adaptive importance sampling and parallelized subset simulation for statistical QoI estimation, rare events analysis, and stochastic gradient-free optimization. The active learning methods include Gaussian Process (GP) surrogates and their training via Adam optimization, design of acquisition functions, and integration with samplers like Monte Carlo, adaptive importance, and parallelized subset simulation. These active learning methods are also designed to work in a batch mode, wherein, the required calls to the full computational model are executed in parallel whenever a user-specified batch size is met. The multifidelity methods in STM are broadly divided into two categories: hierarchical, where a defined hierarchy exists among the low-fidelity models, and peer, where all the low-fidelity models are treated equally. A GP surrogate is used to learn the differences between the low- and high-fidelity models in both multifidelity categories, and acquisition functions from the active learning classes are used to decide whether to rely on a low-fidelity model or call the expensive high-fidelity model. Alongside the software description and usage, applications are also presented to nuclear engineering computational models including a TRISO nuclear fuel particle, a reactor pressure vessel, and a heat-pipe microreactor.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Observed and Imputed Volumetric Soil Water Content Timeseries for the New Mexico Elevation Gradient

Reliable soil water content (SWC) data are essential for understanding dryland ecosystem dynamics, but high-frequency SWC sensors often fail, creating gaps in critical datasets. To address this, we developed a Bayesian mixture model that imputes missing SWC using both linear interpolation and an ecosystem water balance model (SOILWAT2), tested across six AmeriFlux eddy covariance tower sites in the New Mexico Elevation Gradient, demonstrating its effectiveness in reconstructing SWC patterns while providing insights into the factors driving SWC variability. Daily volumetric soil water content (SWC) data are provided as csv-formatted spreadsheets for the six AmeriFlux sites (US-Seg, US-Ses, US-Wjs, US-Mpi, US-Vcp, and US-Vcs). For each site there is an observed SWC file (site_SWC_gapfill.csv) and a file that contains imputed SWC (imputed_SWC_site.csv). The observed SWC files contain temperature corrected sensor values, tower precipitation data, as well as outputs from SOILWAT2 simulations that were used to impute SWC. The imputed files contain the original observed SWC values and the imputed missing SWC values. When SWC was missing from the original data, the missing value was imputed based on the Bayesian imputation mixture model. The posterior mean of all imputed values is reported as "mean_X". When the observed SWC was NOT missing, mean_X = observed SWC value (original data). The standard deviation, 2.5th percentile and the 97.5th percentile for the imputed values are also reported in the imputed files. There are readme text files for each file type explaining the contents of each column.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2024

TChem-atm v1.0

SAND2024-11300O TChem-atm is a software library that was developed to solve complex kinetic models for atmospheric chemistry applications. TChem-atm interface employs a hierarchical parallelism design to exploit the massive parallelism available from modern computing platforms. It also supports gas atmospheric chemistry applications, e.g., the energy exascale earth system model. TChem can be used as a box model or coupled with a climate model to compute the time evolution of gas tracer species. Sandia National Laboratories is a multimission laboratory managed and operated by National Technology & Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International Inc., for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-NA0003525.

Safta, Cosmin↗
DOE OSTI2021

Lattice Clifford fractons and their Chern-Simons-like theory

We use Dirac matrix representations of the Clifford algebra to build fracton models on the lattice and their effective Chern-Simons-like theory. As an example, we build lattice fractons in odd D spatial dimensions and their (D+1) spacetime dimensional effective theory. The model possesses an anti-symmetric K matrix resembling that of hierarchical quantum Hall states. The gauge charges are conserved in sub-dimensional manifolds which ensures the fractonic behavior. The construction extends to any lattice fracton model built from commuting projectors and with tensor products of spin-1/2 degrees of freedom at the sites.

Fontana, Weslei↗
DOE OSTI2021

Identification of crystal plasticity model parameters by multi-objective optimization integrating microstructural evolution and mechanical data

Crystal plasticity models evolve a polycrystalline yield surface using meso-scale descriptions of deformation mechanisms. The activation of deformation mechanisms is governed by crystallography and a set of model parameters, which are typically calibrated through the fitting of mechanical data such as stress–strain curves and elastic lattice strains. Microstructural data such as phase fractions and texture evolution are used for verifying crystal plasticity parameters. In this study, we use a multi-objective genetic algorithm to identify hardening parameters from flow stress curves with an option to incorporate texture into the optimization approach. Robust, generalized objective functions are developed and used to identify sets of parameters pertaining to dislocation density-based hardening laws in visco-plastic and elasto-plastic self-consistent (VPSC and EPSC) homogenization models. First, the parameters are identified for pure Nb directly from texture using an objective function based on generalized spherical harmonics. Since texture evolution is driven by the relative contribution of active slip systems, the parameters governing the evolution of slip resistance ratios can be recovered from fitting discrete textures at a series of strains. Next, a comprehensive set of load reversal data for dual phase (DP) 780 steel is used to fit a hardening law and a back-stress law in EPSC. Finally, parameters pertaining to a complex hardening law for the evolution of slip and twinning in pure α-Ti are identified. Remarkably, using texture as an objective in combination with stress–strain objectives constrains the model of Ti to fully reproduce not only stress–strain and texture evolution but also hierarchical twinning measurements as a function of initial grain size and texture. Furthermore, given an appropriate model fit to representative experimental texture evolution, underlying twin volume fractions contributing to texture evolution can be predicted.

42 ENGINEERING↗
DOE OSTI2026

DP-TwoLevel: two-stage gradient subspace learning for differentially private federated learning

Federated learning (FL) enables collaborative model training across distributed data sources without sharing raw data, but faces fundamental challenges in communication efficiency and privacy. Differentially private (DP) training mitigates information leakage but introduces noise that degrades model performance, especially in high-dimensional settings. We propose DP-TwoLevel, a hierarchical gradient projection method that improves utility under fixed DP constraints by exploiting low-dimensional structure in model updates. Our approach learns a two-level PCA-based representation of gradients and applies DP noise in a reduced-dimensional subspace, thereby lowering the effective noise magnitude while preserving dominant signal components. We evaluate the method across three datasets (MNIST, Fashion-MNIST, CIFAR-10) and three privacy regimes (ϵ∈0.5, 1.0, 2.0). Across nine experimental settings, DP-TwoLevel consistently outperforms DP-FedAvg, achieving an average accuracy improvement of 9.44%, with larger gains observed in lower ϵ(higher-noise) regimes (up to +22.31%). We further analyze scalability across models ranging from 100K to 1.49M parameters and identify a variance-based success criterion: performance remains strong when the projection preserves more than 75% of gradient variance, degrades in a marginal regime (65–75%), and fails below this threshold. Our results demonstrate that structure-aware dimensionality reduction can significantly improve the privacy–utility tradeoff in FL without modifying formal privacy guarantees. We also provide empirical evidence of scaling limitations for global projections and motivate per-layer extensions for larger models.

Kotevska, Olivera [ORNL] (ORCID:0000000316772243)↗