SEARCH · Engineering Papers
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.
An approach to the formulation of fuzzy if-then rules based on clustering objective functions is proposed. The membership functions are then calibrated with the generalized neural networks technique to achieve a desired input-output mapping. The learning procedure is basically a gradient-descent algorithm. A Kalman filter algorithm is used to improve the overall performance.
To address the unprecedented scale of HL-LHC data, the Exa.TrkX project is investigating a variety of machine learning approaches to particle track reconstruction. The most promising of these solutions, graph neural networks (GNN), process the event as a graph that connects track measurements (detector hits corresponding to nodes) with candidate line segments between the hits (corresponding to edges). Detector information can be associated with nodes and edges, enabling a GNN to propagate the embedded parameters around the graph and predict node-, edge- and graph-level observables. Previously, message-passing GNNs have shown success in predicting doublet likelihood, and we here report updates on the state-of-the-art architectures for this task. In addition, the Exa.TrkX project has investigated innovations in both graph construction, and embedded representations, in an effort to achieve fully learned end-to-end track finding. Hence, we present a suite of extensions to the original model, with encouraging results for hitgraph classification. In addition, we explore increased performance by constructing graphs from learned representations which contain non-linear metric structure, allowing for efficient clustering and neighborhood queries of data points. We demonstrate how this framework fits in with both traditional clustering pipelines, and GNN approaches. The embedded graphs feed into high-accuracy doublet and triplet classifiers, or can be used as an end-to-end track classifier by clustering in an embedded space. A set of post-processing methods improve performance with knowledge of the detector physics. Finally, we present numerical results on the TrackML particle tracking challenge dataset, where our framework shows favorable results in both seeding and track finding.
We propose a novel prediction interval (PI) method for uncertainty quantification, which addresses three major issues with the state-of-the-art PI methods. First, existing PI methods require retraining of neural networks (NNs) for every given confidence level and suffer from the crossing issue in calculating multiple PIs. Second, they usually rely on customized loss functions with extra sensitive hyperparameters for which fine tuning is required to achieve a well-calibrated PI. Third, they usually underestimate uncertainties of out-of-distribution (OOD) samples leading to over-confident PIs. Our PI3NN method calculates PIs from linear combinations of three NNs, each of which is independently trained using the standard mean squared error loss. The coefficients of the linear combinations are computed using root-finding algorithms to ensure tight PIs for a given confidence level. We theoretically prove that PI3NN can calculate PIs for a series of confidence levels without retraining NNs and it completely avoids the crossing issue. Additionally, PI3NN does not introduce any unusual hyperparameters resulting in a stable performance. Furthermore, we address OOD identification challenge by introducing an initialization scheme which provides reasonably larger PIs of the OOD samples than those of the in-distribution samples. Benchmark and real-world experiments show that our method outperforms several state-of-the-art approaches with respect to predictive uncertainty quality, robustness, and OOD samples identification.
ABSTRACT We present in this paper one of the largest galaxy morphological classification catalogues to date, including over 20 million galaxies, using the Dark Energy Survey (DES) Year 3 data based on convolutional neural networks (CNNs). Monochromatic i-band DES images with linear, logarithmic, and gradient scales, matched with debiased visual classifications from the Galaxy Zoo 1 (GZ1) catalogue, are used to train our CNN models. With a training set including bright galaxies (16 ≤ i < 18) at low redshift (z < 0.25), we furthermore investigate the limit of the accuracy of our predictions applied to galaxies at fainter magnitude and at higher redshifts. Our final catalogue covers magnitudes 16 ≤ i < 21, and redshifts z < 1.0, and provides predicted probabilities to two galaxy types – ellipticals and spirals (disc galaxies). Our CNN classifications reveal an accuracy of over 99 per cent for bright galaxies when comparing with the GZ1 classifications (i < 18). For fainter galaxies, the visual classification carried out by three of the co-authors shows that the CNN classifier correctly categorizes discy galaxies with rounder and blurred features, which humans often incorrectly visually classify as ellipticals. As a part of the validation, we carry out one of the largest examinations of non-parametric methods, including ∼100 ,000 galaxies with the same coverage of magnitude and redshift as the training set from our catalogue. We find that the Gini coefficient is the best single parameter discriminator between ellipticals and spirals for this data set.
Developing accurate, efficient, and robust closure models is essential in the construction of reduced order models (ROMs) for realistic nonlinear systems, which generally require drastic ROM mode truncations. We propose a deep residual neural network (ResNet) closure learning framework for ROMs of nonlinear systems. The novel ResNet-ROM framework consists of two steps: (i) In the first step, we use ROM projection to filter the given nonlinear system and construct a spatially filtered ROM. This filtered ROM is low-dimensional, but is not closed. (ii) In the second step, we use ResNet to close the filtered ROM, i.e., to model the interaction between the resolved and unresolved ROM modes. We emphasize that in the new ResNet-ROM framework, data is used only to complement classical physical modeling (i.e., only in the closure modeling component), not to completely replace it. We also note that the new ResNet-ROM is built on general ideas of spatial filtering and deep learning and is independent of (restrictive) phenomenological arguments, e.g., of eddy viscosity type. The numerical experiments for the 1D Burgers equation show that the ResNet-ROM is significantly more accurate than the standard projection ROM. The new ResNet-ROM is also more accurate and significantly more efficient than other modern ROM closure models.
A minimization of learning-algorithm completion time is sought in the present optimal-mapping study of the learning process in multilayer feed-forward artificial neural networks (ANNs) for message-passing multicomputers. A novel approximation algorithm for mappings of this kind is derived from observations of the dominance of a parallel ANN algorithm over its communication time. Attention is given to both static and dynamic mapping schemes for systems with static and dynamic background workloads, as well as to experimental results obtained for simulated mappings on multicomputers with dynamic background workloads.
Elements of array programmed with help of ultraviolet light. A 32 x 32 very-large-scale integrated-circuit array of electronic synapses serves as building-block chip for analog neural-network computer. Synaptic weights stored in nonvolatile manner. Makes information content of array invulnerable to loss of power, and, by eliminating need for circuitry to refresh volatile synaptic memory, makes architecture simpler and more compact.
Measuring the structural parameters (size, total brightness, light concentration, etc.) of galaxies is a significant first step towards a quantitative description of different galaxy populations. In this work, we demonstrate that a Bayesian Neural Network (BNN) can be used for the inference, with uncertainty quantification, of such morphological parameters from simulated low-surface-brightness galaxy images. Compared to traditional profile-fitting methods, we show that the uncertainties obtained using BNNs are comparable in magnitude, well-calibrated, and the point estimates of the parameters are closer to the true values. Our method is also significantly faster, which is very important with the advent of the era of large galaxy surveys and big data in astrophysics.
The likelihood ratio is a crucial quantity for statistical inference in science that enables hypothesis testing, construction of confidence intervals, reweighting of distributions, and more. Many modern scientific applications, however, make use of data- or simulation-driven models for which computing the likelihood ratio can be very difficult or even impossible. By applying the so-called “likelihood ratio trick,” approximations of the likelihood ratio may be computed using clever parametrizations of neural network-based classifiers. A number of different neural network setups can be defined to satisfy this procedure, each with varying performance in approximating the likelihood ratio when using finite training data. We present a series of empirical studies detailing the performance of several common loss functionals and parametrizations of the classifier output in approximating the likelihood ratio of two univariate and multivariate Gaussian distributions as well as simulated high-energy particle physics datasets.
This study investigates the feasibility of reconstructing the last closed flux surface in the DIII-D tokamak using neural network models trained on reduced input feature sets, addressing an ill-posed task. Two models are compared: one trained solely on coil currents and another incorporating coil currents, plasma current and loop voltage. The model trained exclusively on coil currents achieved a mean point displacement of $0.04$ m on a held-out test set, while the inclusion of plasma current and loop voltage reduced the error to $0.03$ m. This comparison highlights the trade-offs between input feature complexity and reconstruction accuracy, demonstrating the potential of machine learning algorithms to perform effectively in data-limited environments, such as those expected in fusion power plants due to diagnostic constraints imposed by the presence of blankets and shielding.
One of the major challenges in deep learning is retrieving sufficiently large labeled training datasets, which can become expensive and time consuming to collect. A unique approach to training segmentation is to use Deep Neural Network (DNN) models with a minimal amount of initial labeled training samples. The procedure involves creating synthetic data and using image registration to calculate affine transformations to apply to the synthetic data. The method takes a small dataset and generates a highquality augmented reality synthetic dataset with strong variance while maintaining consistency with real cases. Results illustrate segmentation improvements in various target features and increased average target confidence.
Smart networks such as microgrid (MG) and active distribution networks (ADN) have been recently playing an important role in power system operation. The design and implementation of appropriate protection systems for such networks must be addressed, which imposes new technical challenges. This paper presents the implementation and validation aspects of an adaptive fault detection strategy based on neural networks (NNs) and multiple sampling points for ADN and MG. The solution is implemented on an edge device. Artificial NNs are used to derive a data-driven model that uses only local measurements to detect fault states of the network without the need for communication infrastructure. Multiple sampling points are used to derive a data-driven model, which allows the generalization considering the implementation in physical systems. The adaptive fault detector model is implemented on a Jetson Nano system, which is a single-board computer (SBC) with a small Graphic Processing Unit (GPU) intended to run machine learning loads at the edge. The proposed method is tested in a physical, real-life, low-voltage network located at Universidad del Norte, Colombia. This testing network is based on the IEEE-13 Node Test Feeder scaled down to 220 V. The validation in a simulation environment shows the accuracy and dependability above 99.6%, while the real-time tests show the accuracy and dependability of 95.5% and 100%, respectively. Without hard-to-derive parameters, the easy-to-implement embedded model highlights the potential for real-life applications.
Many lobbyists in Washington have argued that artificial intelligence (AI) is an alternative to manned space activity. In actuality, this is the opposite of the truth, especially as regards artificial neural networks (ANNs), that form of AI which has the greatest hope of mimicking human abilities in learning, ability to interface with sensors and actuators, flexibility and balanced judgement. ANNs and their relation to expert systems (the more traditional form of AI), and the limitations of both technologies are briefly reviewed. A Few highlights of recent work on ANNs, including an NSF-sponsored workshop on ANNs for control applications are given. Current thinking on ANNs for use in certain key areas (the National Aerospace Plane, teleoperation, the control of large structures, fault diagnostics, and docking) which may be crucial to the long term future of man in space is discussed.
Electrical resistance of new solid-state device altered repeatedly by suitable control signals, yet remains at steady value when control signal removed. Resistance set at low value ("on" state), high value ("off" state), or at any convenient intermediate value and left there until new value desired. Circuits of this type particularly useful in nonvolatile, associative electronic memories based on models of neural networks. Such programmable analog memory resistors ideally suited as synaptic interconnects in "self-learning" neural nets. Operation of device depends on electrochromic property of WO3, which when pure is insulator. Potential uses include nonvolatile, erasable, electronically programmable read-only memories.
We evaluate uncertainty quantification (UQ) methods for deep learning applied to liquid argon time projection chamber (LArTPC) physics analysis tasks. As deep learning applications enter widespread usage among physics data analysis, neural networks with reliable estimates of prediction uncertainty and robust performance against overconfidence and out-of-distribution (OOD) samples are critical for their full deployment in analyzing experimental data. While numerous UQ methods have been tested on simple datasets, performance evaluations for more complex tasks and datasets are scarce. Here we assess the application of selected deep learning UQ methods on the task of particle classification using the PiLArNet monte carlo 3D LArTPC point cloud dataset. We observe that UQ methods not only allow for better rejection of prediction mistakes and OOD detection, but also generally achieve higher overall accuracy across different task settings. We assess the precision of uncertainty quantification using different evaluation metrics, such as distributional separation of prediction entropy across correctly and incorrectly identified samples, receiver operating characteristic curves (ROCs), and expected calibration error from observed empirical accuracy. We conclude that ensembling methods can obtain well calibrated classification probabilities and generally perform better than other existing methods in deep learning UQ literature.
We present a novel time-domain method for the detection of faulty bearings that has direct applicability to monitoring the health of the turbo pumps on the Space Shuttle Main Engine. A feed-forward neural network was trained to detect modelled roller bearing faults on the basis of the periodicity of impact pulse trains. The network's performance was dependent upon the number of pulses in the network's input window and the signal-to-noise ratio of the input signal. To test the model's validity, we fit the model's parameters to an actual vibration signal generated by a faulty roller element bearing and applied the network trained on this model to detect faults in actual vibration data. When this network was tested on the actual vibration data, it correctly identified the vibration signal as a fault condition 76 percent of the time.
Abstract. Rivers and river habitats around the world are under sustained pressure from human activities and the changing global environment. Our ability to quantify and manage the river states in a timely manner is critical for protecting the public safety and natural resources. In recent years, vector-based river network models have enabled modeling of large river basins at increasingly fine resolutions, but are computationally demanding. This work presents a multistage, physics-guided, graph neural network (GNN) approach for basin-scale river network learning and streamflow forecasting. During training, we train a GNN model to approximate outputs of a high-resolution vector-based river network model; we then fine-tune the pretrained GNN model with streamflow observations. We further apply a graph-based, data-fusion step to correct prediction biases. The GNN-based framework is first demonstrated over a snow-dominated watershed in the western United States. A series of experiments are performed to test different training and imputation strategies. Results show that the trained GNN model can effectively serve as a surrogate of the process-based model with high accuracy, with median Kling–Gupta efficiency (KGE) greater than 0.97. Application of the graph-based data fusion further reduces mismatch between the GNN model and observations, with as much as 50 % KGE improvement over some cross-validation gages. To improve scalability, a graph-coarsening procedure is introduced and is demonstrated over a much larger basin. Results show that graph coarsening achieves comparable prediction skills at only a fraction of training cost, thus providing important insights into the degree of physical realism needed for developing large-scale GNN-based river network models.
In this paper, Multilayer Feedforward Networks are applied to the robot inverse kinematic problem. The networks are trained with endeffector position and joint angles. After training, performance is measured by having the network generate joint angles for arbitrary endeffector trajectories. A 3-degree-of-freedom (DOF) spatial manipulator is used for the study. It is found that neural networks provide a simple and effective way to both model the manipulator inverse kinematics and circumvent the problems associated with algorithmic solution methods.