Search NASA⌕ Search

SEARCH · Search NASA

Results for “Complex Network Data”

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 1,189 records · Page 66

Scalable Graph Analytics and HPC Operational Enhancement: Parallel Computing and ML/DL Innovations

Parallel computing plays a pivotal role in the efficient processing of large-scale graphs. Complex network analysis stands as a capti- vating research frontier, holding promise across diverse scientific domains such as sociology, biology, online media, and recommenda- tion systems. In this era, Machine Learning (ML) and Deep Learning (DL) have emerged as indispensable tools, underpinning remarkable technological achievements. Within this dynamic landscape, my research revolves around advancing parallel algorithms tailored for large-scale graph operations. To achieve this, I harness the power of cutting-edge technologies including OpenMP, MPI, HIP, and CUDA, on the High-Performance Computing (HPC) platforms to unlock optimal performance. I also apply ML/DL techniques to HPC operational data, to streamline the monitoring and maintenance of supercomputers, alleviating the complexities associated with their upkeep and enhancing user support. My research echoes the syn- ergy between parallel computing, large-scale graph analysis, and ML/DL, improving computational efficiency and user experience.

Sattar, Naw Safrin↗

Understanding Machine Learning in Earth Science: A Natural Language Processing Approach

Machine learning (ML) is being increasingly utilized in Earth science research. Benefits of ML include efficiency, reduction of human error, and ability to extract hidden patterns within data. However, the mutual lack of each other’s domain knowledge by ML and Earth science stands as a barrier to timely and effective implementation. Earth science, in particular, faces challenges in generating sample data, compared to those of traditional ML problems such as face recognition or stock predictions, where data is abundant and not lacking in ground truth, which is necessary for labeling. Earth science data are more varying in formats, such as HDF5 and image resolutions, and are not standardized across instruments, even within a given Earth science discipline. Previous studies have been done to outline the specific challenges that Earth science faces with ML, while others have focused on using existing publications to mine information efficiently. Other resources such as Scikit-Learn have developed decision trees for choosing appropriate machine learning algorithms, but application within Earth science subjects becomes much more complex. For the current study, we propose a methodology and tool that aids in implementation of ML in Earth science using natural language processing (NLP). Our work comprises three main parts: (1) analyzing existing publications related to ML and Earth science, using natural language processing: (2) extracting from the publications information on ML models subjects in Earth Science: and (3) visualizing the extracted relationships as a network graph. The resulting network graph should aid the Earth science communities in applying optimal ML algorithms and guiding data preparation through visualization of similar studies. The network graph and analysis of document similarity will be the basis of our next step, which is to develop a decision tree for selecting optimal machine learning methodologies for specified Earth science applications.

Zheng, Laura↗

Making tomorrow's mistakes today: Evolutionary prototyping for risk reduction and shorter development time

In the early days of JPL's solar system exploration, each spacecraft mission required its own dedicated data system with all software applications written in the mainframe's native assembly language. Although these early telemetry processing systems were a triumph of engineering in their day, since that time the computer industry has advanced to the point where it is now advantageous to replace these systems with more modern technology. The Space Flight Operations Center (SFOC) Prototype group was established in 1985 as a workstation and software laboratory. The charter of the lab was to determine if it was possible to construct a multimission telemetry processing system using commercial, off-the-shelf computers that communicated via networks. The staff of the lab mirrored that of a typical skunk works operation -- a small, multi-disciplinary team with a great deal of autonomy that could get complex tasks done quickly. In an effort to determine which approaches would be useful, the prototype group experimented with all types of operating systems, inter-process communication mechanisms, network protocols, packet size parameters. Out of that pioneering work came the confidence that a multi-mission telemetry processing system could be built using high-level languages running in a heterogeneous, networked workstation environment. Experience revealed that the operating systems on all nodes should be similar (i.e., all VMS or all PC-DOS or all UNIX), and that a unique Data Transport Subsystem tool needed to be built to address the incompatibilities of network standards, byte ordering, and socket buffering. The advantages of building a telemetry processing system based on emerging industry standards were numerous: by employing these standards, we would no longer be locked into a single vendor. When new technology came to market which offered ten times the performance at one eighth the cost, it would be possible to attach the new machine to the network, re-compile the application code, and run. In addition, we would no longer be plagued with lack of manufacturer support when we encountered obscure bugs. And maybe, hopefully, the eternal elusive goal of software portability across different vendors' platforms would finally be available. Some highlights of our prototyping efforts are described.

Friedman, Gary↗

Technology Infusion of CodeSonar into the Space Network Ground Segment (RII07): Software Assurance Symposium Technical Summary

Presents a source code analysis tool (CodeSonar) for use in the Space Network Ground Segment. The Space Network requires 99.9% proficiency and 97.0% availability of systems. Software has historically accounted for an annual average of 28% of the Space Network loss of availability and proficiency. CSCI A and CSCI B account for 42% of the previous eight months of software data loss. The technology infusion of CodeSonar into the Space Network Ground segment is meant to aid in determining the impact of the technology on the project both in the expenditure of effort and the technical results of the technology. Running a CodeSonar analysis and performing a preliminary review of the results averaged 3.5 minutes per finding (approximately 20 hours total). An additional 40 hours is estimated to analyze the 37 findings deemed too complex for the initial review. Using CodeSonar's tools to suppress known non-problems, delta tool runs will not repeat findings that have been marked as non-problems, further reducing the time needed for review. The 'non-interesting' finding rate of 70% is a large number, but filtering, search, and detailed contextual features of CodeSonar reduce the time per finding. Integration of the tool into the build process may also provide further savings by preventing developers from having to configure and operate the tool separately. These preliminary results show the tool to be easy to use and incorporate into the engineering process. These findings also provide significant potential improvements in proficiency and availability on the part of the software. As time-to-fix data become available a better cost trade can be made on person hours saved versus tool cost. Selective factors may be necessary to determine where best to apply CodeSonar to balance cost and benefits.

Benson, Markland J.↗

Evaluation Report on "The Department's Unclassified Cyber Security Program"

Industry experts report that security challenges and threats are continually evolving as malicious activity has become more web-based and attackers are able to rapidly adapt their attack methods. In addition, the number of data breaches continues to rise. In an effort to mitigate and address threats and protect valuable information, the Department of Energy anticipated spending about $275 million in Fiscal Year (FY) 2009 to implement cyber security measures necessary to protect its information technology resources. These systems and data are designed to support the Department's mission and business lines of energy security, nuclear security, scientific discovery and innovation, and environmental responsibility. The Federal Information Security Management Act of 2002 (FISMA) provides direction to agencies on the management and oversight of information security risks, including design and implementation of controls to protect Federal information and systems. As required by FISMA, the Office of Inspector General conducts an annual independent evaluation to determine whether the Department's unclassified cyber security program adequately protects its information systems and data. This memorandum and the attached report present the results of our evaluation for FY 2009. The Department continued to make incremental improvements in its unclassified cyber security program. Our evaluation disclosed that most sites had taken action to address weaknesses previously identified in our FY 2008 evaluation report. They improved certification and accreditation of systems; strengthened configuration management of networks and systems; performed independent assessments; and, developed and/or refined certain policies and procedures. In addition, the Department instituted a centralized incident response organization designed to eliminate duplicative efforts throughout the Department. As we have noted in previous reports, the Department continued to maintain strong network perimeter defenses against malicious intruders and other externals threats. These are positive accomplishments. However, in our judgment, additional action is required to further enhance the Department's unclassified cyber security program and help reduce risks to its systems and data. For example, our current review identified opportunities for improvements in areas such as security planning and testing, systems inventory, access controls, and configuration management. In particular, we issued a number of findings at sites managed by the National Nuclear Security Administration (NNSA). We also identified weaknesses across various Department program elements. Issues that warrant further attention include: (1) Weaknesses such as outdated security plans and not completing annual security control self-assessments were identified at several sites; (2) The Department had not yet resolved systems inventory issues and had yet to deploy a complex-wide automated asset management tool to help track information technology resources and identify interfaces between systems or networks; (3) Although certain improvements had been made to enhance access controls, we noted deficiencies such as a lack of periodic account reviews and inadequate password management at a number of sites; and (4) Previously identified weaknesses in configuration management had been corrected, however, we found problems related to weak administrator account settings and failure to install software patches, as well as incomplete implementation of the Federal Desktop Core Configuration. These internal control weaknesses existed, at least in part, because certain cyber security roles and responsibilities were not clearly delineated. Program officials also had not effectively performed monitoring and review activities essential for evaluating the adequacy of cyber security performance. In some cases, officials had not ensured that weaknesses discovered during audits and other evaluations were recorded and tracked to resolution in the organizations' Plans of Action and Milestones. Our testing disclosed that about 39 percent of existing corrective action milestones had missed estimated remediation dates, with many exceeding planned completion dates by at least one year. As a consequence, the risk of compromise to the Department's information and systems remained higher than necessary. To assist the continuing efforts to improve, we made several recommendations designed to help the Department's managers to strengthen the unclassified cyber security program and, thereby, protect its computer resources from unauthorized modification, loss, or disclosure of information. Due to security considerations, information on specific vulnerabilities and locations has been omitted from this report. Management officials at the sites evaluated were provided with detailed information regarding identified vulnerabilities, and, in many instances, initiated corrective actions.

99 GENERAL AND MISCELLANEOUS↗

Characterizing Tradeoffs in Memory, Accuracy, and Speed for Chemistry Tabulation Techniques

Chemistry tabulation is a common approach in practical simulations of turbulent combustion at engineering scales. Linear interpolants have traditionally been used for accessing precomputed multidimensional tables but suffer from large memory requirements and discontinuous derivatives. Higher-degree interpolants address some of these restrictions but are similarly limited to relatively low-dimensional tabulation. Artificial neural networks (ANNs) can be used to overcome these limitations but cannot guarantee the same accuracy as interpolants and introduce challenges in reproducibility and reliable training. These challenges are enhanced as the physics complexity to be represented within the tabulation increases. Here, we assess the efficiency, accuracy, and memory requirements of Lagrange polynomials, tensor product B-splines, and ANNs as tabulation strategies. We analyze results in the context of nonadiabatic flamelet modeling where higher dimension counts are necessary. While ANNs do not require structuring of data, providing benefits for complex physics representation, interpolation approaches often rely on some structuring of the table. Interpolation using structured table inputs that are not directly related to the variables transported in a simulation can incur additional query costs. This is demonstrated in the present implementation of heat losses. We show that ANNs, despite being difficult to train and reproduce, can be advantageous for high-dimensional, unstructured datasets relevant to nonadiabatic flamelet models. Furthermore we demonstrate that Lagrange polynomials show significant speedup for similar accuracy compared to B-splines.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

Mind the gap: the discrepancy between simulation and reality drives interpretations of the Galactic Center Excess

The Galactic Center Excess (GCE) in GeV gamma rays has been debated for over a decade, with the possibility that it might be due to dark matter annihilation or undetected point sources such as millisecond pulsars (MSPs). This study investigates how the gamma-ray emission model (γEM) used in Galactic center analyses affects the interpretation of the GCE's nature. To address this issue, we construct an ultra-fast and powerful inference pipeline based on convolutional Deep Ensemble Networks. We explore the two main competing hypotheses for the GCE using a set of γEMs with increasing parametric freedom. We calculate the fractional contribution (f src ) of a dim population of MSPs to the total luminosity of the GCE and analyze its dependence on the complexity of the γEM. For the simplest γEM, we obtain fsrc = 0.10 ± 0.07, while the most complex model yields f src = 0.79 ± 0.24. In conclusion, we find that the statement about the nature of the GCE (dark matter or not) strongly depends on the assumed γEM. The quoted results for f src do not account for the additional uncertainty arising from the fact that the observed gamma-ray sky is out-of-distribution concerning the investigated γEM iterations. We quantify the reality gap between our γEMs using deep-learning-based One-Class Deep Support Vector Data Description networks, revealing that all employed γEMs have gaps to reality. Finally, our study casts doubt on the validity of previous conclusions regarding the GCE and dark matter, and underscores the urgent need to account for the reality gap and consider previously overlooked "out of domain" uncertainties in future interpretations.

79 ASTRONOMY AND ASTROPHYSICS↗

The deep-DRT: A deep neural network approach to deconvolve the distribution of relaxation times from multidimensional electrochemical impedance spectroscopy data

Electrochemical impedance spectroscopy (EIS) is an experimental technique ubiquitously used to study electrochemical systems. However, conventional EIS data interpretation through physical and equivalent circuit models is challenging because physical models are problem-specific, and equivalent circuits are often just lumped-element analogs lacking physical meaning. The distribution of relaxation times (DRT) has emerged as a complementary approach to resolve these issues. One drawback of conventional DRT deconvolution is that the EIS data is understood to be (only) a function of frequency (i.e. 1D data) and deconvolved accordingly. This work proposes a novel deconvolution method based on deep neural networks (DNNs), allowing the analysis of multidimensional EIS spectra to bridge data dependency on both frequencies and experimental conditions. Two particularly appealing traits of the deep-DRT method developed in this article are that neither regularization nor specific spacing on the state variables defining the experiment are required. Finally, leveraging DNN to examine complex EIS spectra and their dependence on experimental conditions, this work opens a new research direction in the area of EIS analysis and DRT deconvolution.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

Data augmentation for disruption prediction via robust surrogate models

The goal of this work is to generate large statistically representative data sets to train machine learning models for disruption prediction provided by data from few existing discharges. Such a comprehensive training database is important to achieve satisfying and reliable prediction results in artificial neural network classifiers. Here, we aim for a robust augmentation of the training database for multivariate time series data using Student t process regression. We apply Student t process regression in a state space formulation via Bayesian filtering to tackle challenges imposed by outliers and noise in the training data set and to reduce the computational complexity. Thus, the method can also be used if the time resolution is high. We use an uncorrelated model for each dimension and impose correlations afterwards via colouring transformations. We demonstrate the efficacy of our approach on plasma diagnostics data of three different disruption classes from the DIII-D tokamak. To evaluate if the distribution of the generated data is similar to the training data, we additionally perform statistical analyses using methods from time series analysis, descriptive statistics and classic machine learning clustering algorithms.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗

Stochastic 3D reconstruction of cracked polycrystalline NMC particles using 2D SEM data

Li-ion battery performance is strongly influenced by the 3D microstructure of its cathode particles. Cracks within these particles develop during calendaring and cycling, reducing connectivity but increasing reactive surface, making their impact on battery performance complex. Understanding these contradictory effects requires a quantitative link between particle morphology and battery performance. However, informative 3D imaging techniques are time-consuming, costly and rarely available, such that analyses often have to rely on 2D image data. This paper presents a novel stereological approach for generating virtual 3D cathode particles exhibiting crack networks that are statistically equivalent to those observed in 2D sections of experimentally measured particles. Consequently, 2D image data suffices for deriving a full 3D characterization of cracked cathodes particles. Such virtually generated 3D particles could serve as geometry input for spatially resolved electro-chemo-mechanical simulations to enhance our understanding of structure-property relationships of cathodes in Li-ion batteries.

36 MATERIALS SCIENCE↗

Structure of the SPRY domain of the human RNA helicase DDX1, a putative interaction platform within a DEAD-box protein

The structure of the SPRY domain of the human RNA helicase DDX1 was determined at 2.0 Å resolution. The SPRY domain provides a putative protein–protein interaction platform within DDX1 that differs from other SPRY domains in its structure and conserved regions. The human RNA helicase DDX1 in the DEAD-box family plays an important role in RNA processing and has been associated with HIV-1 replication and tumour progression. Whereas previously described DEAD-box proteins have a structurally conserved core, DDX1 shows a unique structural feature: a large SPRY-domain insertion in its RecA-like consensus fold. SPRY domains are known to function as protein–protein interaction platforms. Here, the crystal structure of the SPRY domain of human DDX1 (hDSPRY) is reported at 2.0 Å resolution. The structure reveals two layers of concave, antiparallel β-sheets that stack onto each other and a third β-sheet beneath the β-sandwich. A comparison with SPRY-domain structures from other eukaryotic proteins showed that the general β-sandwich fold is conserved; however, differences were detected in the loop regions, which were identified in other SPRY domains to be essential for interaction with cognate partners. In contrast, in hDSPRY these loop regions are not strictly conserved across species. Interestingly, though, a conserved patch of positive surface charge is found that may replace the connecting loops as a protein–protein interaction surface. The data presented here comprise the first structural information on DDX1 and provide insights into the unique domain architecture of this DEAD-box protein. By providing the structure of a putative interaction domain of DDX1, this work will serve as a basis for further studies of the interaction network within the hetero-oligomeric complexes of DDX1 and of its recruitment to the HIV-1 Rev protein as a viral replication factor.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗

A Data Processing Pipeline for Socio-Technical Network Analysis [Slides]

With the rapid adoption of emerging technologies, there is a need to catalog and model sociotechnical interdependencies that have been historically used to influence the operation of Critical Infrastructure networks including the impacts of mergers and acquisitions, hostile takeovers, and foreign investment. Our research intends to address this need with two primary contributions. First, we have developed a data curation and processing pipeline to generate sociotechnical networks extracted from a variety of data sources including SEC filings and infrastructure asset databases. The pipeline, implemented in Apache Airflow, extracts and normalizes the representation of entities and relations, specified within ontologies. Second, networks produced by our pipeline enable the development of graph-theoretic metrics that consider the properties of network components in addition to its topology. Measures of network complexity, such as degree distribution, reachability analyses, temporal analysis, and community detection may be adapted to indicate adversarial organizational influence. Our intent is to provide an extensible, machine-actionable approach to quickly communicate such models, reproduce previous results, and adapt them to new, unanticipated situations.

97 MATHEMATICS AND COMPUTING↗

A Review of Physics-Informed Machine Learning in Fluid Mechanics

Physics-informed machine-learning (PIML) enables the integration of domain knowledge with machine learning (ML) algorithms, which results in higher data efficiency and more stable predictions. This provides opportunities for augmenting—and even replacing—high-fidelity numerical simulations of complex turbulent flows, which are often expensive due to the requirement of high temporal and spatial resolution. In this review, we (i) provide an introduction and historical perspective of ML methods, in particular neural networks (NN), (ii) examine existing PIML applications to fluid mechanics problems, especially in complex high Reynolds number flows, (iii) demonstrate the utility of PIML techniques through a case study, and (iv) discuss the challenges and opportunities of developing PIML for fluid mechanics.

42 ENGINEERING↗

Satellite Precipitation and Cloud Experiment (SPACE)

An experiment plan was formulated to support an investigation of the precipitation processes associated with mesoscale systems and the interaction of rain producing cloud complexes with the meso and macro-scale environment. The field program necessary to accomplish the scientific goals is designed to incorporate satellite information (VIS and IR as well as VAS data), special radiosonde and surface observing capabilities, radar observations of storms development, special observations from aircraft and ground based lightning measurements. The observing network comprises rawinsonde stations located in northern Alabama and eastern Tennessee. A high density raingauge network and the MSFC lightning location network also cover the area. Coordination with NASA U2 and ER2 aircraft programs and with the operational requirements of the microburst severe thunderstorm program can assure the most efficient utilization of the observing resources.

Arnold, J. E.↗

Advancing Methodologies for Applying Machine Learning and Evaluating Spatiotemporal Models of Fine Particulate Matter (PM 2.5 ) Using Satellite Data Over Large Regions

Reconstructing the distribution of fine particulate matter (PM 2.5 ) in space and time, even far from ground monitoring sites, is an important exposure science contribution to epidemiologic analyses of PM 2.5 health impacts. Flexible statistical methods for prediction have demonstrated the integration of satellite observations with other predictors, yet these algorithms are susceptible to overfitting the spatiotemporal structure of the training datasets. We present a new approach for predicting PM 2.5 using machine-learning methods and evaluating prediction models for the goal of making predictions where they were not previously available. We apply extreme gradient boosting (XGBoost) modeling to predict daily PM 2.5 on a 1 x 1 km 2 resolution for a 13 state region in the Northeastern USA for the years 2000–2015 using satellite-derived aerosol optical depth and implement a recursive feature selection to develop a parsimonious model. We demonstrate excellent predictions of withheld observations but also contrast an RMSE of 3.11 μg/m 3 in our spatial cross-validation withholding nearby sites versus an overfit RMSE of 2.10 μg/m 3 using a more conventional random ten-fold splitting of the dataset. As the field of exposure science moves forward with the use of advanced machine-learning approaches for spatiotemporal modeling of air pollutants, our results show the importance of addressing data leakage in training, overfitting to spatiotemporal structure, and the impact of the predominance of ground monitoring sites in dense urban sub-networks on model evaluation. The strengths of our resultant modeling approach for exposure in epidemiologic studies of PM 2.5 include improved efficiency, parsimony, and interpretability with robust validation while still accommodating complex spatiotemporal relationships.

air pollution↗

Real-Time Prediction of Power System Frequency in FNET: A State Space Approach

This paper proposes a novel approach to predict power frequency by applying a state-space model to describe the time-varying nature of power systems. It introduces the Expectation maximization (EM) and prediction error minimization (PEM) algorithms to dynamically estimate the parameters of the model. This paper discusses how the proposed models can be used to ensure the efficiency and reliability of power systems in Frequency Monitoring Network (FNET), if serious frequency fluctuation or measurement failure occur at some nodes; this is achieved without requiring the exact model of complex power systems. Our approach leads to an easy online implementation with high precision and short response time that are key to effective frequency control. We randomly pick a set of frequency data for one power station in FNET and use it to estimate and predict the power frequency based on past measurements. Several computer simulations are provided to evaluate the method. Simulation results showed that the proposed technique could achieve good performance regarding the frequency monitoring with very limited measurement input information.

Djouadi, Seddik M↗

Wide‐Field Bond Quality Evaluation Using Frequency Domain Thermoreflectance with Deep Neural Network Feature Reconstruction

Heterogeneous integration of microelectronic components provides a pathway to improve circuit/component performance; however, this comes with assembly challenges, in particular due to complex interfaces via subsurface bump bonds. The ability of these bonds to transmit electrical signals and conduct heat to the carrier substrate limits component performance. In this work, hyperspectral frequency‐domain thermoreflectance (FDTR) imaging is demonstrated as a robust technique for evaluating the quality of subsurface indium bump bonds in a surrogate microelectronic sample. By performing microscale FDTR imaging with coarse motion image stitching, thermal phase maps that cover a 4 mm by 4 mm field‐of‐view with subsurface feature sensitivity at depths greater than 50 µm are obtained. The resulting FDTR hyperspectral data contains more than three million pixels and reveal the quality of subsurface microbump arrays. Wide‐field analysis of bonded versus gap regions is enabled by deep neural network feature reconstruction, that after training, rapidly provides an interpretable representation of bond quality. Utility of noisy higher frequency FDTR phase maps, i.e., near the computationally predicted sensing depth limit, results in an average prediction error of 11%. Taken together, FDTR with neural network‐based analysis demonstrates subsurface bond monitoring at length scales relevant for heterogeneously integrated microelectronics.

FDTR↗

Quantifying uncertainty for deep learning based forecasting and flow-reconstruction using neural architecture search ensembles

Classical problems in computational physics such as data-driven forecasting and signal reconstruction from sparse sensors have recently seen an explosion in deep neural network (DNN) based algorithmic approaches. However, most DNN models do not provide uncertainty estimates, which are crucial for establishing the trustworthiness of these techniques in downstream decision making tasks and scenarios. In recent years, ensemble-based methods have achieved significant success for the uncertainty quantification in DNNs on a number of benchmark problems. However, their performance on real-world applications remains under-explored. In this work, we present an automated approach to DNN discovery and demonstrate how this may also be utilized for ensemble-based uncertainty quantification. Specifically, we propose the use of a scalable neural and hyperparameter architecture search for discovering an ensemble of DNN models for complex dynamical systems. We highlight how the proposed method not only discovers high-performing neural network ensembles for our tasks, but also quantifies uncertainty seamlessly. This is achieved by using genetic algorithms and Bayesian optimization for sampling the search space of neural network architectures and hyperparameters. Subsequently, a model selection approach is used to identify candidate models for an ensemble set construction. Afterwards, a variance decomposition approach is used to estimate the uncertainty of the predictions from the ensemble. We demonstrate the feasibility of this framework for two tasks — forecasting from historical data and flow reconstruction from sparse sensors for the sea-surface temperature. In conclusion, we demonstrate superior performance from the ensemble in contrast with individual high-performing models and other benchmarks.

Deep ensembles↗