Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “graph neural network”

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 55 records · Page 3
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Addressing GPU memory limitations for Graph Neural Networks in High-Energy Physics applications

Introduction Reconstructing low-level particle tracks in neutrino physics can address some of the most fundamental questions about the universe. However, processing petabytes of raw data using deep learning techniques poses a challenging problem in the field of High Energy Physics (HEP). In the Exa.TrkX Project, an illustrative HEP application, preprocessed simulation data is fed into a state-of-art Graph Neural Network (GNN) model, accelerated by GPUs. However, limited GPU memory often leads to Out-of-Memory (OOM) exceptions during training, due to the large size of models and datasets. This problem is exacerbated when deploying models on High-Performance Computing (HPC) systems designed for large-scale applications. Methods We observe a high workload imbalance issue during GNN model training caused by the irregular sizes of input graph samples in HEP datasets, contributing to OOM exceptions. We aim to scale GNNs on HPC systems, by prioritizing workload balance in graph inputs while maintaining model accuracy. Our paper introduces diverse balancing strategies aimed at decreasing the maximum GPU memory footprint and avoiding the OOM exception, across various datasets. Results Our experiments showcase memory reduction of up to 32.14% compared to the baseline. We also demonstrate the proposed strategies can avoid OOM in application. Additionally, we create a distributed multi-GPU implementation using these samplers to demonstrate the scalability of these techniques on the HEP dataset. Discussion By assessing the performance of these strategies as data loading samplers across multiple datasets, we can gauge their effectiveness in both single-GPU and distributed environments. Our experiments, conducted on datasets of varying sizes and across multiple GPUs, broaden the applicability of our work to various GNN applications that handle input datasets with irregular graph sizes.

Lee, Claire Songhyun↗
DOE OSTI2024

MDLoader: A Hybrid Model-Driven Data Loader for Distributed Graph Neural Network Training

Scalable data management is essential for processing large scientific dataset on HPC platforms for distributed deep learning. In-memory distributed storage is preferred for its speed, enabling rapid, random, and frequent data access required by stochastic optimizers. Processes use one-sided or collective communication to fetch remote data, with optimal performance depending on (i) dataset characteristics, (ii) training scale, and (iii) interconnection network. Empirical analysis shows collective communication excels with larger mini-batch sizes and/or fewer processes, whereas one-sided communication outperforms at larger scales. We propose MDLoader, a hybrid in-memory data loader for distributed graph neural network training. MDLoader features a model-driven performance estimator that dynamically selects between one-sided and collective communication at the beginning of training using Tree of Parzen Estimators (TPE). Evaluations on NERSC Perlmutter and OLCF Summit show MDLoader outperforms single-backend loaders by up to 2.83 × and predicts the suitable communication method with 96.3% (Perlmutter) and 94.3% (Summit) success rate.

Bae, Jonghyun↗
DOE OSTI2024

Using graph neural networks to reconstruct charged pion showers in the CMS High Granularity Calorimeter

A novel method to reconstruct the energy of hadronic showersin the CMS High Granularity Calorimeter (HGCAL) is presented. TheHGCAL is a sampling calorimeter with very fine transverse andlongitudinal granularity. The active media are silicon sensors andscintillator tiles readout by SiPMs and the absorbers are acombination of lead and Cu/CuW in the electromagnetic section, andsteel in the hadronic section. The shower reconstruction method isbased on graph neural networks and it makes use of a dynamicreduction network architecture. It is shown that the algorithm isable to capture and mitigate the main effects that normally hinderthe reconstruction of hadronic showers using classicalreconstruction methods, by compensating for fluctuations in themultiplicity, energy, and spatial distributions of the shower'sconstituents. The performance of the algorithm is evaluated usingtest beam data collected in 2018 prototype of the CMS HGCALaccompanied by a section of the CALICE AHCAL prototype. Thecapability of the method to mitigate the impact of energy leakagefrom the calorimeter is also demonstrated.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
DOE OSTI2025

Spin-informed universal graph neural networks for simulating magnetic ordering

The screening and discovery of magnetic materials are hindered by the computational cost of first-principles density-functional theory (DFT) calculations required to find the ground state magnetic ordering. Although universal machine-learning interatomic potentials (uMLIPs), also known as atomistic foundation models, offer high-fidelity models of many atomistic systems with significant speedup, they currently lack the inputs required for predicting magnetic ordering. In this work, we present a data-efficient, spin-informed graph neural network framework that incorporates spin degrees of freedom as inputs and preserves physical symmetries, extending the functionality of uMLIPs to simulate magnetic orderings. This framework speeds up DFT calculations through better initial guesses for magnetic moments, determines the ground-state ordering of bulk materials and even generalizes to magnetic ordering in surfaces. Furthermore, we implement a closed-loop anomaly detection approach that effectively addresses the classic "chicken-and-egg" problem of creating a high-quality dataset while developing a uMLIP, unearthing anomalies in large benchmark datasets and boosting model accuracy.

Xu, Wenbin↗
DOE OSTI2024

Moiré Patterns in Pt Overlayers on Gold: A Graph Neural Network Interatomic Potential Study

Overlayer structures in bimetallic catalysts are relevant to a variety of catalytic reactions, particularly in electrocatalysis for fuel cell applications. Previous computational studies largely consider these overlayer structures to be those of a pseudomorphic overlayer, where there is a 1:1 atomic ratio between the overlayer and the support metal. Our previous work based on density functional theory (DFT) has shown that there exist nonstoichiometric overlayer structures that are more stable than the stoichiometric ones. Here, in this work, we developed a graph neural network interatomic potential (GNN-IP) to analyze structures formed in Pt overlayers on Au(111). The GNN-IP was used to explore properties of nonstoichiometric overlayers at length scales that are prohibitively expensive to pursue with planewave DFT calculations. In particular, we examined large Pt islands on top of a 48 × 48 Au(111) unit cell to explore the influence of the rotational angle (α) between the Pt overlayer and support Au(111) on both the stability of and the preferred atomic density in the overlayer. Island structures with smaller rotational angles between the Pt overlayer and Au(111) tend to be more stable. Further, smaller rotational angles tend to result in lower atomic density in the Pt overlayer.

chemical structure↗
DOE OSTI2025

Expediting field-effect transistor chemical sensor design with neuromorphic spiking graph neural networks

Improving the sensitive and selective detection of analytes in a variety of applications requires accelerating the rational design of field-effect transistor (FET) chemical sensors. Achieving high-performance detection relies on identifying optimal probe materials that can effectively interact with target analytes, a process traditionally driven by chemical intuition and time-consuming trial-and-error methods. To address the difficulties in probe screening for FET sensor development, this work presents a methodology that combines neuromorphic machine learning (ML) architectures, specifically a hybrid spiking graph neural network (SGNN), with an enriched dataset of physicochemical properties through semi-automated data extraction using large language models. Achieving a classification accuracy of 0.89 in predicting sensor sensitivity categories, the SGNN model outperformed traditional ML techniques by leveraging its ability to capture both global physicochemical properties and sparse topological features through a hybrid modeling framework. Next-generation sensor design was informed by the actionable insights into the connections between material properties and sensing performance offered by the SGNN framework. Through virtual screening for the detection of per- and polyfluoroalkyl substances (PFAS) as a use case, the effectiveness of the SGNN model was further validated. Density functional theory simulations confirmed graphene as a promising active material for PFAS detection as suggested by the SGNN framework. By bridging gaps in predictive modeling and data availability, this integrated approach provides a strong foundation for accelerating advancements in FET sensor design and innovation.

Ferreira, Rodrigo Pires [Univ. of Chicago, IL (Uni↗
DOE OSTI2025

Defect Diffusion Graph Neural Networks for Materials Discovery in High-Temperature Energy Applications

Here, the migration of crystallographic defects dictates material properties and performance for a plethora of technological applications. Density functional theory (DFT)-based nudged elastic band (NEB) calculations are a powerful computational technique for predicting defect migration activation energy barriers, yet they become prohibitively expensive for high-throughput screening of defect diffusivities. Without introducing hand-crafted (i.e., chemistry- or structure-specific) descriptors, we propose a generalized deep learning approach to train surrogate models for NEB energies of vacancy migration by hybridizing graph neural networks with transformer encoders and simply using pristine host structures as input. With sufficient training data, computationally efficient and simultaneous inference of vacancy defect thermodynamics and migration activation energies can be obtained to compute temperature-dependent vacancy diffusivities and to down-select candidates for more thorough DFT analysis or experiments. Thus, as we specifically demonstrate for potential water-splitting materials, candidates with desired defect thermodynamics, kinetics, and host stability properties can be more rapidly targeted from open-source databases of experimentally validated or hypothetical materials.

14 SOLAR ENERGY↗
DOE OSTI2025

HEPOM: Using Graph Neural Networks for the Accelerated Predictions of Hydrolysis Free Energies in Different pH Conditions

Hydrolysis is a fundamental family of chemical reactions where water facilitates the cleavage of bonds. The process is ubiquitous in biological and chemical systems, owing to water’s remarkable versatility as a solvent. However, accurately predicting the feasibility of hydrolysis through computational techniques is a difficult task, as subtle changes in reactant structure like heteroatom substitutions or neighboring functional groups can influence the reaction outcome. Furthermore, hydrolysis is sensitive to the pH of the aqueous medium, and the same reaction can have different reaction properties at different pH conditions. In this work, we have combined reaction templates and high-throughput ab initio calculations to construct a diverse data set of hydrolysis free energies. The developed framework automatically identifies reaction centers, generates hydrolysis products, and utilizes a trained graph neural network (GNN) model to predict ΔG values for all potential hydrolysis reactions in a given molecule. The long-term goal of the work is to develop a data-driven, computational tool for high-throughput screening of pH-specific hydrolytic stability and the rapid prediction of reaction products, which can then be applied in a wide array of applications including chemical recycling of polymers and ion-conducting membranes for clean energy generation and storage.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2026

pyFLANK, a graph neural network based null distribution inference model for F ST outlier detection

Detecting genomic regions under selection is essential for understanding how populations adapt to different environments, yet it remains challenging due to the confounding effects of demographic history and linkage disequilibrium (LD). Fixation index (F ST ) is a widely used statistic to identify genomic regions under adaptation. However, identifying genes under selection by defining F ST outliers often remains challenging, owing to confounding effects of underlying demographic history. Traditional methods assume independence among loci and rely on simple demographic models, while newer models perform much better but are computationally expensive and not easily scalable. Here, we present pyFLANK, an open-source and automated Python implementation which detects F ST outliers using a null distribution inferred from quasi-independent loci. Our tool integrates three approaches to identify loci obeying a null distribution: graph neural network (GNN) inference, linkage disequilibrium (LD)-based inference, and user-defined input. Because pyFLANK uses GNN-based inference of quasi-independent loci, it yields a more accurate null model with less need for user parameter input. In simulation experiments, pyFLANK achieved lower false positive rates than current methods while maintaining comparable detection power, indicating that its refined null model better distinguishes true adaptive loci from background variation. The GNN-based model, in particular, detected additional loci associated with phenotypic variance that were not identified by existing methods. Assessments of simulation and real data from different species demonstrate that pyFLANK achieves lower false positive rates compared with other commonly used F ST outlier detectors, while maintaining comparable detection power and excellent computational performance, providing a robust and user-friendly tool for identifying loci under divergent selection. It extends existing F ST outlier frameworks by incorporating explicit LD-aware strategies for null model calibration. The method is intended as a practical and scalable complement to existing genome scan approaches.

FST↗
DOE OSTI2024

A Graph Neural Network Surrogate Model for hls4ml

Recent advancements in use of machine learning (ML) techniques on field-programmable gate arrays (FPGAs) have allowed for the implementation of embedded neural networks with extremely low latency. This is invaluable for particle detectors at the Large Hadron Collider, where latency and used area are strictly bounded. The hls4ml framework is a procedure that converts trained ML model software to a synthesis result to can be used on an FPGA. However, running the pipeline is a time-consuming procedure, and there is a strong risk of failure. In particular, it may not be possible to successfully convert a model into a synthesis result, or the resource consumption of the model may exceed the resources of the target FPGA. To aid with this development, we introduce wa-hls4ml, a surrogate model using a graph neural network to emulate the structure of the source models. The goal is to estimate the chance of success and resource consumption of a given model when passed through the hls4ml pipeline, without needing to run the pipeline.

Plotnikov, Dennis↗
DOE OSTI2025

Enabling Real-Time Communication in Multi-Agent Systems: A Graph Neural Network Based Approach

Global connectivity enables effective coordination in Multi-Agent Systems (MAS). Solving these connection problems under hardware constraints is an NP-hard non-Euclidean Degree Constrained Minimum Spanning Tree (DCMST) problem. Prior MAS controllers coordinate team movement for task completion and collision avoidance; some considering Line-of-Sight (LOS) maintenance but prioritizing flexibility over guarantees. Evolutionary Algorithms (EA) have been shown to find good solutions for DCMST, but their performance degrades with larger populations required to support a large MAS. We present a method based on edge graph attention networks, trained offline to reduce online computation times. Empirical comparisons with greedy polynomial-time solvers and EA show that our method leverages latent graph information to consistently find constraint-satisfying solutions in less time.

connectivity maintenance↗
DOE OSTI2025

Thermodynamic Modeling of Complex Solid Solutions in the Lu-H-N System via Graph Neural Network Accelerated Monte Carlo Simulations

Metal hydrides are important across diverse applications, such as hydrogen storage, batteries, gas sensors, nuclear reactions, and high-temperature superconductivity. Previous computational studies of metal hydrides under extreme pressures, e.g., 𝑂⁡(10 2 ) ⁢GPa, usually treat them as stoichiometric compounds without considering interstitial lattice disorder. As pressures become more moderate in the 𝑂⁡(10 0 ) ⁢GPa and below range, hydrogen disorder at interstitial lattice sites becomes prominent, e.g., in the N-doped Lu hydride that was recently claimed superconducting near 1 GPa. Further adding compositional complexity from alloying and/or multielement interstitial occupation makes elucidating pressure- and temperature-dependent observables intractable by first-principles calculations alone. We therefore propose a lattice graph neural-network surrogate modeling approach to predict configuration- and pressure-dependent equation-of-state properties. Their efficiency permits Monte Carlo simulations to calculate Gibbs energies and pressure-dependent phase diagrams, thereby revealing insights into the synthesis conditions required for achieving desired phase equilibria. We demonstrate this concept for the compositionally complex cubic Lu(H,N,Va) 3 system where three constituents (hydrogen, nitrogen and vacancy) have disordered multielement interstitial occupancies and insights into pressure-dependent phase equilibria are critically needed, e.g., N-doping levels can significantly lower dehydrogenation temperatures and provide a new strategy to optimize hydrogen-storage alloys. This work can improve the thermodynamic understanding of the Lu-H-N system and help rational synthesis of N-doped Lu hydrides, but more generally demonstrates an efficient approach to model pressure-dependent thermodynamics of multicomponent solid solutions.

Monte Carlo methods↗
DOE OSTI2025

Contributions of vegetation heterogeneity within tower footprint to CO 2 flux estimations through graph neural network modeling

Net ecosystem exchange of CO 2 (Fc) measured directly by eddy covariance towers is based on various assumptions, including large, flat and homogenous land cover type. In reality, often a tower site is not large enough for flux measurements, and landscapes consist of patches of different land cover types within the flux footprint. In addition, some portions of fluxes are contributed by different cover types when a footprint exceeds the size of the target ecosystem. The contributions of non-dominant patches to Fc are often ignored. Here, in this study, we propose a novel integrated modeling framework that combines random forest (RF) and XGBoost with a residual correction module based on a deep graph convolutional network (DeeperGCN) to simulate Fc for seven flux measurement sites in southwest Michigan. High-resolution remote sensing vegetation indices, soil properties, meteorological variables, and footprint-weighted spatial features were used as model inputs at three spatial resolutions (10 m, 20 m, 30 m), and their importance in predicting Fc with DeeperGCN was assessed. We found that residual correction using DeeperGCN significantly improved prediction accuracy, with the R 2 increasing from 0.9098 to 0.9479 for RF and from 0.9235 to 0.9433 for XGBoost. At site level, the maximum improvement in R 2 reached 0.1617. Paired t-tests confirmed that these improvements were statistically significant (p < 0.05). Among all predictors, leaf area index and incoming shortwave radiation emerged as the dominant drivers of spatial residual variation, followed by precipitation, relative humidity, and selected vegetation indices. The 20 m resolution yielded the best balance between model performance and computational efficiency. In conclusion, our modeling framework effectively captures both spatial heterogeneity and nonlinear interactions, offering a robust solution for spatially explicit flux modeling in structurally diverse ecosystems beyond the study sites.

footprint model↗
DOE OSTI2025

Physics-Informed Graph Neural Networks for Collaborative Dynamic Reconfiguration and Voltage Regulation in Unbalanced Distribution Systems

Network reconfiguration has long been employed as a strategic approach to minimize power distribution system losses and effectively regulate voltage levels. Tap-changing voltage regulators are also critical for controlling bus voltages, especially in accommodating the increasing integration of distributed energy resources (DERs) with intermittent outputs. This paper introduces novel methodologies to address the challenges of dynamic reconfiguration and optimal tap setting in unbalanced three-phase distribution systems. We propose an approximated mixed-integer quadratically constrained program (MIQCP) to model dynamic reconfiguration, along with a pioneering formulation for voltage regulator (VR) tap-setting based on Special Ordered Set type 1 (SOS1). To mitigate computational complexity, we propose a physics-informed spatial-temporal graph convolutional network (STGCN) with an integrated link classifier. The proposed approach enables efficient solution generation by fixing specific variables in the MIQCP instance and solving the simplified sub-MIP using an MIP solver. Numerical studies demonstrate the superior prediction accuracy of our STGCN model compared to baseline neural network models, resulting in reduced DER curtailment and voltage deviation with shorter computation time.

dynamic reconfiguration↗