Search NASA⌕ Search

SEARCH · Search NASA

Results for “sparse representation”

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 217 records · Page 12

Low-rank factorization of electron integral tensors and its application in electronic structure theory

In this letter, we introduce the reverse Cuthill-McKee (RCM) algorithm, which is often used for the bandwidth reduction of sparse tensors, to transform the two-electron integral tensors to their block diagonal forms. By further applying the pivoted Cholesky decomposition (CD) on each of the diagonal blocks, we are able to represent the high-dimensional two-electron integral tensors in terms of permutation matrices and low-rank Cholesky vectors. This representation facilitates the low-rank factorization of the high-dimensional tensor contractions that are usually encountered in post-Hartree-Fock calculations. In this letter, we discuss the second-order Møller-Plesset (MP2) method and linear coupled- cluster model with doubles (L-CCD) as two simple examples to demonstrate the efficiency of the RCM-CD technique in representing two-electron integrals in a compact form.

electronic structure theory↗

Cloud condensation nuclei concentrations derived from the CAMS reanalysis

Determining number concentrations of cloud condensation nuclei (CCN) is one of the first steps in the chain in analysis of cloud droplet formation, the direct microphysical link between aerosols and cloud droplets, and a process key for aerosol–cloud interactions (ACI). However, due to sparse coverage of in situ measurements and difficulties associated with retrievals from satellites, a global exploration of their magnitude, source as well as temporal and spatial distribution cannot be easily obtained. Thus, a better representation of CCN numbers is one of the goals for quantifying ACI processes and achieving uncertainty-reduced estimates of their associated radiative forcing. Here, we introduce a new CCN dataset which is derived based on aerosol mass mixing ratios from the latest Copernicus Atmosphere Monitoring Service reanalysis (CAMSRA) in a diagnostic model that uses CAMSRA aerosol properties and a simplified kappa-Köhler framework suitable for global models. The emitted aerosols in CAMSRA are not only based on input from emission inventories using aerosol observations, they also have a strong tie to satellite-retrieved aerosol optical depth (AOD) as this is assimilated as a constraining factor in the reanalysis. Furthermore, the reanalysis interpolates for cases of poor or missing retrievals and thus allows for a full spatiotemporal quantification of CCN numbers. The derived CCN dataset captures the general trend and spatial and temporal distribution of total CCN number concentrations and CCN from different aerosol species. A brief evaluation with ground-based in situ measurements demonstrates the improvement of the modelled CCN over the sole use of AOD as a proxy for CCN as the overall correlation coefficient improved from 0.37 to 0.71. However, we find the modelled CCN from CAMSRA to be generally high biased and find a particular erroneous overestimation at one heavily polluted site which emphasises the need for further validation. The CCN dataset (https://doi.org/10.26050/WDCC/QUAERERE_CCNCAMS_v1, Block, 2023), which is now freely available to users, features 3-D CCN number concentrations of global coverage for various supersaturations and aerosol species covering the years 2003–2021 with daily frequency. This dataset is one of its kind as it offers lots of opportunities to be used for evaluation in models and in ACI studies.

54 ENVIRONMENTAL SCIENCES↗

TH-E-18A-01: Developments in Monte Carlo Methods for Medical Imaging

Monte Carlo simulation methods are widely used in medical physics research and are starting to be implemented in clinical applications such as radiation therapy planning systems. Monte Carlo simulations offer the capability to accurately estimate quantities of interest that are challenging to measure experimentally while taking into account the realistic anatomy of an individual patient. Traditionally, practical application of Monte Carlo simulation codes in diagnostic imaging was limited by the need for large computational resources or long execution times. However, recent advancements in high-performance computing hardware, combined with a new generation of Monte Carlo simulation algorithms and novel postprocessing methods, are allowing for the computation of relevant imaging parameters of interest such as patient organ doses and scatter-to-primaryratios in radiographic projections in just a few seconds using affordable computational resources. Programmable Graphics Processing Units (GPUs), for example, provide a convenient, affordable platform for parallelized Monte Carlo executions that yield simulation times on the order of 10{sup 7} xray/ s. Even with GPU acceleration, however, Monte Carlo simulation times can be prohibitive for routine clinical practice. To reduce simulation times further, variance reduction techniques can be used to alter the probabilistic models underlying the x-ray tracking process, resulting in lower variance in the results without biasing the estimates. Other complementary strategies for further reductions in computation time are denoising of the Monte Carlo estimates and estimating (scoring) the quantity of interest at a sparse set of sampling locations (e.g. at a small number of detector pixels in a scatter simulation) followed by interpolation. Beyond reduction of the computational resources required for performing Monte Carlo simulations in medical imaging, the use of accurate representations of patient anatomy is crucial to the virtual generation of medical images and accurate estimation of radiation dose and other imaging parameters. For this, detailed computational phantoms of the patient anatomy must be utilized and implemented within the radiation transport code. Computational phantoms presently come in one of three format types, and in one of four morphometric categories. Format types include stylized (mathematical equation-based), voxel (segmented CT/MR images), and hybrid (NURBS and polygon mesh surfaces). Morphometric categories include reference (small library of phantoms by age at 50th height/weight percentile), patient-dependent (larger library of phantoms at various combinations of height/weight percentiles), patient-sculpted (phantoms altered to match the patient's unique outer body contour), and finally, patient-specific (an exact representation of the patient with respect to both body contour and internal anatomy). The existence and availability of these phantoms represents a very important advance for the simulation of realistic medical imaging applications using Monte Carlo methods. New Monte Carlo simulation codes need to be thoroughly validated before they can be used to perform novel research. Ideally, the validation process would involve comparison of results with those of an experimental measurement, but accurate replication of experimental conditions can be very challenging. It is very common to validate new Monte Carlo simulations by replicating previously published simulation results of similar experiments. This process, however, is commonly problematic due to the lack of sufficient information in the published reports of previous work so as to be able to replicate the simulation in detail. To aid in this process, the AAPM Task Group 195 prepared a report in which six different imaging research experiments commonly performed using Monte Carlo simulations are described and their results provided. The simulation conditions of all six cases are provided in full detail, with all necessary data on material composition, source, geometry, scoring and other parameters provided. The results of these simulations when performed with the four most common publicly available Monte Carlo packages are also provided in tabular form. The Task Group 195 Report will be useful for researchers needing to validate their Monte Carlo work, and for trainees needing to learn Monte Carlo simulation methods. In this symposium we will review the recent advancements in highperformance computing hardware enabling the reduction in computational resources needed for Monte Carlo simulations in medical imaging. We will review variance reduction techniques commonly applied in Monte Carlo simulations of medical imaging systems and present implementation strategies for efficient combination of these techniques with GPU acceleration. Trade-offs involved in Monte Carlo acceleration by means of denoising and “sparse sampling” will be discussed. A method for rapid scatter correction in cone-beam CT (<5 min/scan) will be presented as an illustration of the simulation speeds achievable with optimized Monte Carlo simulations. We will also discuss the development, availability, and capability of the various combinations of computational phantoms for Monte Carlo simulation of medical imaging systems. Finally, we will review some examples of experimental validation of Monte Carlo simulations and will present the AAPM Task Group 195 Report. Learning Objectives: Describe the advances in hardware available for performing Monte Carlo simulations in high performance computing environments. Explain variance reduction, denoising and sparse sampling techniques available for reduction of computational time needed for Monte Carlo simulations of medical imaging. List and compare the computational anthropomorphic phantoms currently available for more accurate assessment of medical imaging parameters in Monte Carlo simulations. Describe experimental methods used for validation of Monte Carlo simulations in medical imaging. Describe the AAPM Task Group 195 Report and its use for validation and teaching of Monte Carlo simulations in medical imaging.

60 APPLIED LIFE SCIENCES↗

Combining Tensor Decompositions and Graph Analytics to Provide Cyber Situational Awareness at HPC Scale

This paper describes MADHAT (Multidimensional Anomaly Detection fusing HPC, Analytics, and Tensors), an integrated workflow that demonstrates the applicability of HPC resources to the problem of maintaining cyber situational awareness. MADHAT combines two high-performance packages: ENSIGN for large-scale sparse tensor decompositions and HAGGLE for graph analytics. Tensor decompositions isolate coherent patterns of network behavior in ways that common clustering methods based on distance metrics cannot. Parallelized graph analysis then uses directed queries on a representation that combines the elements of identified patterns with other available information (such as additional log fields, domain knowledge, network topology, whitelists and blacklists, prior feedback, and published alerts) to confirm or reject a threat hypothesis, collect context, and raise alerts. MADHAT was developed using the collaborative HPC Architecture for Cyber Situational Awareness (HACSAW) research environment and evaluated on structured network sensor logs collected from Defense Research and Engineering Network (DREN) sites using HPC resources at the U.S. Army Engineer Research and Development Center DoD Supercomputing Resource Center (ERDC DSRC). To date, MADHAT has analyzed logs with over 650 million entries.

Cyber security, Tensor decomposition, graph analyt↗

A Novel Centrality Measure for Network-wide Cyber Vulnerability Assessment

In this work we propose a novel formulation that models the attack and compromise on a cyber network as a combination of two parts - direct compromise of a host and the compromise occurring through the spread of the attack on the network from a compromised host. The model parameters for the nodes are a concise representation of the host profiles that can include the risky behaviors of the associated human users while the model parameters for the edges are based on the existence of vulnerabilities between each pair of connected hosts. The edge models relate to the summary representations of the corresponding attack-graphs. This results in a formulation based on Random Walk with Restart (RWR) and the resulting centrality metric can be solved for in an efficient manner through the use of sparse linear solvers. Thus the formulation goes beyond mere topological considerations in centrality computations by summarizing the host profiles and the attack graphs into the model parameters. The computational efficiency of the method also allows us to also quantify the uncertainty in the centrality measure through Monte Carlo analysis.

Centrality↗

Self-consistent calculation of nuclear photoabsorption cross sections: Finite amplitude method with Skyrme functionals in the three-dimensional real space

The finite amplitude method (FAM), which we have recently proposed [T. Nakatsukasa, T. Inakura, and K. Yabana, Phys. Rev. C 76, 024318 (2007)], significantly simplifies the fully self-consistent calculation of the random-phase approximation (RPA). This article presents a computational scheme of FAM suitable for systematic investigation and shows its performance for realistic Skyrme energy functionals. We adopt the mixed representation in which the forward and backward RPA amplitudes are represented by index of hole orbitals and of the spatial grid points for the three-dimensional real space. We solve a linear algebraic problem with a sparse non-Hermitian matrix, using an iterative method. We show results of the dipole response for selected spherical and deformed nuclei. The calculated peak energies of the giant dipole resonance well agree with experiments for heavy nuclei. However, they are systematically underestimated for light nuclei. We also discuss the width of the giant dipole resonance in the fully self-consistent RPA calculation.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗

Human Sensorimotor Cortex Control of Directly Measured Vocal Tract Movements during Vowel Production

During speech production, we make vocal tract movements with remarkable precision and speed. Our understanding of how the human brain achieves such proficient control is limited, in part due to the challenge of simultaneously acquiring high-resolution neural recordings and detailed vocal tract measurements. To overcome this challenge, we combined ultrasound and video monitoring of the supralaryngeal articulators (lips, jaw, and tongue) with electrocorticographic recordings from the cortical surface of 4 subjects (3 female, 1 male) to investigate how neural activity in the ventral sensory-motor cortex (vSMC) relates to measured articulator movement kinematics (position, speed, velocity, acceleration) during the production of English vowels. We found that high-gamma activity at many individual vSMC electrodes strongly encoded the kinematics of one or more articulators, but less so for vowel formants and vowel identity. Neural population decoding methods further revealed the structure of kinematic features that distinguish vowels. Encoding of articulator kinematics was sparsely distributed across time and primarily occurred during the time of vowel onset and offset. In contrast, encoding was low during the steady-state portion of the vowel, despite sustained neural activity at some electrodes. Significant representations were found for all kinematic parameters, but speed was the most robust. These findings enabled by direct vocal tract monitoring demonstrate novel insights into the representation of articulatory kinematic parameters encoded in the vSMC during speech production.

47 OTHER INSTRUMENTATION↗

Latent Twins

Over the past decade, scientific machine learning has transformed the development of mathematical and computational frameworks for analyzing, modeling, and predicting complex systems. From inverse problems to numerical partial differential equations (PDEs), dynamical systems, and model reduction, these advances have pushed the boundaries of what can be simulated. Yet they have often progressed in parallel, with representation learning and algorithmic solution methods evolving largely as separate pipelines. With Latent Twins, we propose a unifying mathematical framework that creates a hidden surrogate in latent space for the underlying equations. Whereas digital twins mirror physical systems in the digital world, Latent Twins mirror mathematical systems in a learned latent space governed by operators. Through this lens, classical modeling, inversion, model reduction, and operator approximation all emerge as special cases of a single principle. We establish the fundamental approximation properties of Latent Twins for both ordinary differential equations (ODEs) and PDEs and demonstrate the framework across three representative settings: (i) canonical ODEs, capturing diverse dynamical regimes; (ii) a PDE benchmark using the shallow-water equations, contrasting Latent Twin simulations with deep operator network and forecasts with a four-dimensional variational method baseline; and (iii) a challenging real-data geopotential reanalysis dataset, reconstructing and forecasting from sparse, noisy observations. Latent Twins provide a compact, interpretable surrogate for solution operators that evaluate across arbitrary time gaps in a single-shot, while remaining compatible with scientific pipelines such as assimilation, control, and uncertainty quantification. Looking forward, this framework offers scalable, theory-grounded surrogates that bridge data-driven representation learning and classical scientific modeling across disciplines.

Latent Twins↗

Spectrotemporal CT data acquisition and reconstruction at low dose

Purpose: X-ray computed tomography (CT) is widely used, both clinically and preclinically, for fast, high-resolution anatomic imaging; however, compelling opportunities exist to expand its use in functional imaging applications. For instance, spectral information combined with nanoparticle contrast agents enables quantification of tissue perfusion levels, while temporal information details cardiac and respiratory dynamics. The authors propose and demonstrate a projection acquisition and reconstruction strategy for 5D CT (3D + dual energy + time) which recovers spectral and temporal information without substantially increasing radiation dose or sampling time relative to anatomic imaging protocols. Methods: The authors approach the 5D reconstruction problem within the framework of low-rank and sparse matrix decomposition. Unlike previous work on rank-sparsity constrained CT reconstruction, the authors establish an explicit rank-sparse signal model to describe the spectral and temporal dimensions. The spectral dimension is represented as a well-sampled time and energy averaged image plus regularly undersampled principal components describing the spectral contrast. The temporal dimension is represented as the same time and energy averaged reconstruction plus contiguous, spatially sparse, and irregularly sampled temporal contrast images. Using a nonlinear, image domain filtration approach, the authors refer to as rank-sparse kernel regression, the authors transfer image structure from the well-sampled time and energy averaged reconstruction to the spectral and temporal contrast images. This regularization strategy strictly constrains the reconstruction problem while approximately separating the temporal and spectral dimensions. Separability results in a highly compressed representation for the 5D data in which projections are shared between the temporal and spectral reconstruction subproblems, enabling substantial undersampling. The authors solved the 5D reconstruction problem using the split Bregman method and GPU-based implementations of backprojection, reprojection, and kernel regression. Using a preclinical mouse model, the authors apply the proposed algorithm to study myocardial injury following radiation treatment of breast cancer. Results: Quantitative 5D simulations are performed using the MOBY mouse phantom. Twenty data sets (ten cardiac phases, two energies) are reconstructed with 88 μm, isotropic voxels from 450 total projections acquired over a single 360° rotation. In vivo 5D myocardial injury data sets acquired in two mice injected with gold and iodine nanoparticles are also reconstructed with 20 data sets per mouse using the same acquisition parameters (dose: ∼60 mGy). For both the simulations and the in vivo data, the reconstruction quality is sufficient to perform material decomposition into gold and iodine maps to localize the extent of myocardial injury (gold accumulation) and to measure cardiac functional metrics (vascular iodine). Their 5D CT imaging protocol represents a 95% reduction in radiation dose per cardiac phase and energy and a 40-fold decrease in projection sampling time relative to their standard imaging protocol. Conclusions: Their 5D CT data acquisition and reconstruction protocol efficiently exploits the rank-sparse nature of spectral and temporal CT data to provide high-fidelity reconstruction results without increased radiation dose or sampling time.

60 APPLIED LIFE SCIENCES↗

Machine learning inversion of interatomic force constants from single-crystal inelastic neutron scattering

Atomic vibrations govern many macroscopic properties of materials, but experiments to comprehensively probe them remain challenging. Inelastic neutron scattering (INS) is a powerful technique to map phonon dispersions in crystals, especially when leveraging modern time-of-flight (ToF) spectrometers with large detectors. However, efficiently and robustly extracting interatomic force constants (FCs) parameterizing phonon dynamics from experimental spectra remains a bottleneck due to the complexity and high dimensionality of ToF INS datasets. Here, we present a machine learning approach for the direct inversion of FCs from single-crystal INS measurements. The framework leverages synthetic training data generated using universal machine-learned force fields and an efficient physics-based forward model. We benchmark two neural architectures–one emphasizing structured latent representation learning and the other direct, supervised spectral regression–across simulated datasets for two materials under idealized and noisy conditions. The latent-representation model is subsequently applied to experimental single-crystal INS data on germanium. The model is shown to reproduce FCs derived from both first-principles simulations and from iterative optimization, and furthermore achieves reliable inference even from sparse, single-orientation measurements representing short data acquisitions. Analysis of the learned latent space reveals semantically continuous and physically interpretable encodings that support strong cross-domain generalization. By bridging theoretical and experimental domains, we establish a path toward rapid inversion of experimental spectra and data-driven interpretation of temperature-dependent lattice dynamics.

42 ENGINEERING↗

Qubit Lattice Algorithms Based on the Schrödinger-Dirac Representation of Maxwell Equations and Their Extensions

It is well known that Maxwell equations can be expressed in a unitary Schrodinger-Dirac representation for homogeneous media. However, difficulties arise when considering inhomogeneous media. A Dyson map points to a unitary field qubit basis, but the standard qubit lattice algorithm of interleaved unitary collision-stream operators must be augmented by some sparse non-unitary potential operators that recover the derivatives on the refractive indices. Here, the effect of the steepness of these derivatives on two-dimensional scattering is examined with simulations showing quite complex wavefronts emitted due to transmissions/reflections within the dielectric objects. Maxwell equations are extended to handle dissipation using Kraus operators. Then, our theoretical algorithms are extended to these open quantum systems. A quantum circuit diagram is presented as well as estimates on the required number of quantum gates for implementation on a quantum computer.

2D electromagnetic scattering↗

On the Recoverability of Reactor Dynamics from Point Kinetics Data using SINDYc

Data driven models for reactor dynamics tend to face a few notable challenges. First, the broad range of timescales present across the various feedback mechanisms. Next, high standards for safety and importance of model performance in all operational domains. Finally, the correlations between state variables observed in most transients leads to difficulties when methods attempt to attribute certain dynamic phenomena to a particular cause. The current paper seeks to support efforts towards incorporating physics knowledge into one particular data-driven method for finding reactor dynamics called ``Sparse Identification of Nonlinear Dynamics with Control (SINDYc). The incorporation of physical knowledge into SINDYc will help address the challenges listed above by giving the model an initial understanding of the system being modeled. The demonstration of an exact representation of a set of point reactor dynamics equations in SINDYc is provided. Then, this allows for further discussion with mathematical justification as to why SINDYc, and other data-driven methods, may be difficult to apply to nuclear reactor dynamics due to correlations in the state variables. A particularly useful result of this work is the set of candidate functions required in SINDYc to exactly represent the reactor dynamics under a point approximation. In the future, SINDYc can be integrated with point models for reactor dynamics before being applied to the physical system to yield higher accuracy.

22 - GENERAL STUDIES OF NUCLEAR REACTORS↗

Artificial Intelligence Guided Studies of van der Waals Magnets

A materials informatics framework to explore a large number of candidate van der Waals (vdW) materials is developed. In particular, in this study a large space of monolayer transition metal halides is investigated by combining high-throughput density functional theory calculations and artificial intelligence (AI) to accelerate the discovery of stable materials and the prediction of their magnetic properties. Here, the formation energy is used as a proxy for chemical stability. Semi-supervised learning is harnessed to mitigate the challenges of sparsely labeled materials data in order to improve the performance of AI models. This approach creates avenues for the rapid discovery of chemically stable vdW magnets by leveraging the ability of AI to recognize patterns in data, to learn mathematical representations of materials from data and to predict materials properties. Using this approach, previously unexplored vdW magnetic materials with potential applications in data storage and spintronics are identified.

36 MATERIALS SCIENCE↗

Qubit Lattice Algorithms based on the Schrodinger-Dirac representation of Maxwell Equations and their Extensions

It is well known that Maxwell equations can be expressed in a unitary Schrodinger-Dirac representation for homogeneous media. However, difficulties arise when considering inhomoge- neous media. A Dyson map points to a unitary field qubit basis, but the standard qubit lattice algorithm of interleaved unitary collision-stream operators must be augmented by some sparse non-unitary potential operators that recover the derivatives on the refractive indices. The effect of the steepness of these derivatives on two dimensional scattering is examined with simulations showing quite complex wavefronts emitted due to transmissions/reflections within the dielectric objects. Maxwell equations are extended to handle dissipation using Kraus operators. Then, our theoretical algorithms are extended to these open quantum systems. A quantum circuit diagram is presented as well as estimates on the required number of quantum gates for implementation on a quantum computer.

Vahala, George↗

Fast and converged classical simulations of evidence for the utility of quantum computing before fault tolerance

A recent quantum simulation of observables of the kicked Ising model on 127 qubits implemented circuits that exceed the capabilities of exact classical simulation. We show that several approximate classical methods, based on sparse Pauli dynamics and tensor network algorithms, can simulate these observables orders of magnitude faster than the quantum experiment and can also be systematically converged beyond the experimental accuracy. Our most accurate technique combines a mixed Schrödinger and Heisenberg tensor network representation with the Bethe free entropy relation of belief propagation to compute expectation values with an effective wave function–operator sandwich bond dimension >16,000,000, achieving an absolute accuracy, without extrapolation, in the observables of <0.01, which is converged for many practical purposes. We thereby identify inaccuracies in the experimental extrapolations and suggest how future experiments can be implemented to increase the classical hardness.

Science & Technology - Other Topics↗

Indigo: A Domain-Specific Language for Fast, Portable Image Reconstruction

© 2018 IEEE. Linear operators used in iterative methods like conjugate gradient have typically been implemented either as ''matrix-driven'' subroutines backed by explicit sparse or dense matrices, or as ''matrix-free'' subroutines that implement specific linear operations directly (e.g. FFTs). The matrix-driven approach is generally more portable because it can target widely-Available BLAS libraries, but it can be inefficient in terms of time and space complexity. In contrast, the matrix-free approach is more performant because it leverages structure in operations, but it requires each operator be re-implemented on each new platform. To increase performance and portability, we propose a hybrid approach that represents linear operators as expression trees. Leaf nodes in the tree are either matrix-free or matrix-driven operators, and interior nodes represent mathematical compositions (sums, products, transposes) or structural compositions (stacks, block diagonals, etc.) of the leaf operators. This representation enables expert-guided reordering and fusion transformations that can improve performance or reduce memory pressure. We implement our approach in a domain-specific language called Indigo. We assess Indigo on image reconstruction problems arising in four application areas: magnetic resonance imaging, ptychography, magnetic particle imaging, and fluorescent microscopy. We give performance results from vendor BLAS libraries, and we introduce specializations to Sparse BLAS routines that achieve near-Roofline performance on multi-core, many-core, and GPU systems.

Driscoll, Michael↗

Knowledge-guided graph machine learning for spatially distributed prediction of daily discharge and nitrogen export dynamics

Spatially distributed prediction of streamflow and nitrogen export dynamics is essential for precision management of agricultural watersheds. While temporal deep learning models such as Long Short-Term Memory (LSTM) have shown strong performance at basin scales, their ability to generalize spatially is limited by insufficient representation of spatial dependencies and flow paths, particularly under data-scarce conditions. To address this gap, we propose HydroGraphNet, a knowledge-guided graph machine learning framework that integrates process-based knowledge and explicit spatial learning into temporal modeling. This framework incorporates directed graph topology to encode watershed connectivity and upstream inflows, with mass balance constraints to improve physical consistency. To enhance generalization in sparsely monitored regions, HydroGraphNet is pretrained on synthetic data generated by the SWAT+ (Soil and Water Assessment Tool Plus) model. We evaluated HydroGraphNet in the Upper Sangamon River Basin (44 HUC-12 subwatersheds, 2001–2020) against two LSTM baselines: a lumped basin-level model and a distributed variant. When benchmarked on SWAT+ simulations in pretraining, HydroGraphNet improved test NSEs by 8.9% (discharge) and 13.7% (NO₃–N load) in temporal extrapolation, and by 27.1% and 34.7% in spatial extrapolation, relative to the Lumped LSTM baseline. After fine-tuning with USGS monitoring data, the model achieved mean test NSE (KGE) scores of 0.768 (0.861) for discharge and 0.626 (0.664) for NO₃–N load, substantially outperforming baselines. Attribution analysis further highlighted the importance of upstream inflow representation and graph-based spatial learning in capturing cross-subwatershed dependencies. The model also reproduced seasonal hydrological and biogeochemical patterns consistent with known processes, demonstrating its robustness and process fidelity for spatially distributed prediction. Altogether, HydroGraphNet advances the integration of physical knowledge and spatially explicit learning in hydrological modeling, offering a generalizable framework for distributed modeling to support spatially targeted water quality management in data-scarce watersheds.

54 ENVIRONMENTAL SCIENCES↗