Physical Sciences and Mathematics Commons^™

Hold Paramount The Health, Safety, And Welfare Of The Public And The Planet, Daniel B. Oerther

Civil, Architectural and Environmental Engineering Faculty Research & Creative Works

No abstract provided.

Computational Characterization Of A Recently Identified Disinfection Byproduct In Drinking Water: The Chloronitramide Anion And Its Monohydrate Complex, Thufail M. Ismail, Max R. Tucker, Gregory S. Tschumper

Chemistry Faculty Research & Creative Works

The chloronitramide anion (ClNNO₂⁻) was recently identified as disinfection byproduct in drinking water, and various quantum chemistry methods (CCSD(T), MP2, B3LYP, ωB97XD, and M06-2X) are employed in this study to characterize this negatively charged species and its monohydrate complex. Both vertical and adiabatic quantities indicate the excess electron is bound by 3–4 eV. Four ClNNO₂⁻ monohydrate minima were identified, each exhibiting double ionic hydrogen bonding. The computed ClNNO₂⁻⋯H₂O interaction approaches 15 kcal/mol, and it induces pronounced vibrational frequency shifts in both fragments. Natural bond orbital analysis shows a redistribution of …

A Note On Sufficient Dimension Folding For Regression Mean Function With Categorical Predictors, Bilin Zeng, Akim Adekpedjou, Xuerong Meggie Wen

Mathematics and Statistics Faculty Research & Creative Works

Multi-dimensional arrays are referred to as tensors. Tensor-valued predictors are commonly encountered in modern biomedical applications, such as electroencephalogram (EEG), magnetic resonance imaging (MRI), functional MRI (fMRI), diffusion-weighted MRI, and longitudinal health data. In survival analysis, it is both important and challenging to integrate clinically relevant information, such as gender, age, and disease state along with medical imaging tensor data or longitudinal health data to predict disease outcomes. Most existing higher-order sufficient dimension reduction regressions for matrix- or array-valued data focus solely on tensor data, often neglecting established clinical covariates that are readily available and known to have predictive value. …

Electric Dipole Forbidden, Quadrupole Allowed Transitions In The Pure Rotational Spectrum Of Cyclopropylchloromethyldifluorosilane, Alexander R. Davies, Abanob G. Hanna, Alma Lutas, Gamil A. Guirgis, S. A. Cooke, Garry S. Grubbs

Chemistry Faculty Research & Creative Works

In a recent publication, some electric dipole forbidden, quadrupole allowed ΔJ = +2 and x-type transitions were observed in the chirped-pulse Fourier transform microwave spectrum of two conformations of cyclopropylchloromethyldifluorosilane. Many of these transitions arise from a handful of mixed states and mechanisms are proposed through which these transitions become weakly allowed. Observations of electric dipole forbidden, quadrupole allowed transitions in rotational spectra are unusual for molecules which contain a chlorine nucleus owing to the small quadrupole moment of ³⁵Cl and ³⁷Cl; thus, we believe we are the first to observe x-type transitions arising from perturbations caused by …

Leveraging Molecular Mechanisms Of Desorption To Enhance Pfas Bioavailability In Contaminated Soils, Husam Kafeenah, Margaret D. Taiwo, Patrica Ivy Agorsor, Michael O. Eze

Chemistry Faculty Research & Creative Works

Per- and polyfluoroalkyl substances (PFAS) are known for their strong binding properties to soil matrices owing to their amphiphilic properties. While most studies focus on the search for novel PFAS bio degraders, our knowledge of sustainable biomolecules that could drive PFAS desorption and make them more bioavailable is limited. This study investigated the effectiveness of rhamnolipids and organic acids in desorbing PFAS compounds from contaminated soil. Rhamnolipids (25 mg/L) significantly enhanced the release of PFAS compounds from soil, achieving up to 90 % desorption. Acetic acid provided 60–90 % desorption efficiency for most of the PFAS studied, except for PFDA …

Hybrid Data-Driven Cement-Stabilized Soil Design: An Integration Of Machine Learning, Multi-Objective Optimization, And Life Cycle Assessment, Chikezie Chimere Onyekwena, Yunli Li, Ikenna J. Okeke, Ubani Obinna Uzodimma, Monday Uchenna Okoronkwo, Wenping Wu

Chemical and Biochemical Engineering Faculty Research & Creative Works

Soil stabilization is crucial in geotechnical engineering, yet conventional methods are often time-consuming, resource-intensive, and environmentally unsustainable. Despite growing interest in Machine Learning (ML) and optimization tools for mix design, few studies integrate these methods with decision-making techniques and environmental assessment to support practical implementation. This study proposes a hybrid data-driven framework for predicting strength, optimizing mix compositions, and evaluating environmental impacts via life cycle assessment of cement-stabilized soft soils. Six ML models were evaluated, and the top-performing eXtreme Gradient Boosting (XGB) model was further improved using the Grey Wolf Optimizer (GWO). The optimized XGB-GWO model, integrated with a polynomial …

Performance And Mechanistic Insights Into Cement Systems Modified With Wastewater-Recovered Struvite, Ugochukwu Ewuzie, Rupack R. Halder, Abdulkareem O. Yusuf, Abiodun A. Saka, Godwin I. Ogbuehi, Titus C. Egbosiuba, Damilola A. Daramola, Monday Uchenna Okoronkwo

Chemical and Biochemical Engineering Faculty Research & Creative Works

Struvite, the stable hydration product and primary strength phase in magnesium ammonium phosphate cement (MAPC), derived from wastewater treatment, has recently been utilized as a sustainable additive to Portland cement (PC). However, its impacts on cement hydration kinetics, pore refinement, rheology, and the mechanisms underlying these processes have not been comprehensively studied. This study developed Portland cement-struvite (PCS) systems by replacing PC with 3–20 % struvite (ST wt.%: PCS3–PCS20) and evaluated these processes using isothermal calorimetry, 3D micro-computed tomography (μXCT), time-dependent rheometry, X-ray diffraction (XRD), and Fourier-transform infrared spectroscopy (FTIR), and the Krstulović-Dabić (K-D) model. The FTIR/XRD confirmed the coexistence …

Machine Learning In Peak Demand Forecasting: Foundations, Trends, And Insights, Shuang Dai, Fanlin Meng, Hongsheng Dai, Qian Wang, Xizhong Chen, Wenlei Bai, Peizhi Shi, Richard Allmendinger, Yuchen Zhang, Jian Liu

Electrical and Computer Engineering Faculty Research & Creative Works

Peak demand forecasting involves predicting the maximum electricity demand within a specific period, which plays a key role in maintaining the efficiency and stability of power systems. The rapid evolution of power systems, driven by advanced metering infrastructure, local energy applications such as electric vehicles, and the increasing adoption of intermittent renewable energy, has introduced greater randomness and reduced predictability in peak demand. Given the pressing need to address more diverse implementation requirements across different contexts, accurate and reliable peak demand forecasting has become increasingly important. To the best of our knowledge, this study is the first to provide a …

Heterogeneous Catalysis Of Large Biomolecules: Insights From Platinum Particle Size In Nad+Regeneration, Jianwei Li, Samuel Robertshaw, Shouying Huang, Shelley D. Minteer, Xiaodong Wang

Chemistry Faculty Research & Creative Works

Nicotinamide adenine dinucleotide (NAD⁺) cofactor regeneration is essential for enabling dehydrogenase-promoted biosynthesis for value-added chemicals. Heterogeneous catalytic cofactor regeneration, using supported metal catalysts, is an emerging approach and has shown great promise. However, mechanistic insight remains largely unexplored. In this work, a series of silica-supported platinum (Pt) catalysts have been prepared for NAD⁺ cofactor regeneration, to understand the roles of Pt particle size and structure. A turnover frequency (TOF) 'volcano plot' was obtained for Pt clusters in the range of 2.2–7.1 nm, with the maximum TOF (136 h⁻¹) observed at 5.6 nm. Selective Pt site …

Ultrasonic Extraction-Based Analysis Of Persistent Organic Pollutants In Blubber From False Killer Whales, Michael O. Eze, Eva Borras, Mitchell M. Mccartney, Don R. Bergfelt, Kristi L. West, Sarah E. Hooper, Cristina E. Davis

Chemistry Faculty Research & Creative Works

In view of the toxic effects of persistent organic pollutants (POPs), fast and effective assessment of their concentrations in marine mammals is important for understanding individual and population-level health impacts. This study developed an ultrasonic-based method that is less time-consuming, uses minimal solvent, and thus is more sustainable than the gold standard Soxhlet method for accurate analysis of organochlorine pesticides (OCs), polychlorinated biphenyls (PCBs), and benzene hexachlorides (BHC) in false killer whale blubber. This method was developed by comparing concentrations of POPs obtained using the traditional Soxhlet and novel ultrasonic extraction methods using blubber from false killer whales (n = …

Synergistic V2ctₓ Mxene–Pani Hybrid With Expanded Interlayers For Ultrastable And High-Rate Pseudocapacitive Energy Storage, Amideddin Nouralishahi, Maryam Sharifi Paroushi, Mansour Razavi, Amarachi Clare Nnachor, Harish Singh, Manashi Nath

Chemistry Faculty Research & Creative Works

Recently, MXene-conducting polymer hybrids have emerged as promising electrode materials for sustainable energy storage applications, owing to their impressive electrochemical properties. Herein, we report the synthesis of vanadium carbide MXene nanoparticles (V2CT_x-MXene) using innovative Spark Plasma Sintering (SPS) technology followed by exfoliation steps. The V2CT_x nanoparticles were incorporated with PANI (MXene-PANI) by electrochemical polymerization of aniline monomers in the presence of V2CT_x nanolayers, to be used as a highly efficient material for charge storage application. PANI nanofibers form a conductive and porous architecture, which intercalates the V2CT_x nanoflakes. The resulting structure increases the interlayer spacing …

Rogue Waves In Extended Gross-Pitaevskii Models With A Lee-Huang-Yang Correction, Sathyanarayanan Chandramouli, S. I. Mistakidis, G. C. Katsimiga, D. J. Ratliff, D. J. Frantzeskakis, P. G. Kevrekidis

Physics Faculty Research & Creative Works

We explore the existence and dynamical generation of rogue waves (RWs) within a one-dimensional quantum droplet-bearing environment. RWs are computed by deploying a space-time fixed point scheme to the relevant extended Gross-Pitaevskii equation (eGPE). Parametric regions where the ensuing RWs are different from their counterparts in the nonlinear Schrödinger equation are identified. To corroborate the controllable generation—relevant to ultracold atom experiments—of these rogue patterns, we exploit two different protocols. The first is based on interfering dam break flows emanating from Riemann initial conditions, and the second refers to the gradient catastrophe of a spatially localized waveform. A multitude of possible …

Pciafl: Personalized And Class Imbalance-Aware Federated Learning For Driver Behavior Classification, Osho Osho, Shubh Garg, Suchetana Chakraborty, Sajal K. Das

Computer Science Faculty Research & Creative Works

Automated understanding of driver behavior from vehicular kinematics is vital for safety-aware intelligent transportation systems. However, centralized cloud processing suffers from latency, scalability, and privacy issues. Federated Learning (FL) provides a decentralized alternative but faces two major challenges: (i) non-IID client data due to heterogeneous driving styles and sensors, and (ii) severe class imbalance, as risky behaviors are inherently rare. In this work, we propose a personalized FL framework that uses a shared CNN-LSTM backbone with client-adaptive classifiers and incorporates a cost-sensitive loss to address behavior skew. Evaluated on the UAH-DriveSet dataset, our method achieves 92.60% accuracy and 91.68% macro-F1, …

Datamut: Deterministic Algorithms For Time-Delay Attack Detection In Multi-Hop Uav Networks, Keiwan Soltani, Federico Corò, Punyasha Chatterjee, Sajal K. Das

Computer Science Faculty Research & Creative Works

Unmanned Aerial Vehicles (UAVs), also known as drones, have gained popularity in various fields such as agriculture, emergency response, and search and rescue operations. UAV networks are susceptible to potential security threats, such as wormhole attacks, jamming, spoofing, and false data injection. Time-Delay Attack (TDA) is a unique attack in which malicious UAVs intentionally delay packet forwarding, posing significant threats, especially in time-sensitive applications. It is challenging to distinguish malicious delay from benign network delay due to the dynamic nature of UAV networks, intermittent wireless connectivity, or the Store-Carry-Forward (SCF) mechanism during multi-hop communication. Some existing works propose machine learning-based …

Rescue: Routing Under Evolving Stochastic Congestion And Uncertain Spread In Wildfire Emergencies, Sowjanya Tammali, Arindam Khanda, Anurag Satpathy, S. M. Shovan, Sajal K. Das

Computer Science Faculty Research & Creative Works

Wildfires cause unpredictable spread and panic-driven congestion, posing severe challenges to evacuation planning. We present RESCUE (Routing under Evolving Stochastic Congestion and Uncertain Spread in Wildfire Emergencies), a dynamic, risk-aware framework that models the road network as a time-varying weighted graph. RESCUE operates in two stages: (i) a preprocessing phase integrating fire forecasts, traffic density, and distance to assign edge weights, and (ii) a real-time routing phase that adaptively updates paths using a multi-granular strategy distinguishing macro-level disruptions (e.g., rapid spread) from micro-level changes (e.g., local congestion). Two stochastic edge-cost functions are introduced: the Edge-Fire Risk Function (EFRF), estimating road …

Enhanced Superconductivity And Vortex Dynamics In Quasi-1d Tas2 Nanowires, Mathew Pollard, Visakha Ho, Clarissa Wisner, Eric W. Bohannan, Yew San Hor

Chemistry Faculty Research & Creative Works

We report the synthesis of high-quality 2H-TaS₂ nanowires via a controlled two-step conversion process from TaS₃ precursors, achieving robust superconductivity with a transition temperature T c ≈ 3.6K, which is significantly higher than bulk 2H-TaS₂ (T c ≈ 0.8K). Structural and compositional analyses confirm phase purity and preserved one-dimensional morphology, while magneto transport measurements reveal an enhanced upper critical field μ 0 H c2 (2K)≈5 T, far exceeding the bulk value (_μ0H_c2 (0)≈1.17T), attributed to dimensional confinement and suppression of charge-density wave order. Magnetic characterization demonstrates complex vortex dynamics, including flux jumps and a …

Explaining The Unseen: Multimodal Vision-Language Reasoning For Situational Awareness In Underground Mining Disasters, Mizanur Rahman Jewel, Mohamed Elmahallawy, Sanjay Kumar Madria, Samuel Frimpong

Computer Science Faculty Research & Creative Works

Underground mining disasters produce pervasive darkness, dust, and collapses that obscure vision and make situational awareness difficult for humans and conventional systems. To address this, we propose MDSE, Multimodal Disaster Situation Explainer, a novel vision-language framework that automatically generates detailed textual explanations of post-disaster underground scenes. MDSE has three-fold innovations: (i) Context-Aware Cross-Attention for robust alignment of visual and textual features even under severe degradation; (ii) Segmentation-aware dual pathway visual encoding that fuses global and region-specific embeddings; and (iii) Resource-Efficient Transformer-Based Language Model for expressive caption generation with minimal compute cost. To support this task, we present the Underground Mine …

Sadqn-Based Residual Energy-Aware Beamforming For Lora-Enabled Rf Energy Harvesting For Disaster-Tolerant Underground Mining Networks, Hilary Kelechi Anabi, Samuel Frimpong, Sanjay Madria

Mining Engineering Faculty Research & Creative Works

The end-to-end efficiency of radio-frequency (RF)-powered wireless communication networks (WPCNs) in post-disaster underground mine environments can be enhanced through adaptive beamforming. The primary challenges in such scenarios include (i) identifying the most energy-constrained nodes, i.e., nodes with the lowest residual energy to prevent the loss of tracking and localization functionality; (ii) avoiding reliance on the computationally intensive channel state information (CSI) acquisition process; and (iii) ensuring long-range RF wireless power transfer (LoRa-RFWPT). To address these issues, this paper introduces an adaptive and safety-aware deep reinforcement learning (DRL) framework for energy beamforming in LoRa-enabled underground disaster networks. Specifically, we develop a …

Meshless Collocation Methods For Time-Dependent Nonlocal Problems Based On Radial Basis Functions, Qiao Zhuang, Yanzhi Zhang, Zhongqiang Zhang

Mathematics and Statistics Faculty Research & Creative Works

We present radial basis function (RBF) collocation methods for time-dependent space fractional problems on general bounded domains. Building on a recently developed approach for accurately computing the integral fractional Laplacian of any RBF, we design collocation schemes for fractional heat and Stokes equations using extended-domain techniques. In particular, we propose a numerical Leray projection method for fractional Stokes problems, where both the discrete projection operator and the collocation scheme are formulated on extended domains to handle complex domains. Numerical results demonstrate the effectiveness of the proposed methods in solving time-dependent nonlocal problems on complex domains.

Safety Aware Continual Reinforcement Learning-Based Output Tracking Control Of Nonlinear Continuous-Time Systems, Irfan Ganie, Sarangapani Jagannathan

Electrical and Computer Engineering Faculty Research & Creative Works

An output feedback (OF)-based control scheme utilizing both a scalable multilayer neural network (MNN) observer and actor–critic MNN via integral reinforcement learning (IRL)/adaptive dynamics programming (ADP) approach for a class of nonlinear systems with output constraints is introduced. The proposed observer, critic, and actor MNN weight updates are derived using a singular value decomposition (SVD) of MNN activation function gradient along with output error, Bellman and control input errors, respectively. Next, the approach incorporates continual learning (CL), utilizing a penalty function in the weight update laws for both actor–critic MNNs to consolidate knowledge from previous tasks and enhance learning in …

Effect Of Ionic Strength On Gelation Time And Strength Of Amps-Based Polymer Gels, Maryam Sharifi Paroushi, Xuyang Tian, Baojun Bai, Thomas P. Schuman, Yin Zhang, Mingzhen Wei

Geosciences and Geological and Petroleum Engineering Faculty Research & Creative Works

Polymer gel treatment has been widely applied for improving sweep efficiency and controlling excessive water and gas production. Their performance depends on gelation time and final gel strength. In most studies, brine salinity is used to describe the effect of formation water on gel behavior. However, changing salinity also changes ionic strength and ion composition at the same time. Because of this coupling, it is difficult to identify the mechanisms controlling gelation, which has led to inconsistent trends in the literature. Increasing salinity has been reported to either slow or accelerate gelation and to weaken or strengthen gels depending on …

Fourier Pseudospectral Methods For The Variable-Order Space Fractional Wave Equations, Yanzhi Zhang, Xiaofei Zhao, Shiping Zhou

Mathematics and Statistics Faculty Research & Creative Works

In this paper, we propose Fourier pseudospectral methods to solve the variable-order space fractional wave equation and develop an accelerated matrix-free approach for its effective implementation. In constant-order cases, fast algorithms can be designed via the fast Fourier transforms (FFTs), and the computational cost at each time step is O(Nlog⁡N) with N the total number of spatial points. In variable-order cases, however, the spatial dependence in the power s(x) leads to the failure of inverse FFTs. While the direct matrix-vector multiplication approach becomes impractical due to excessive memory requirements. Hence, we propose an accelerated matrix-free approach for effective implementation in …

Empirical-Based Model Of Spatio-Temporal Errors, Godwin Naaba Ndaa

Masters Theses

This research develops an empirical model to characterize spatial-temporal InSAR errors and improve the accuracy of deformation time-series analysis. Using standardized Sentinel-1 HyP3 products and MintPy, the study quantifies how correlated noise affects velocity precision and validates the results against continuous GNSS velocities.

Residual velocities are near-Gaussian, with σ ~0.92–2.04 cm/yr and ~1 cm/yr on average. Variograms show power-law spatial structure with a non-zero nugget implicating troposphere and decorrelation while errors drop exponentially with more acquisitions; spatial uncertainty is strongly affected by unwrapping errors, coherence, and tropospheric noise, not simply troposphere.

A comparative assessment of on-demand cloud processing with other …

Analyzing Sleep Architecture And Brain State Transitions Via Hidden Markov Models On Fmri Data, Brileigh Jay Cates

Masters Theses

Sleep is associated with systematic changes in brain activity and functional connectivity observable in functional magnetic resonance imagining (fMRI) signals. Because subjects often fall asleep during resting-state experiments, the absence of vigilance monitoring can confound the interpretation of resting-state dynamics. Although electroencephalography (EEG) is the gold standard for sleep staging, simultaneous EEG-fMRI acquisition is not always feasible.

This study investigates whether sleep stages can be inferred directly from fMRI using a probabilistic latent-state framework. Hidden Markov Models (HMMs) are applied to blood-oxygen-level-dependent (BOLD) time series to identify latent brain states and their temporal transitions. Inferred states are aligned with EEG-derived …

Fairrfl: Fair And Robust Federated Learning In The Presence Of Selfish Clients, Andrea Augello, Ashish Gupta, Giuseppe Lo Re, Sajal K. Das

Computer Science Faculty Research & Creative Works

Federated Learning (FL) is a paradigm that enables collaborative machine learning without disclosing the local data of the participants. However, in real-world FL deployment scenarios, some unscrupolous clients may alter the training process to skew the global model towards their local optimum, unfairly prioritizing their data distribution. Their influence can degrade overall model performance for normal clients and reduce fairness in the system. We call this novel category of misbehaving clients 'selfish'. This work proposes a Fair and Robust strategy for aggregation in the Federated Learning (FL) server to mitigate the effect of Selfish clients (FairRFL). FairRFL incorporates a novel …

Partially Penalized Anisotropic Trilinear Ife-Pic Methods For Dc Plasma Transport Problems, Jiahui Li, Guangqing Xia, Yajie Han, Ziping Wang, Chang Lu, Xiaoming He

Mathematics and Statistics Faculty Research & Creative Works

Implicit and hybrid particle-in-cell methods are widely used for efficient simulation of DC discharge plasma transport. However, their computations require solving anisotropic elliptic equations and face challenges related to mesh geometry, non-axisymmetry, and complex interfaces. Moreover, the accuracy of particle trajectories is critical for plasma etching and erosion studies, where errors near interfaces can significantly impact simulation results. To address these challenges, this paper proposes a three-dimensional anisotropic trilinear partially penalized immersed finite element (ATPPIFE) method, which captures interfaces on Cartesian meshes and effectively reduces discontinuities at interface element faces, ensuring that particle trajectories better align with real-world behavior. Building …

Serum Biomarker Trajectory Clusters Predict Functional Outcome And Quality Of Life For Traumatic Brain Injury, Thanh Son Do, Chantal Carnes, Zhihui Yang, Firas Kobeissy, Hamad Yadikar, Gayla R. Olbricht, Olli Tenovuo, Jussi P. Posti, Ewout W. Steyerberg, Lindsay Wilson, Nicole Von Steinbüchel, Endre Czeiter, Andras Buki, David K. Menon

Mathematics and Statistics Faculty Research & Creative Works

Serum brain-enriched biomarkers are increasingly employed in the clinical evaluation of traumatic brain injury (TBI) to assist with triage, neuroimaging decisions, and prognostication. However, the potential of temporal biomarker trajectories to inform disease monitoring and long-term outcomes remains underexplored. We aim to identify distinct biomarker trajectory (TRAJ) profiles in traumatic brain injury patients and to examine their associations with long-term clinical outcomes. The study included 373, CT-positive Intensive Care Unit (ICU) traumatic brain injury patients (256 with initial Glasgow Coma Scale 3–12) from the Collaborative European Neurotrauma Effectiveness Research in TBI (CENTER-TBI) core study who had at least two serum …

Digital Twin Freshness Maximization In Edge Computing, Jing Li, Jianping Wang, Weifa Liang, Quan Chen, Sajal K. Das, Xiaohua Jia

Computer Science Faculty Research & Creative Works

Mobile Edge Computing (MEC) shifts powerful computing resource provisioning from remote powerful data centers to the edge of core networks. Meanwhile, Digital Twin (DT) has surfaced as a promising technology to provide comprehensive and dynamic descriptions of physical objects in cyberspace with bidirectional and real-time interactions. Moreover, Internet of Things (IoT) devices have contributed abundant, heterogeneous and continuous data from interconnected devices to the explosion of DTs. With technologies evolution, there is an increasing necessity to address the freshness of both DT states and DT data, through timely synchronizations between DTs and their objects in a highly dynamic IoT environment. …

On-Device Artificial Intelligence Solutions With Applications To Smart Environments, Fabrizio De Vita, Dario Bruneo, Sajal K. Das

Computer Science Faculty Research & Creative Works

Recent advances in Artificial Intelligence (AI) and the increasing availability of computational power have accelerated the diffusion of Intelligent Cyber-Physical Systems (ICPSs), enabling smart applications with reasoning capabilities. However, the limited resources of embedded and Edge devices significantly constrain the complexity of deep learning models that can be effectively deployed. Traditional approaches rely on cloud-based training and edge-only inference, a paradigm that becomes inadequate when low latency, privacy, security, and high customization are required. In this context, On-device AI is emerging as a new paradigm in which both training and inference are performed directly on the device, avoiding data transfer …

Pahina: Precision-Aware Hierarchical In-Network Aggregation For Edge Distributed Training, Yingpu Nian, Bo Yi, Qiang He, Xingwei Wang, Geyong Min, Keqin Li, Sajal K. Das

Computer Science Faculty Research & Creative Works

The rise of edge intelligence is driving distributed machine learning toward a new paradigm of edge-collaborative computing. To overcome the severe communication bottleneck in this paradigm, In-Network Aggregation is a critical enabling technology. However, its effectiveness is fundamentally undermined by the profound resource heterogeneity of edge networks. Specifically, edge devices, adapting to hardware constraints, operate at varying numerical precisions, leading to significant data inflation as gradients are aggregated. Compounding this, unevenly distributed network resources and traditional, precision-oblivious routing strategies often misallocate critical, high-precision gradients to low-quality paths. This mismatch creates severe network congestion, crippling the efficiency of distributed training. To …