Vertical gradient in atmospheric particle phase state: a case study over the alaskan arctic oil fields
Explore the source record for details and available documents.
54 ENVIRONMENTAL SCIENCES↗
SEARCH · Search NASA
Results for “gradient”
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.
RCEMIP-II: mock-Walker simulations as phase II of the radiative–convective equilibrium model intercomparison project
Abstract. The radiative–convective equilibrium (RCE) model intercomparison project (RCEMIP) leveraged the simplicity of RCE to focus attention on moist convective processes and their interactions with radiation and circulation across a wide range of model types including cloud-resolving models (CRMs), general circulation models (GCMs), single-column models, global cloud-resolving models, and large-eddy simulations. While several robust results emerged across the spectrum of models that participated in the first phase of RCEMIP (RCEMIP-I), two points that stand out are (1) the strikingly large diversity in simulated climate states and (2) the strong imprint of convective self-aggregation on the climate state. However, the lack of consensus in the structure of self-aggregation and its response to warming is a barrier to understanding. Gaining a deeper understanding of convective aggregation and tropical climate will require reducing the degrees of freedom with which convection can vary. Therefore, we propose phase II of RCEMIP (RCEMIP-II) that utilizes a prescribed sinusoidal sea surface temperature (SST) pattern to provide a constraint on the structure of convection and move one critical step up the model hierarchy. This so-called “mock-Walker” configuration generates features that resemble observed tropical circulations. The specification of the mock-Walker protocol for RCEMIP-II is described, along with example results from one CRM and one GCM. RCEMIP-II will consist of five required simulations: three simulations with the same three mean SSTs as in RCEMIP-I but with an SST gradient and two additional simulations at one of the mean SSTs with different values of the SST gradients. We also test the sensitivity to the imposed SST gradient and the domain size. Under weak SST gradients, unforced self-aggregation emerges across the entire domain, similar to what was found in RCEMIP. As the SST gradient increases, the convective region narrows and is more confined to the warmest SSTs. At warmer mean SSTs and stronger SST gradients, low-frequency variability in the convective aggregation emerges, suggesting that simulations of at least 200 d may be needed to achieve robust equilibrium statistics in this configuration. Simulations with different domain sizes generally have similar mean statistics and convective structures, depending on the value of the SST gradient. The prescribed SST boundary condition is the only difference in the set-up between RCEMIP-II and RCEMIP-I, which enables comparison between the two; however, we also welcome participation in RCEMIP-II from models that did not participate in RCEMIP-I.
Geology↗
Dynamic Excitonic Beam Switching with Atomically‐Thin Binary Blazed Gratings
Abstract Beam steering metasurfaces are ultra‐compact optical coatings that offer on‐demand redirection of optical power to specific diffraction orders. To achieve this, spatial gradients are commonly introduced in the phase of light scattered by plasmon or Mie resonant nanoparticles within the metasurface grating's unit cell. However, these phase gradients are oftentimes difficult to tune post‐fabrication. Recently, excitons in monolayer 2D semiconductors have emerged as a new metasurface building block, due to their strong and electrically‐tunable resonant light‐matter interaction. These 2D excitonic metasurfaces offer the tantalizing prospect of beam switching within a single monolayer. Here, it is demonstrated how the 2D analog of binary blazed gratings enables such beam switching by mere nanopatterning of a large monolayer WS 2 , even though nanoscale ribbons of WS 2 do not support geometrical resonances. By introducing a gradient in the nanoribbon width within the metasurface unit cell, an amplitude gradient combined with a small phase gradient in the scattered fields results in asymmetric diffraction efficiencies. Using a scattered‐field analysis, it is shown that these gradients can be further engineered via interference effects with the substrate reflection. Finally, the electrical tunability of the exciton resonance is leveraged to achieve selective and dynamic beam switching with an atomically‐thin metasurface.
Guarneri, Ludovica [Van der Waals‐Zeeman Institute↗
Simulation of triggering and evolution of ELM by pellet injection in EAST under BOUT++ framework
A BOUT ++ three-field magnetohydrodynamic model is employed to study the triggering and evolution of edge localized mode (ELM) by Li pellets injected along the outer mid-plane in the EAST configuration. The linear simulation shows that compared with a large deposition on the pedestal top (scenario I), a smaller deposition within the steep-gradient pedestal region (scenario II) can stimulate much larger linear growth rates of all-n peeling-ballooning modes (PBMs). The nonlinear simulation shows that there exists a pellet size threshold for ELM triggering for two deposition locations; the threshold for scenario I predicted in the present study matches the EAST observation well. Comparison of the two scenarios reveals that a smaller deposition is sufficient to trigger an ELM in a much shorter time in scenario II, whose ELM size is comparable to that in scenario I. This conclusion confirms previous DIII-D and ASDEX-Upgrade observations, suggesting that the steep-gradient pedestal region is a favorable deposition location for ELM triggering with minimum pellet size. Simulation analyses also find that the positive radial gradient of the hump-like pressure profile in the outer mid-plane induced by the pellet deposition plays a different role in the two scenarios. In scenario I, the force resulting from the gradient hinders the outflow of core plasmas and in return, the perturbation is suppressed from spreading inwards after ELM crashes. In scenario II, with a sizable deposition, the gradient results in another competitive perturbation growth region during the linear phase, thus dispersing the free energy and reducing the efficiency of destabilizing PBMs by pellet injection. The suppressing effect of saturated zonal flow on other modes, the short ELM fast crash phase, and the restricting transport effect of the positive radial pressure gradient work together to constrain the pedestal energy loss, especially when the pellet deposition amount is high.
70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
TRANSFER LEARNING FOR FIELD EMISSION MITIGATION IN CEBAF SRF CAVITIES
The Continuous Electron Beam Accelerator Facility (CEBAF) at Jefferson Lab operates hundreds of super-conducting radio frequency (SRF) cavities in its two linear accelerators (linacs). Field emission (FE) is an ongoing operational challenge in higher gradient SRF cavities. FE generates high levels of neutron and gamma radiation leading to damaged accelerator hardware and a radiation hazard environment. During machine development periods, we performed gradient scans to record data capturing the relationship between cavity gradients and radiation levels measured throughout the linacs. However, the field emission environment at CEBAF varies considerably over time as the configuration of the radio frequency (RF) gradients changes and due to the changing behaviour of field emitters. An artificial intelligence/machine learning (AI/ML) approach with transfer learning could be a valuable tool to mitigate FE and lower the radiation levels. In this work, we mainly focus on leveraging the RF trip data gathered during CEBAF operations. We develop a transfer learning-based surrogate model for radiation detector readings given RF cavity gradients to track the CEBAF?s changing configuration and environment. Then, we could use the developed model as an optimization process for redistributing the RF gradients within a linac to minimize radiation levels.
Ahammed, K.↗
TRANSFER LEARNING FOR FIELD EMISSION MITIGATION IN CEBAF SRF CAVITIES
The Continuous Electron Beam Accelerator Facility (CEBAF) at Jefferson Lab operates hundreds of super-conducting radio frequency (SRF) cavities in its two linear accelerators (linacs). Field emission (FE) is an ongoing operational challenge in higher gradient SRF cavities. FE generates high levels of neutron and gamma radiation leading to damaged accelerator hardware and a radiation hazard environment. During machine development periods, we performed gradient scans to record data capturing the relationship between cavity gradients and radiation levels measured throughout the linacs. However, the field emission environment at CEBAF varies considerably over time as the configuration of the radio frequency (RF) gradients changes and due to the changing behaviour of field emitters. An artificial intelligence/machine learning (AI/ML) approach with transfer learning could be a valuable tool to mitigate FE and lower the radiation levels. In this work, we mainly focus on leveraging the RF trip data gathered during CEBAF operations. We develop a transfer learning-based surrogate model for radiation detector readings given RF cavity gradients to track the CEBAF?s changing configuration and environment. Then, we could use the developed model as an optimization process for redistributing the RF gradients within a linac to minimize radiation levels.
Ahammed, K.↗
Ambient ion focusing from a field-free region to a detector: enhanced signal for explosives and drug detection with mass spectrometry
This study demonstrates ion focusing at ambient pressure and increased ion signal by creating a voltage gradient from a field-free region to a detector, thereby improving the detection of chemicals, such as explosives and drugs. At ambient pressure, ion loss and resulting signal reduction pose challenges that limit detection sensitivity in analytical instruments. Techniques to increase sensitivity, such as atmospheric flow tube-mass spectrometry (AFT-MS), extend ion-molecule reaction times but result in significant overall ion loss due to diffusion. Ion manipulation techniques, though challenging at ambient pressure, can mitigate these losses by concentrating ions toward the detector inlet. Using SIMION, ion trajectories were modeled with a voltage gradient applied between a flow tube and a detector, revealing ion focusing at ambient pressure. Experimental verification with an atmospheric flow tube employed both mass spectrometry and Faraday plate detectors to measure ion beam profiles across varying flow rates, tube diameters, and voltage gradients. Application of a voltage gradient effectively directed ions to the axial center of the flow tube, narrowed ion beam width, and increased signal intensity by 5 to 10 times compared to conditions without a voltage gradient. This ion focusing approach shows promise for improving sensitivity in ambient-pressure instruments. This technique has the potential to enhance detection levels in security and forensic applications, with particular benefits for field-portable devices used at checkpoints to identify explosives and drugs.
ambient pressure↗
A real-time distributed solid oxide electrolysis cell (SOEC) model for cyber-physical simulation
System integration and dynamic operability between SOEC and balance-of-plant (BoP) components are major technical challenges before realizing rapid load following of SOEC systems. Cyber-physical simulation (CPS) is a leading-edge digital engineering approach and is regarded as the next step beyond Digital Twins. CPS approach can be used to research SOEC system integration and develop dynamic controls prior to actual pilot testing without using a real SOEC. To seamlessly couple with BoP hardware and access non-observable operational parameters (e.g., local temperature gradient) during transients, a distributed one-dimensional (1D) real-time SOEC model was developed. Its real-time execution was demonstrated for 20 to 640 nodes at the fixed time step of 5 ms. A higher excess air ratio enabled smaller local temperature gradients on SOEC solid materials and faster transients upon current density step change from 0.15 to 0.55 A cm -2 . During the transients, the magnitude of the peak temperature gradient nearly doubled in 10 s from -3.5 to -5.9 °C cm -1 . This represents a significant operating risk that can impact the dynamic operability of SOEC systems. In addition, the local temperature gradient was found to change directions on all nodes in SOEC solid materials, with the greatest impact on the upstream nodes. The SOEC model was also tested at the thermal neutral voltage using actual process air flow parameters as variable model inputs. Variable process air temperatures were found to induce alternating local temperature gradients on SOEC solid materials. These are new operational mechanisms for SOEC degradation relevant for load following operational modes yet distinct from previous reports. To mitigate these unfavorable features, the SOEC can be operated at voltages that are slightly (±20 mV) deviated from the thermal neutral voltage. Here, the corresponding net thermal energy change was less than 1.6% of the electric power consumption. This 1D real-time SOEC model established the basis of cyber-physical simulation of SOEC hybrid systems.
24 POWER TRANSMISSION AND DISTRIBUTION↗
Comparative analysis of plasticity-based GND density estimation methods in crystal plasticity finite element models
In crystal plasticity finite element (CPFE) simulations, accurately quantifying geometrically necessary dislocations (GNDs) is critical for capturing strain gradients in polycrystals. We compare different methods for quantifying GNDs, all of which originate from the Nye tensor, which is computed as the curl of the plastic deformation gradient. The projection technique directly decomposes the Nye tensor onto individual screw and edge dislocation components to compute GNDs. This approach requires converting a nine-component Nye tensor into densities for a larger number of dislocation systems, a fundamentally underdetermined (non-unique) process, which is resolved using L2 minimization. In contrast, when employing CPFE analysis, one could directly compute dislocation densities on each slip system using shear gradients. Projection and slip gradient methods are compared with respect to their prediction of GNDs with changing grain size, strain, and grain neighborhoods, including multigrain junctions. Although these techniques match analytical GND densities for single slip, single crystal deformation, and are consistent with anticipated overall GND trends, we find that the GND densities from projection techniques are significantly lower than those predicted from CPFE-based slip gradients in polycrystals. A suggested improvement of only using the active dislocation systems in the projection technique almost entirely resolved this mismatch.
Crystal plasticity↗
Impact of molten salt inflow on the temperature distribution in thermal energy storage tanks at startup for central receiver concentrating solar power plants
Concentrating Solar Power (CSP) systems with molten salt thermal energy storage (TES) tanks are one of the most promising, renewable-based energy conversion technologies for larger-scale power generation. The TES tank is one of the most critical components in CSP plants due to its high-temperature operation (up to 565 °C), daily thermal cycling, and intermittent solar radiation conditions. The plant startup is one of the most challenging operation conditions that could lead to damaging thermal gradients due to low salt inventory levels. In this study an analytical model for the sparger ring was developed and integrated with a detailed computational fluid dynamics model of a commercial-scaled molten salt tank. The integrated model allows an accurate representation of the tank operation to evaluate the effect of molten salt inflow on the mixing process. The tank filling process during plant startup was analyzed considering sparger rings with variations in design features, including inlet orifice configurations, number of orifices, direction of the inlets, and orifice diameter. The results demonstrated that higher temperature gradients are obtained in the tank floor during the plant startup. A sparger ring configuration with a predetermined orifice inlet inclination (30°, 45° and 60°) leads to significant temperature differences in the floor, between 57 °C and 62 °C, but better homogeneity in the temperature of the salt inventory. Lower salt inflow velocities result in a more homogeneous floor temperature, with maximum temperature differences under 38 °C. The sparger ring configuration with 52 orifices of 1-in. diameter and vertical flow showed better homogeneity in the temperature differences as a function of the salt level and lower temperature gradients in the tank floor. Because large temperature gradients in the tank's floor have been identified as one of the main factors contributing to tank failures, assessing various sparger ring design features is fundamental to determining proper inflow conditions that lead to low-temperature gradients and reducing failure susceptibility.
14 SOLAR ENERGY↗
Numerical modeling of plasma assisted deflagration to detonation transition in a microscale channel
Here, this work numerically studies the plasma assisted deflagration to detonation transition (DDT) of H 2 /O 2 mixtures in a microscale channel with detailed chemistry and transport. The results show that the DDT onset time is non-monotonically dependent on the discharge pulse number. The DDT is accelerated with small pulse numbers, whereas retarded with large ones. Two different DDT regimes, respectively at a small and large plasma discharge number, via acoustic choking of the burned gas and plasma-enhanced reactivity gradient without acoustic choking, are observed. Without plasma discharge, pronounced pressure and temperature gradients in front of the flame are generated by acoustic compression after the choking of the burned gas, triggering DDT via autoignition. With small plasma pulse numbers, the plasma-generated species enhance the ignition kinetics and lead to an increased reactivity in the boundary layer. After the choking of the burned gas, the plasma-enhanced reactivity advances the sequence of autoignition near the wall, strengthens ignition-shock wave coupling, and accelerates DDT. However, with a large discharge pulse number, a direct autoignition initiating DDT can occur without the acoustic choking of the burned gas due to the strongly accelerated reactivity and elevated temperature. In this case, DDT onset is retarded because the elevated temperature increases sonic velocity and the increased reactivity accelerates fuel oxidation in front of the flame, decelerating the formation of a leading shock and subsequent pressure buildup ahead of the flame. The present modeling reveals that no matter with or without plasma discharge, DDT is initiated by autoignition in thermal, pressure, and reactivity gradient fields via the Zel'dovich gradient mechanism. The acoustic choking of the burned gas may not be the necessary condition of DDT with strong plasma-enhanced reactivity gradient. This work provides an answer to the experimentally observed non-monotonic DDT onset time by plasma, which provides guidance to control DDT in advanced detonation engines and fire safety of hydrogen-fueled catalytic reactors in microchannels by non-equilibrium plasma discharge.
33 ADVANCED PROPULSION SYSTEMS↗
Evaluating methods to reduce duration of near-threshold fatigue crack growth rate measurements for low-alloy steels in hydrogen gas
Measurement of the near-threshold fatigue crack growth rate (da/dN) vs. stress-intensity factor range (∆K) relationship in hydrogen gas is essential for maximizing the calculated design fatigue life of high-pressure hydrogen storage vessels. However, such measurements are rarely performed, since the low cyclic loading frequencies applied in standard practice lead to prohibitively protracted test durations. The objective of this study was to demonstrate two means for reducing test durations when measuring near-threshold da/dN vs. ΔK relationships under decreasing ΔK for low-alloy pressure vessel steels in hydrogen gas: 1) imposing steeper K-gradients relative to the recommended limits in standards such as ASTM E647, and 2) increasing cyclic loading frequency relative to typical values applied during fatigue crack growth testing of low-alloy steels in hydrogen gas. Recognizing that steeper K-gradients could amplify loading-history effects, test methods employing this approach were designed to mitigate such effects by either maintaining constant K max or gradually increasing the K-gradient as the threshold was approached. Although the varying K-gradient method was vulnerable to loading-history effects in the form of plasticity-induced crack closure, particularly at lower stress ratio (R) and higher starting K max values, these effects could be compensated by applying the adjusted compliance ratio (ACR) method. Here, it was demonstrated that steeper K-gradients in concert with increased cyclic loading frequency reduced the duration of near-threshold fatigue crack growth tests in hydrogen gas by more than 99% relative to standard practices.
Fatigue threshold↗
Molecular concentration field design using closed-form steady-state solutions
Control over spatial concentration fields represents a fundamental challenge in designing synthetic biological systems and programmable soft materials. While nature creates morphogen gradients that orchestrate complex developmental processes, synthetic approaches have largely relied on empirical optimization and computationally intensive simulations. Here, we present an analytical framework for steady-state concentration fields generated by finite-sized localized sources in diffusion–degradation systems and derive closed-form solutions for one-, two-, and three-dimensional geometries. By expressing these solutions in dimensionless form, we show that gradient steepness and spatial structure are organized by the Thiele modulus, which captures the competition between diffusion and degradation length scales. The analysis reveals distinct design regimes: in degradation-dominated systems, gradient shape is governed by exponential decay and becomes dimension-independent, whereas in diffusion-dominated systems, gradient magnitude and extent follow dimension-dependent power-law scaling. Building on these results, we introduce a quantitative design strategy that uses threshold-based criteria to program concentration ranges by tuning physically accessible parameters, most directly the production rate, while holding transport and degradation properties fixed. Comparisons with numerical solutions and reported experimental systems demonstrate consistency with the predicted scaling behavior. Together, this work provides a generalizable and physically transparent framework for designing steady-state concentration fields in synthetic biological and soft matter systems, enabling predictive control of gradient-mediated organization without reliance on extensive numerical optimization.
Kim, Dong Woo [Johns Hopkins University, Baltimore↗
Turbulence suppression at extreme plasma densities on DIII-D and EAST
Recent high-poloidal-beta (high-βP) experiments on DIII-D and EAST have made coordinated breakthroughs for high confinement quality at high density near the Greenwald limit. Density gradient amplification of turbulence suppression at high βP can explain both of these achievements. Experiments on DIII-D have achieved Greenwald fraction (fGr = line-averaged density/Greenwald density) above 1 simultaneously with normalized energy confinement (H98y2) around 1.5, as required in fusion reactor designs but never before verified in tokamak experiments with the divertor configuration. A synergy between increased H98y2 and fGr is observed with strong gas puffing, due to the build-up of an internal transport barrier at large radius in the temperature and density channels. Transport simulations reveal that the favorable trend of reduced turbulent energy transport at higher density is only expected when increasing the density gradient at high local safety factor and high β, thus at high βP to ensure strong α-stabilization. These conditions are crucial to many conceptual designs for steady-state reactors. New experiments on EAST have nearly doubled the ion temperature at fGr ∼ 0.9, consistent with predict-first modeling results based on the same physics revealed from the DIII-D analysis. All previous EAST long-pulse H-modes have Ti ≪ Te near plasma axis. Transport modeling indicates that the profiles are limited by ion-temperature-gradient modes at mid-radius. The modeling also suggested potential solutions, including reducing magnetic shear, enhancing density gradients, and higher impurity concentration. Following this guidance, EAST experiments directly show a strong enhancement of Ti achieved with a combination of a second plasma current ramp-up, a density gradient increase, and a Zeff perturbation by a short pulse (100 ms) of impurity injection, as predicted by the earlier modeling.
70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
Exploring the transition from continuous turbulence fluctuations to bursting ELMs in high SOL density regimes
BOUT++ turbulence simulations of the DIII-D reveal that the density profile between the separatrix and pedestal plays a crucial role in the dynamics of edge localized modes (ELMs) and edge plasma turbulent transport. Nonlinear simulations demonstrate that small ELMs in the DIII-D hybrid scenario under high SOL density conditions are predominantly driven by local ballooning modes near the separatrix, stabilizing global instabilities while enhancing localized pressure fluctuations. A key control parameters for ELM dynamics is the separatrix-to-pedestal density ratio, n e,sep /n e,ped . A high ratio indicates a shallow gradient, favoring small ELMs, while a lower ratio signals a steep gradient, which increases the likelihood of large ELMs. Comprehensive parameter scans, including n e,sep /n e,ped , density gradient profiles near the separatrix, and resistivity, reveal the critical role of these parameters in shaping transitions between turbulence-driven transport and ELM bursting. The scans demonstrate that in high SOL density regimes, small ELMs can result from either global resistive MHD instabilities or local ballooning modes near the separatrix, depending on the steepness of the separatrix density gradient. These findings also highlight the transition from continuous turbulence to small ELMs. The post-crash peak in pressure fluctuations, δP rms serves as a critical metric for identifying transition from continuous turbulence fluctuations to ELM bursting. Larger δP rms values correlate with ELM bursts driven by local or global instabilities, whereas smaller values indicate turbulence-dominated transport. Drift-Alfvén and resistive ballooning turbulence enhance the entrainment of fluctuations from the pedestal to the SOL, contributing to the complex interplay of dynamics in this regime. These findings emphasize the importance of separatrix density shaping and pedestal gradient control for optimizing ELM behavior in ITER and future fusion devices.
Li, Nami [Lawrence Livermore National Laboratory (↗
Random coordinate descent: A simple alternative for optimizing parameterized quantum circuits
Variational quantum algorithms rely on the optimization of parameterized quantum circuits in noisy settings. The commonly used back-propagation procedure in classical machine learning is not directly applicable in this setting due to the collapse of quantum states after measurements. Thus, gradient estimations constitute a significant overhead in a gradient-based optimization of such quantum circuits. This paper introduces a random coordinate descent algorithm as a practical and easy-to-implement alternative to the full gradient descent algorithm. This algorithm only requires one partial derivative at each iteration. Motivated by the behavior of measurement noise in the practical optimization of parameterized quantum circuits, this paper presents an optimization problem setting that is amenable to analysis. Under this setting, the random coordinate descent algorithm exhibits the same level of stochastic stability as the full gradient approach, making it as resilient to noise. The complexity of the random coordinate descent method is generally no worse than that of the gradient descent and can be much better for various quantum optimization problems with anisotropic Lipschitz constants. Theoretical analysis and extensive numerical experiments validate our findings. Published by the American Physical Society 2024
Ding, Zhiyan (ORCID:000000018863403X)↗
Scalar flux transport models for self-similar turbulent mixing
A common approach to closing turbulent species flux in multicomponent Reynolds-averaged Navier-Stokes models is to use the standard gradient diffusion approximation. While such an approach has been shown to work well when applied to many canonical turbulent mixing configurations, a gradient diffusion approach is fundamentally limited in its ability to capture complex phenomena such as countergradient transport. For this reason, complicated mixing applications may benefit by treating the turbulent diffusivity with a model transport equation in a manner analogous to second-moment momentum closure in Reynolds-stress transport models. Here, the present work explores the development and application of two different scalar flux transport (SFT) models. Self-similarity constraints are derived for these models, and they are evaluated against gradient-diffusion-based models in several one- and two-dimensional problems of turbulent mixing. It is found that the new SFT models out-perform gradient diffusion models in problems involving rapid acceleration reversal and in problems involving anisotropic transport of materials. In addition, it is found that even a hybrid-SFT approach, in which an SFT equation is utilized along with a gradient diffusion closure, provides some measure of improvement over models that transport the mass flux rather than the scalar flux.
Reynolds-averaged Navier Stokes↗
FedCSpc: A Cross-Silo Federated Learning System With Error-Bounded Lossy Parameter Compression
Cross-Silo federated learning is widely used for scaling deep neural network (DNN) training over data silos from different locations worldwide while guaranteeing data privacy. Communication has been identified as the main bottleneck when training large-scale models due to large-volume model parameters and gradient transmission across public networks with limited bandwidth. Most previous works focus on gradient compression, while limited work tries to compress parameters that can not be ignored and extremely affect communication performance during the training. Here, to bridge this gap, we propose FedCSpc: an efficient cross-silo federated learning system with an XAI-driven adaptive parameter compression strategy for large-scale model training. Our work substantially differs from existing gradient compression techniques due to the distinct data features of gradient and parameter. The key contributions of this paper are fourfold. (1) Our designed FedCSpc proposes to compress the parameter during the training using the state-of-the-art error-bounded lossy compressor – SZ3. (2) We develop an adaptive compression error bound adjustment algorithm to guarantee the model accuracy effectively. (3) We exploit an efficient approach to utilize the idle CPU resources of clients to compress the parameters. (4) We perform a comprehensive evaluation with a wide range of models and benchmarks on a GPU cluster with 65 GPUs. Results show that FedCSpc can achieve the same model accuracy as FedAvg while reducing the data volume of parameters and gradients in communication by up to 7.39× and 288×, respectively. With 32 clients on a 4 Gb size model, FedCSpc significantly outperforms FedAvg in wall-clock time in the emulated WAN environment (at the bandwidth of 1 Gbps or lower without loss of generality).
SZ3↗