Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “random initialization of 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 37 records · Page 2
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

TNet: A Model-Constrained Tikhonov Network Approach for Inverse Problems

Deep learning (DL), in particular deep neural networks, by default is purely data-driven and in general does not require physics. This is the strength of DL but also one of its key limitations when applied to science and engineering problems in which underlying physical properties—such as stability, conservation, and positivity—and accuracy are required. DL methods in their original forms are often not capable of respecting the underlying mathematical models or achieving desired accuracy even in big-data regimes. On the other hand, many data-driven science and engineering problems, such as inverse problems, typically have limited experimental or observational data, and DL would overfit the data in this case. Leveraging information encoded in the underlying mathematical models, we argue, not only compensates for missing information in low data regimes but also provides opportunities to equip DL methods with the underlying physics, hence promoting better generalization. This paper develops a model-constrained DL approach and its variant TNet—a Tikhonov neural network—which are capable of learning not only information hidden in the training data but also in the underlying mathematical models to solve inverse problems governed by partial differential equations in low data regimes. We provide the constructions and some theoretical results for the proposed approaches for both linear and nonlinear inverse problems. Since TNet is designed to learn inverse solutions with Tikhonov regularization, it is interpretable: in fact it recovers Tikhonov solutions for linear cases while potentially approximating Tikhonov solutions for nonlinear inverse problems. We also prove that data randomization can enhance not only the smoothness of the networks but also their generalizations. Comprehensive numerical results confirm the theoretical findings and show that with even as little as 1 training data sample for one-dimensional (1D) deconvolution, 5 for an inverse 2D heat conductivity problem, 100 for inverse initial conditions for a time-dependent 2D Burgers’s equation, and 50 for inverse initial conditions for 2D Navier–Stokes equations, TNet solutions can be as accurate as Tikhonov solutions while being several orders of magnitude faster. Furthermore, this is possible owing to the model-constrained term, replications, and randomization.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

Deep active learning for classifying cancer pathology reports

Abstract Background Automated text classification has many important applications in the clinical setting; however, obtaining labelled data for training machine learning and deep learning models is often difficult and expensive. Active learning techniques may mitigate this challenge by reducing the amount of labelled data required to effectively train a model. In this study, we analyze the effectiveness of 11 active learning algorithms on classifying subsite and histology from cancer pathology reports using a Convolutional Neural Network as the text classification model. Results We compare the performance of each active learning strategy using two differently sized datasets and two different classification tasks. Our results show that on all tasks and dataset sizes, all active learning strategies except diversity-sampling strategies outperformed random sampling, i.e., no active learning. On our large dataset (15K initial labelled samples, adding 15K additional labelled samples each iteration of active learning), there was no clear winner between the different active learning strategies. On our small dataset (1K initial labelled samples, adding 1K additional labelled samples each iteration of active learning), marginal and ratio uncertainty sampling performed better than all other active learning techniques. We found that compared to random sampling, active learning strongly helps performance on rare classes by focusing on underrepresented classes. Conclusions Active learning can save annotation cost by helping human annotators efficiently and intelligently select which samples to label. Our results show that a dataset constructed using effective active learning techniques requires less than half the amount of labelled data to achieve the same performance as a dataset constructed using random sampling.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2024

Zero-Power Analog Optical Processing

The motivation behind this research is the growing challenge of handling the massive amounts of data generated by modern imaging systems. Conventional digital image processing techniques are struggling to keep pace with the demands of high-resolution and high-speed imaging systems for remote sensing due to their high-power consumption and data storage requirements. We present a novel approach based on analog photonics to address this challenge. The proposed system utilizes a silicon-photonics-based image encoder positioned after image formation and initial optical-to-electrical conversion. The photonic encoder compresses image data using a passive disordered photonic structure to perform kernel-type random projections of the raw data. The compressed data is then processed by a back-end neural network, which reconstructs the original image with high fidelity (structural similarity exceeding 90%). Our proposed approach has the potential to compress images with ~ 1000X lower power consumption compared to digital approaches with data rates exceeding 1 terapixel/second.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

FY21 Progress Report: SRNL Analysis of ICCWR LCM and WAMS data for Corrosion and Cracking

The development of algorithms for machine learning and data analysis for the 3013 Surveillance Program is a collaborative effort by the Savannah River National Laboratory (SRNL) and the University of South Carolina (USC). For corrosion detection, Laser Confocal Microscope (LCM) or Wide Area 3D Measurement System (WAMS) data is extracted from large binary files, with software written to convert the data to physical attributes (e.g., height, color and grayscale values; all as functions of a location in a plane projection). A user-friendly Matlab Graphical User Interface (GUI) that reads data from either LCM or WAMS files was developed to integrate input data with software developed for processing and evaluation. The GUI can selectively download binary data, interrogate data attributes, label data, flag significant features, execute Machine Learning (ML) algorithms, output parameters for trained ML algorithms, report ML model accuracy with respect to labeled data, and generate graphical representations for various analyses. Features can be called out by user-specified thresholds, manual labeling or machine learning algorithms when they have been completed. The ability to rapidly label data is important because of the volume of data required for training machine learning algorithms. The GUI has the flexibility to allow addition of improved ML algorithms, methods for data visualization, and statistical computations. Statistical analyses via the GUI include areas of pits within a defined range of pit depths, correlations between Red-Green-Blue (RGB) or grayscale intensity and relative surface height, covariances between values associated with features, and feature histograms. The development of supervised machine learning algorithms, however, has been hindered by a lack of training data. The machine learning algorithms for crack identification are being refined but require improvements to the true positive rate for crack detection. This shortcoming is an artifact of the limited training data currently available, perhaps more so than the structure of the neural networks. At present, the best results are had from a consensus over an ensemble of randomly generated Deep Neural Network (DNN) or Convolutional Neural Network (CNN) algorithms. Although the consensus accuracy method has yielded optimum true positive and true negative rates in excess of 80%, additional validation testing is necessary. In addition to the suite of LCM data that was initially used, and which represents the majority of the work presented in this report, WAMS image data was also reviewed at a preliminary level. The review included a comparison between image resolution and dynamic range for each method. WAMS (ZON file) image data was found to have a pixel pitch of 3.69μm compared to 1 μm for the LCM (vk4 file) data, which implies a lower resolution for the WAMS images. Conversely, the ratio of dynamic range of the WAMS data to the LCM data was approximately 41:20 for height data, suggesting that information from WAMS should more accurately determine the depth of pits. At present, the significance of the greater dynamic range of the WAMS data relative to the LCM data has not yet been evaluated.

12 MANAGEMENT OF RADIOACTIVE AND NON-RADIOACTIVE W↗
DOE OSTI2024

Newton-Raphson AC Power Flow Convergence Based on Deep Learning Initialization and Homotopy Continuation

Power flow forms the basis of many power system studies. With the increased penetration of renewable energy, grid planners tend to perform multiple power flow simulations under various operating conditions and not just selected snapshots at peak or light load conditions. Getting a converged AC power flow (ACPF) case remains a significant challenge for grid planners especially in large power grid networks. This paper proposes a two-stage approach to improve Newton-Raphson ACPF convergence and was applied to a 6102 bus Electric Reliability Council of Texas (ERCOT) system. The first stage utilizes a deep learning-based initializer with data re-training. Here a deep neural network (DNN) initializer is developed to provide better initial voltage magnitude and angle guesses to aid in power flow convergence. This is because Newton-Raphson ACPF is quite sensitive to the initial conditions and bad initialization could lead to divergence. The DNN initializer includes a data re-training framework that improves the initializer's performance when faced with limited training data. The DNN initializer successfully solved 3,285 cases out of 3,899 non-converging dispatch and performed better than random forest and DC power flow initialization methods. ACPF cases not solved in this first stage are then passed through a hot-starting algorithm based on homotopy continuation with switched shunt control. The hot-starting algorithm successfully converged 416 cases out of the remaining 614 non-converging ACPF dispatch. In conclusion, the combined two-stage approach achieved a 94.9% success rate, by converging a total of 3,701 cases out of the initial 3,899 unsolved cases.

Deep learning↗
DOE OSTI2021

CyBERT: Cybersecurity Claim Classification by Fine-Tuning the BERT Language Model

We introduce CyBERT, a cybersecurity feature claims classifier based on bidirectional encoder representations from transformers and a key component in our semi-automated cybersecurity vetting for industrial control systems (ICS). To train CyBERT, we created a corpus of labeled sequences from ICS device documentation collected across a wide range of vendors and devices. This corpus provides the foundation for fine-tuning BERT’s language model, including a prediction-guided relabeling process. We propose an approach to obtain optimal hyperparameters, including the learning rate, the number of dense layers, and their configuration, to increase the accuracy of our classifier. Fine-tuning all hyperparameters of the resulting model led to an increase in classification accuracy from 76% obtained with BertForSequenceClassification’s original architecture to 94.4% obtained with CyBERT. Furthermore, we evaluated CyBERT for the impact of randomness in the initialization, training, and data-sampling phases. CyBERT demonstrated a standard deviation of ±0.6% during validation across 100 random seed values. Finally, we also compared the performance of CyBERT to other well-established language models including GPT2, ULMFiT, and ELMo, as well as neural network models such as CNN, LSTM, and BiLSTM. The results showed that CyBERT outperforms these models on the validation accuracy and the F1 score, validating CyBERT’s robustness and accuracy as a cybersecurity feature claims classifier.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Transferring predictions of formation energy across lattices of increasing size*

In this study, we show the transferability of graph convolutional neural network (GCNN) predictions of the formation energy of the nickel-platinum solid solution alloy across atomic structures of increasing sizes. The original dataset was generated with the large-scale atomic/molecular massively parallel simulator using the second nearest-neighbor modified embedded-atom method empirical interatomic potential. Geometry optimization was performed on the initially randomly generated face centered cubic crystal structures and the formation energy has been calculated at each step of the geometry optimization, with configurations spanning the whole compositional range. Using data from various steps of the geometry optimization, we first trained our open-source, scalable implementation of GCNN called HydraGNN on a lattice of 256 atoms, which accounts well for the short-range interactions. Using this data, we predicted the formation energy for lattices of 864 atoms and 2048 atoms, which resulted in lower-than-expected accuracy due to the long-range interactions present in these larger lattices. We accounted for the long-range interactions by including a small amount of training data representative for those two larger sizes, whereupon the predictions of HydraGNN scaled linearly with the size of the lattice. Therefore, our strategy ensured scalability while reducing significantly the computational cost of training on larger lattice sizes.

36 MATERIALS SCIENCE↗
NASA NTRS2004

Shortcomings with Tree-Structured Edge Encodings for Neural Networks

In evolutionary algorithms a common method for encoding neural networks is to use a tree structured assembly procedure for constructing them. Since node operators have difficulties in specifying edge weights and these operators are execution-order dependent, an alternative is to use edge operators. Here we identify three problems with edge operators: in the initialization phase most randomly created genotypes produce an incorrect number of inputs and outputs; variation operators can easily change the number of input/output (I/O) units; and units have a connectivity bias based on their order of creation. Instead of creating I/O nodes as part of the construction process we propose using parameterized operators to connect to preexisting I/O units. Results from experiments show that these parameterized operators greatly improve the probability of creating and maintaining networks with the correct number of I/O units, remove the connectivity bias with I/O units and produce better controllers for a goal-scoring task.

Hornby, Gregory S.↗
NASA NTRS2016

Psychophysiological Sensing and State Classification for Attention Management in Commercial Aviation

Attention-related human performance limiting states (AHPLS) can cause pilots to lose airplane state awareness (ASA), and their detection is important to improving commercial aviation safety. The Commercial Aviation Safety Team found that the majority of recent international commercial aviation accidents attributable to loss of control inflight involved flight crew loss of airplane state awareness, and that distraction of various forms was involved in all of them. Research on AHPLS, including channelized attention, diverted attention, startle / surprise, and confirmation bias, has been recommended in a Safety Enhancement (SE) entitled "Training for Attention Management." To accomplish the detection of such cognitive and psychophysiological states, a broad suite of sensors has been implemented to simultaneously measure their physiological markers during high fidelity flight simulation human subject studies. Pilot participants were asked to perform benchmark tasks and experimental flight scenarios designed to induce AHPLS. Pattern classification was employed to distinguish the AHPLS induced by the benchmark tasks. Unimodal classification using pre-processed electroencephalography (EEG) signals as input features to extreme gradient boosting, random forest and deep neural network multiclass classifiers was implemented. Multi-modal classification using galvanic skin response (GSR) in addition to the same EEG signals and using the same types of classifiers produced increased accuracy with respect to the unimodal case (90 percent vs. 86 percent), although only via the deep neural network classifier. These initial results are a first step toward the goal of demonstrating simultaneous real time classification of multiple states using multiple sensing modalities in high-fidelity flight simulators. This detection is intended to support and inform training methods under development to mitigate the loss of ASA and thus reduce accidents and incidents.

Harrivel, Angela R.↗
DOE OSTI2020

Trainability of ReLU networks and Data-dependent Initialization

Here, in this paper, we study the trainability of rectified linear unit (ReLU) networks at initialization. A ReLU neuron is said to be dead if it only outputs a constant for any input. Two death states of neurons are introduced–tentative and permanent death. A network is then said to be trainable if the number of permanently dead neurons is sufficiently small for a learning task. We refer to the probability of a randomly initialized network being trainable as trainability. We show that a network being trainable is a necessary condition for successful training, and the trainability serves as an upper bound of training success rates. In order to quantify the trainability, we study the probability distribution of the number of active neurons at initialization. In many applications, overspecified or overparameterized neural networks are successfully employed and shown to be trained effectively. With the notion of trainability, we show that overparameterization is both a necessary and a sufficient condition for achieving a zero training loss. Furthermore, we propose a data-dependent initialization method in an overparameterized setting. Numerical examples are provided to demonstrate the effectiveness of the method and our theoretical findings.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Absence of Barren Plateaus and Scaling of Gradients in the Energy Optimization of Isometric Tensor Network States

Abstract Vanishing gradients can pose substantial obstacles for high-dimensional optimization problems. Here we consider energy minimization problems for quantum many-body systems with extensive Hamiltonians and finite-range interactions, which can be studied on classical computers or in the form of variational quantum eigensolvers on quantum computers. Barren plateaus correspond to scenarios where the average amplitude of the energy gradient decreases exponentially with increasing system size. This occurs, for example, for quantum neural networks and for brickwall quantum circuits when the depth increases polynomially in the system size. Here we prove that the variational optimization problems for matrix product states, tree tensor networks, and the multiscale entanglement renormalization ansatz are free of barren plateaus. The derived scaling properties for the gradient variance provide an analytical guarantee for the trainability of randomly initialized tensor network states (TNS) and motivate certain initialization schemes. In a suitable representation, unitary tensors that parametrize the TNS are sampled according to the uniform Haar measure. We employ a Riemannian formulation of the gradient based optimizations which simplifies the analytical evaluation.

Barthel, Thomas↗
DOE OSTI2020

Survey of Image Denoising Methods for Medical Image Classification

Medical imaging devices, such as X-ray machines, inherently produce images that suffer from visual noise. Our objectives were to (i.) determine the effect of image denoising on a medical image classification task, and (ii.) determine if there exists a correlation between image denoising performance and medical image classification performance. We performed the medical image classification task on chest X-rays using the DenseNet-121 convolutional neural network (CNN) and used the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) metrics as the image denoising performance measures. We first found that different denoising methods can make a statistically significant difference in classification performance for select labels. We also found that denoising methods affect fine-tuned models more than randomly-initialized models and that fine-tuned models have significantly higher and more uniform performance than randomly-initialized models. Lastly, we found that there is no significant correlation between PSNR and SSIM values and classification performance for our task.

Michael, Peter↗
DOE OSTI2022

Benchmarking structural evolution methods for training of machine learned interatomic potentials

When creating training data for machine-learned interatomic potentials (MLIPs), it is common to create initial structures and evolve them using molecular dynamics (MD) to sample a larger configuration space. Here, we benchmark two other modalities of evolving structures, contour exploration (CE) and dimer-method (DM) searches against MD for their ability to produce diverse and robust density functional theory training data sets for MLIPs. We also discuss the generation of initial structures which are either from known structures or from random structures in detail to further formalize the structure-sourcing processes in the future. The polymorph-rich zirconium-oxygen composition space is used as a rigorous benchmark system for comparing the performance of MLIPs trained on structures generated from these structural evolution methods. Using Behler–Parrinello neural networks as our MLIP models, we find that CE and the DM searches are generally superior to MD in terms of spatial descriptor diversity and statistical accuracy.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗
DOE OSTI2023

1-D coupled surface flow and transport equations revisited via the physics-informed neural network approach

The de Saint-Venant equation (SVE) and advection–diffusion equation (ADE) are commonly employed to solve solute transport problems in surface water. In this work, we propose a mesh-free method based on the physics-informed neural network (PINN) to solve the one dimensional (1-D) SVE, ADE, and the coupled SVE and ADE (SVE-ADE) under various initial and boundary conditions. The PINN model extends the architecture of deep neural networks (DNNs) with implementation of loss function, which are additionally subject to constraints imposed by the physical laws of SVE and ADE, along with their initial and boundary conditions. In such a manner, PINNs can be quickly steered to the true solution while obeying the physical laws. The results of PINN model are compared with the analytical and/or numerical solutions under various conditions to investigate its accuracy and efficiency in solving the SVE, ADE, and SVE-ADE. Our results indicate PINN can accurately simulate the shock wave morphology and avoid numerical dissipation in unsteady flow condition. The PINN method outweighed traditional numerical methods in several aspects, including its ability to function with small amounts of data, no grid discretization, and random selection of sampling points, etc. Additionally, the PINN method is also suitable for solving inverse problems with sparse and noisy data. With 1% noise and 2000 initial and boundary condition points (N u ), the errors of the estimated flow rate (v) and diffusion coefficient (D) are 0.003% and 0.105%, respectively, which indicate the accuracy and robustness of the proposed method. Finally, our results indicate the capability and robustness of the proposed PINN methodology for solving multi-physics problems, irrespective of the presence of sparse and noisy data in the training dataset.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2026

Online LIBS–ML Framework for Dynamic Characterization of Heterogeneous Waste-Derived Gasification Feedstocks

LIBS−ML framework for real time feedstock characterization during continuous conveyor transport Heterogeneous waste derived feedstocks (e.g., waste coal, biomass and blends) introduce rapid variability in heating value and ash chemistry that affect gasifier operation, yet conventional laboratory characterization techniques are too slow to support proactive control. To address this gap, this study reports on an online, in situ, dynamic characterization framework that couple’s laser-induced breakdown spectroscopy (LIBS) with leakage safe machine learning (ML) regression to deliver real time, decision quality predictions of gasifier relevant properties. A controlled sample matrix spanning two different waste coals, two different biomasses, and engineered blends under two particle size conditions were constructed and benchmarked using standardized laboratory analyses for proximate/ultimate properties and ash composition. LIBS spectra were acquired dynamically as material flowed on a conveyor belt, using high energy 1064 nm laser ablation and shot averaging to improve repeatability and precision. Supervised regression models (multi layer perceptron (MLP) /artificial neural network (ANN), random forest (RF), and support vector regression (SVR)) and an optimized weighted ensemble were trained on emission line feature sets using nested cross validation with Bayesian hyperparameter tuning and validated against an independent hold out set. The proposed LIBS−ML workflow achieves near laboratory predictive fidelity across parametric targets (including higher heating value (HHV), ash content, fixed carbon, sulfur, major ash forming oxides, and initial deformation temperature (IDT)), with the weighted ensemble providing a robust default predictor under dynamic measurement conditions. These results demonstrate a practical pathway for real time feedstock characterization that can enable feedforward adjustments and more resilient gasifier operation for variable quality waste derived fuels.

Biomass↗
DOE OSTI2026

Structure-aware Initialization via Numerical Continuation and Informed Priors

Scientific machine learning (SciML) often operates in ill-conditioned, weakly identifiable regimes due to limited data or indirect observations. In such settings, optimization and inference are highly sensitive to the starting point, making initialization--often under-reported--a consequential degree of freedom. Random initialization is not a neutral default as it induces an implicit prior over candidate solutions and can systematically bias the result, producing large run-to-run variability. Here, we formalize this view by treating initialization as a hidden confounder in SciML and develop a unifying theory for structure-aware initialization via numerical continuation, constructing warm starts from related problem instances. Across representative tasks, including physics-informed neural networks, maximum likelihood estimation, and variational inference, warm starts have been shown to consistently reduce optimization effort and improve reliability.

Data integrity↗
DOE OSTI2020

Learning in Modal Space: Solving Time-Dependent Stochastic PDEs Using Physics-Informed Neural Networks

One of the open problems in scientific computing is the long-time integration of nonlinear stochastic partial differential equations (SPDEs), especially with arbitrary initial data. We address this problem by taking advantage of recent advances in scientific machine learning and the spectral dynamically orthogonal (DO) and borthogonal (BO) methods for representing stochastic processes. The recently introduced DO/BO methods reduce the SPDE to solving a system of deterministic PDEs and a system of stochastic ordinary differential equations. Specifically, we propose two new physics-informed neural networks (PINNs) for solving time-dependent SPDEs, namely the neural network (NN)-DO/BO methods. The proposed methods incorporate the DO/BO constraints into the loss function (along with the modal decomposition of the SPDE) with an implicit form instead of generating explicit expressions for the temporal derivatives of the DO/BO modes. Hence, the NN-DO/BO methods can overcome some of the drawbacks of the original DO/BO methods. For example, we do not need the assumption that the covariance matrix of the random coefficients is invertible as in the original DO method, and we can remove the assumption of no eigenvalue crossing as in the original BO method. Moreover, the NN-DO/BO methods can be used to solve time-dependent stochastic inverse problems with the same formulation and same computational complexity as for forward problems. Furthermore, we demonstrate the capability of the proposed methods via several numerical examples, namely: (1) A linear stochastic advection equation with deterministic initial condition: we obtain good results with the proposed methods, while the original DO/BO methods cannot be applied directly in this case. (2) Long-time integration of the stochastic Burgers' equation: we show the good performance of NN-DO/BO methods, especially the effectiveness of the NN-BO approach for such problems with many eigenvalue crossings during the whole time evolution, while the original BO method fails. (3) Nonlinear reaction diffusion equation: we consider both the forward problem and the inverse problems, including very noisy initial point values, to investigate the flexibility of the NN-DO/BO methods in handling inverse and mixed type problems. Taken together, these simulation results demonstrate that the NN-DO/BO methods can be employed to effectively quantify uncertainty propagation in a wide range of physical problems, but future work should address the efficiency issue of PINNs for forward problems.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Deep material network via a quilting strategy: visualization for explainability and recursive training for improved accuracy

Recent developments integrating micromechanics and neural networks offer promising paths for rapid predictions of the response of heterogeneous materials with similar accuracy as direct numerical simulations. The deep material network is one such approaches, featuring a multi-layer network and micromechanics building blocks trained on anisotropic linear elastic properties. Once trained, the network acts as a reduced-order model, which can extrapolate the material’s behavior to more general constitutive laws, including nonlinear behaviors, without the need to be retrained. However, current training methods initialize network parameters randomly, incurring inevitable training and calibration errors. Here, we introduce a way to visualize the network parameters as an analogous unit cell and use this visualization to “quilt” patches of shallower networks to initialize deeper networks for a recursive training strategy. The result is an improvement in the accuracy and calibration performance of the network and an intuitive visual representation of the network for better explainability.

97 MATHEMATICS AND COMPUTING↗