Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Deep Neural Networks”

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 379 records · Page 21

A Classification of G -invariant Shallow Neural Networks

When trying to fit a deep neural network (DNN) to a G-invariant target function with G a group, it only makes sense to constrain the DNN to be G-invariant as well. However, there can be many different ways to do this, thus raising the problem of “G-invariant neural architecture design”: What is the optimal Ginvariant architecture for a given problem? Before we can consider the optimization problem itself, we must understand the search space, the architectures in it, and how they relate to one another. In this paper, we take a first step towards this goal; we prove a theorem that gives a classification of all G-invariant single-hidden-layer or “shallow” neural network (G-SNN) architectures with ReLU activation for any finite orthogonal group G, and we prove a second theorem that characterizes the inclusion maps or “network morphisms” between the architectures that can be leveraged during neural architecture search (NAS). The proof is based on a correspondence of every G-SNN to a signed permutation representation of G acting on the hidden neurons; the classification is equivalently given in terms of the first cohomology classes of G, thus admitting a topological interpretation. The G-SNN architectures corresponding to nontrivial cohomology classes have, to our knowledge, never been explicitly identified in the literature previously. Using a code implementation, we enumerate the G-SNN architectures for some example groups G and visualize their structure. Lastly, we prove that architectures corresponding to inequivalent cohomology classes coincide in function space only when their weight matrices are zero, and we discuss the implications of this for NAS.

Agrawal, Devanshu↗

Machine learning coupled multi-scale modeling for redox flow batteries

The reaction distribution in macro or device-scale has been studied for redox flow batteries. The reaction distribution on electrode pore-scale structure however is not well understood, lacking especially on how the reaction distribution on the pore-scale may impact the overall performance of a flow battery. This study introduces for the first time a framework of a multi-scale model that provides understanding of the relationship between the pore-scale electrode structure reaction and the device-scale electrochemical reaction uniformity within the flow battery. A reduced order model is constructed based on 128 pore-scale simulations, which provide a quantitative relationship between the battery operation conditions (inlet velocity, current density, inlet concentration) and the surface reaction uniformity for the pore-scale sample. The multi-scale framework upscales this pore-scale surface reaction uniformity to device-scale combined uniformity. Based on the multi-scale model, a time-varying optimization of the inlet velocity is established, leading to significant reduction on pump power consumption with targeted surface reaction uniformity. The multi-scale model establishes the critical link between the micro-structure of a flow battery component and its performance at the macro-scale, therefore providing rationale for further operational or material optimization.

flow batteries, machine learning, multi-scale mode↗

Closure Learning for Nonlinear Model Reduction Using Deep Residual Neural Network

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.

Xie, Xuping↗

Evaluating Automated Face Identity-Masking Methods with Human Perception and a Deep Convolutional Neural Network

Face de-identification (or “masking”) algorithms have been developed in response to the prevalent use of video recordings in public places. Here, we evaluated the success of face identity masking for human perceivers and a deep convolutional neural network (DCNN). Eight de-identification algorithms were applied to videos of drivers’ faces, while they actively operated a motor vehicle. These masks were pre-selected to be applicable to low-quality video and to maintain coarse information about facial actions. Humans studied high-resolution images to learn driver identities and were tested on their recognition of active drivers in low-resolution videos. Faces in the videos were either unmasked or were masked by one of the eight algorithms. When participants were tested immediately after learning (Experiment 1), all masks reduced identification, with six of eight masks reducing identification to extremely poor performance. In a second experiment, two of the most effective masks were tested after a delay of 7 or 28 days. The delay did not further reduce identification of the masked faces. In all masked conditions, participants maintained stringent decision criteria, with low confidence in recognition, further indicating the effectiveness of the masks. Next, the DCNN performed an identity-matching task between high-resolution images and masked videos—a task analogous to that done by humans. The pattern of accuracy for the DCNN mirrored some, but not all, aspects of human performance, highlighting the need to test the effectiveness of identity masking for both humans and machines. The DCNN was also tested on its ability to match identity between masked and unmasked versions of the same video, based only on the face. DCNN performance for the eight masks offers insight into the nature of the information in faces that is coded in these networks.

97 MATHEMATICS AND COMPUTING↗

Computer-aided detection using non-convolutional neural network Gaussian processes

Deep convolutional neural networks (CNNs) have in recent years achieved record-breaking performance on many image classification tasks and are therefore well-suited for computer aided detection (CAD). The need for uncertainty quantification for CAD motivates the need for a probabilistic framework for deep learning. The most well-known probabilistic neural network model is the Bayesian neural network (BNN), but BNNs are notoriously difficult to sample for large complex network architectures, and as such their use is restricted to small problems. It is known that the limit of BNNs as their widths increase toward infinity is a Gaussian process (GP), and there has been considerable research interest in these infinitely wide BNNs. Recently, this classic result has been extended to deep architectures in what is termed the neural network Gaussian process (NNGP) model. In this work, we implement an NNGP model and apply it to the ChestXRay14 dataset at the full resolution of 1024x1024 pixels. Even without any convolutional aspects to the network architecture and without any data augmentation, our five layer deep NNGP model outperforms other non-convolutional models and therefore helps to narrow the performance gap between non-convolutional and convolutional models. Our NNGP model is fully Bayesian and therefore offers uncertainty information through its predictive variance that can be used to formulate a predictive confidence measure. We show that the performance of the NNGP model is significantly boosted after low-confidence predictions are rejected, suggesting that convolution is most beneficial only for these low-confidence examples. Finally, our results indicate that an extremely large fully-connected neural network with appropriate regularization could perform as well as the NNGP if not for the computational bottleneck resulting from the large number of model parameters.

Agrawal, Devanshu↗

Prolongation of SMAP to Spatiotemporally Seamless Coverage of Continental U.S. Using a Deep Learning Neural Network

The Soil Moisture Active Passive (SMAP) mission has delivered valuable sensing of surface soil moisture since 2015. However, it has a short time span and irregular revisit schedules. Utilizing a state-of-the-art time series deep learning neural network, Long Short-Term Memory (LSTM), we created a system that predicts SMAP level-3 moisture product with atmospheric forcings, model-simulated moisture, and static physiographic attributes as inputs. The system removes most of the bias with model simulations and improves predicted moisture climatology, achieving small test root-mean-square errors (<0.035) and high-correlation coefficients >0.87 for over 75% of Continental United States, including the forested southeast. As the first application of LSTM in hydrology, we show the proposed network avoids overfitting and is robust for both temporal and spatial extrapolation tests. LSTM generalizes well across regions with distinct climates and environmental settings. With high fidelity to SMAP, LSTM shows great potential for hindcasting, data assimilation, and weather forecasting.

54 ENVIRONMENTAL SCIENCES↗

A Deep Gravity model for mobility flows generation

The movements of individuals within and among cities influence critical aspects of our society, such as well-being, the spreading of epidemics, and the quality of the environment. When information about mobility flows is not available for a particular region of interest, we must rely on mathematical models to generate them. In this work, we propose Deep Gravity, an effective model to generate flow probabilities that exploits many features (e.g., land use, road network, transport, food, health facilities) extracted from voluntary geographic data, and uses deep neural networks to discover non-linear relationships between those features and mobility flows. Our experiments, conducted on mobility flows in England, Italy, and New York State, show that Deep Gravity achieves a significant increase in performance, especially in densely populated regions of interest, with respect to the classic gravity model and models that do not use deep neural networks or geographic data. Deep Gravity has good generalization capability, generating realistic flows also for geographic areas for which there is no data availability for training. Finally, we show how flows generated by Deep Gravity may be explained in terms of the geographic features and highlight crucial differences among the three considered countries interpreting the model’s prediction with explainable AI techniques.

97 MATHEMATICS AND COMPUTING↗

Radar Super Resolution using a Deep Convolutional Neural Network

Super-resolution involves synthetically increasing the resolution of gridded data beyond its native resolution. Typically, this is done using interpolation schemes, which estimate sub-grid scale values from neighboring data, and perform the same operation everywhere regardless of the large-scale context, or by requiring a network of radars with overlapping fields of view. Recently, significant progress has been made in single image super resolution using convolutional neural networks. Conceptually, a neural network may be able to learn relations between large scale precipitation features and the associated sub-pixel scale variability and outperform interpolation schemes. Here, we use a deep convolutional neural network to artificially enhance the resolution of NEXRAD PPI scans. The model is trained on 6-months of reflectivity observations from the Langley Hill WA (KLGX) radar, and we find that it substantially outperforms common interpolation schemes for x4 and x8 resolution increases based on several objective error and perceptual quality metrics.

radar, machine learning, super resolution, Remote ↗

Neuromorphic Accelerator for Deep Spiking Neural Networks with NVM Crossbar Arrays

In this paper, we present a scalable digital hardware accelerator based on non-volatile memory arrays capable of realizing deep convolutional spiking neural networks (SNNs). Our design studies are conducted using a compact model for spin-transfer torque random access memory (STT-RAM) devices. Large networks are realized by tiling multiple cores which communicate by transmitting spike packets via an on-chip routing network. Compared to an equivalent SRAM based core design, we show that the STT-RAM based design achieves nearly 15X higher GSOPS (Synaptic Operations per Second) per Watt per mm 2 making it a promising platform for realizing systems with significant area and power limitations.

Kulkarni, Shruti↗

Development and Evaluation of a General Drag Model for Gas-Solid Flows via Deep Learning

This project presents the development and evaluation of a general drag model for gas–solid multiphase flows using deep learning techniques. A comprehensive database of more than 4,000 experimental and numerical data points for spherical and non spherical particles was compiled, incorporating geometric features such as sphericity, aspect ratio, and orientation. Several predictive approaches—including traditional em pirical correlations, machine learning, and deep neural networks—were benchmarked, with the proposed Drag Coefficient Correlation-aided Deep Neural Network (DCC DNN) demonstrating superior accuracy. To account for particle–particle interactions, additional drag data were generated using CFD-based simulations of packed and flu idized beds, leading to the development of a retrained model capable of incorporat ing volume fraction effects. Integration of the trained model with the MFiX CFD solver was achieved using FTorch, enabling drag predictions during discrete element method (DEM) simulations. Validation against experimental data for single particles and fluidized beds confirmed the model’s improved predictive ability, particularly for non-spherical geometries. While the model performed strongly under fluidized con ditions, limitations remained in unfluidized regimes, suggesting a need for expanded datasets. Overall, this study demonstrates the feasibility of combining deep learning with physics-informed CFD to improve drag modeling for gas–solid flows, with promis ing implications for scaling multiphase simulations in industrial applications.

42 ENGINEERING↗

Distilling particle knowledge for fast reconstruction at high-energy physics experiments

Knowledge distillation is a form of model compression that allows artificial neural networks of different sizes to learn from one another. Its main application is the compactification of large deep neural networks to free up computational resources, in particular on edge devices. In this article, we consider proton-proton collisions at the High-Luminosity Large Hadron Collider (HL-LHC) and demonstrate a successful knowledge transfer from an event-level graph neural network (GNN) to a particle-level small deep neural network (DNN). Our algorithm, DistillNet, is a DNN that is trained to learn about the provenance of particles, as provided by the soft labels that are the GNN outputs, to predict whether or not a particle originates from the primary interaction vertex. The results indicate that for this problem, which is one of the main challenges at the HL-LHC, there is minimal loss during the transfer of knowledge to the small student network, while improving significantly the computational resource needs compared to the teacher. This is demonstrated for the distilled student network on a CPU, as well as for a quantized and pruned student network deployed on a field programmable gate array. Our study proves that knowledge transfer between networks of different complexity can be used for fast artificial intelligence (AI) in high-energy physics that improves the expressiveness of observables over non-AI-based reconstruction algorithms. Such an approach can become essential at the HL-LHC experiments, e.g. to comply with the resource budget of their trigger stages.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

Low-dose x-ray tomography through a deep convolutional neural network

Synchrotron-based X-ray tomography offers the potential of rapid large-scale reconstructions of the interiors of materials and biological tissue at fine resolution. However, for radiation sensitive samples, there remain fundamental trade-offs between damaging samples during longer acquisition times and reducing signals with shorter acquisition times. We present a deep convolutional neural network (CNN) method that increases the acquired X-ray tomographic signal by at least a factor of 10 during low-dose fast acquisition by improving the quality of recorded projections. Short exposure time projections enhanced with CNN show similar signal to noise ratios as compared with long exposure time projections and much lower noise and more structural information than low-dose fats acquisition without CNN. We optimized this approach using simulated samples and further validated on experimental nano-computed tomography data of radiation sensitive mouse brains acquired with a transmission X-ray microscopy. We demonstrate that automated algorithms can reliably trace brain structures in datasets collected with low dose-CNN. As a result, this method can be applied to other tomographic or scanning based X-ray imaging techniques and has great potential for studying faster dynamics in specimens.

60 APPLIED LIFE SCIENCES↗

Transfer learning for analysis of collective and non-collective Thomson scattering spectra

Thomson scattering (TS) diagnostics provide reliable, minimally perturbative measurements of fundamental plasma parameters, such as electron density (⁠n e ) and electron temperature (⁠T e ⁠). Deep neural networks can provide accurate estimates of ⁠n e and T e when conventional fitting algorithms may fail, such as when TS spectra are dominated by noise, or when fast analysis is required for real-time operation. Although deep neural networks typically require large training sets, transfer learning can improve model performance on a target task with limited data by leveraging pre-trained models from related source tasks, where select hidden layers are further trained using target data. We present five architecturally diverse deep neural networks, pre-trained on synthetic TS data and adapted for experimentally measured TS data, to evaluate the efficacy of transfer learning in estimating n e and T e in both the collective and non-collective scattering regimes. We evaluate errors in n e and T e estimates as a function of training set size for models trained with and without transfer learning, and we observe decreases in model error from transfer learning when the training set contains ≲ 200 experimentally measured spectra.

Artificial neural networks↗

Wire-cell 3D pattern recognition techniques for neutrino event reconstruction in large LArTPCs: algorithm description and quantitative evaluation with MicroBooNE simulation

Wire-Cell is a 3D event reconstruction package for liquid argon time projection chambers. Through geometry, time, and drifted charge from multiple readout wire planes, 3D space points with associated charge are reconstructed prior to the pattern recognition stage. Pattern recognition techniques, including track trajectory and dQ/dx (ionization charge per unit length) fitting, 3D neutrino vertex fitting, track and shower separation, particle-level clustering, and particle identification are then applied on these 3D space points as well as the original 2D projection measurements. A deep neural network is developed to enhance the reconstruction of the neutrino interaction vertex. Compared to traditional algorithms, the deep neural network boosts the vertex efficiency by a relative 30% for charged-current ν e interactions. Therefore, this pattern recognition achieves 80–90% reconstruction efficiencies for primary leptons, after a 65.8% (72.9%) vertex efficiency for charged-current ν e (ν μ ) interactions. Based on the resulting reconstructed particles and their kinematics, we also achieve 15-20% energy reconstruction resolutions for charged-current neutrino interactions.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

SU-C-207B-07: Deep Convolutional Neural Network Image Matching for Ultrasound Guidance in Radiotherapy

Purpose: Robust matching of ultrasound images is a challenging problem as images of the same anatomy often present non-trivial differences. This poses an obstacle for ultrasound guidance in radiotherapy. Thus our objective is to overcome this obstacle by designing and evaluating an image blocks matching framework based on a two channel deep convolutional neural network. Methods: We extend to 3D an algorithmic structure previously introduced for 2D image feature learning [1]. To obtain the similarity between two 3D image blocks A and B, the 3D image blocks are divided into 2D patches Ai and Bi. The similarity is then calculated as the average similarity score of Ai and Bi. The neural network was then trained with public non-medical image pairs, and subsequently evaluated on ultrasound image blocks for the following scenarios: (S1) same image blocks with/without shifts (A and A-shift-x); (S2) non-related random block pairs; (S3) ground truth registration matched pairs of different ultrasound images with/without shifts (A-i and A-reg-i-shift-x). Results: For S1 the similarity scores of A and A-shift-x were 32.63, 18.38, 12.95, 9.23, 2.15 and 0.43 for x=ranging from 0 mm to 10 mm in 2 mm increments. For S2 the average similarity score for non-related block pairs was −1.15. For S3 the average similarity score of ground truth registration matched blocks A-i and A-reg-i-shift-0 (1≤i≤5) was 12.37. After translating A-reg-i-shift-0 by 0 mm, 2 mm, 4 mm, 6 mm, 8 mm, and 10 mm, the average similarity scores of A-i and A-reg-i-shift-x were 11.04, 8.42, 4.56, 2.27, and 0.29 respectively. Conclusion: The proposed method correctly assigns highest similarity to corresponding 3D ultrasound image blocks despite differences in image content and thus can form the basis for ultrasound image registration and tracking.[1] Zagoruyko, Komodakis, “Learning to compare image patches via convolutional neural networks', IEEE CVPR 2015,pp.4353–4361.

60 APPLIED LIFE SCIENCES↗

Identification of Grand-design and Flocculent spirals from SDSS using deep convolutional neural network

ABSTRACT Spiral galaxies can be classified into the Grand-designs and Flocculents based on the nature of their spiral arms. The Grand-designs exhibit almost continuous and high contrast spiral arms and are believed to be driven by stationary density waves, while the Flocculents have patchy and low-contrast spiral features and are primarily stochastic in origin. We train a deep convolutional neural network model to classify spirals into Grand-designs and Flocculents, with a testing accuracy of $\mathrm{97.2{{\ \rm per\ cent}}}$. We then use the above model for classifying 1354 spirals from the SDSS. Out of these, 721 were identified as Flocculents, and the rest as Grand-designs. Interestingly, we find the mean asymptotic rotational velocities of our newly classified Grand-designs and Flocculents are 218 ± 86 and 146 ± 67 km s−1, respectively, indicating that the Grand-designs are mostly the high-mass and the Flocculents the intermediate-mass spirals. This is further corroborated by the observation that the mean morphological indices of the Grand-designs and Flocculents are 2.6 ± 1.8 and 4.7 ± 1.9, respectively, implying that the Flocculents primarily consist of a late-type galaxy population in contrast to the Grand-designs. Finally, an almost equal fraction of bars ∼0.3 in both the classes of spiral galaxies reveals that the presence of a bar component does not regulate the type of spiral arm hosted by a galaxy. Our results may have important implications for formation and evolution of spiral arms in galaxies.

Astronomy & Astrophysics↗

Inception Based Deep Convolutional Neural Network for Remaining Useful Life Estimation of Turbofan Engines

Accurate estimation of the remaining useful life (RUL) is a key component of condition based maintenance (CBM) and prognosis and health management (PHM). Data-based models for the estimation of RUL are of particular interest because expert knowledge of systems is not always available and physical modeling is often not feasible. In this paper, a deep convolutional neural network (CNN) architecture is investigated for its ability to estimate the RUL of turbofan engines. The input to the model is a window of time series data collected from the engine under test. Inputting raw sensor data allows features to be learned instead of manually determined. To incorporate the ability to detect features of differing lengths, inception modules are used in the neural network architecture. The model is trained and tested using the new Commercial Modular Aero-Propulsion System Simulation (N-CMAPSS) data set and high prognosis accuracy was achieved. The developed model was used in the 2021 PHM Society Data Challenge and received second place, further validating its ability to accurately estimate RUL.

DeVol, Nathaniel↗