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 equivariant graph neural network model enables the rapid and accurate prediction of complete fourth-rank elasticity tensors of inorganic materials, facilitating the discovery of materials with exceptional mechanical properties.
This work demonstrates the applicability of partially Bayesian neural networks (PBNNs) to active and transfer learning within materials science and chemistry domains, combining reliable uncertainty quantification (UQ) with NN flexibility.
Here, we apply deep neural networks (DNN) to data from the EXO-200 experiment. In the studied cases, the DNN is able to reconstruct the relevant parameters—total energy and position—directly from raw digitized waveforms, with minimal exceptions. For the first time, the developed algorithms are evaluated on real detector calibration data. The accuracy of reconstruction either reaches or exceeds what was achieved by the conventional approaches developed by EXO-200 over the course of the experiment. Most existing DNN approaches to event reconstruction and classification in particle physics are trained on Monte Carlo simulated events. Such algorithms are inherently limited by the accuracy of the simulation. We describe a unique approach that, in an experiment such as EXO-200, allows to successfully perform certain reconstruction and analysis tasks by training the network on waveforms from experimental data, either reducing or eliminating the reliance on the Monte Carlo.
A tau lepton identification algorithm,DeepTau, based on convolutional neural network techniques, has been developed in the CMS experiment to discriminate reconstructed hadronic decays of tau leptons (τ h ) from quark or gluon jets and electrons and muons that are misreconstructed as τ h candidates. The latest version of this algorithm, v2.5, includes domain adaptation by backpropagation, a technique that reduces discrepancies between collision data and simulation in the region with the highest purity of genuine τh candidates. Additionally, a refined training workflow improves classification performance with respect to the previous version of the algorithm, with a reduction of 30–50% in the probability for quark and gluon jets to be misidentified as τ h candidates for given reconstruction and identification efficiencies. This paper presents the novel improvements introduced in theDeepTau algorithm and evaluates its performance in LHC proton-proton collision data at √(s) = 13 and 13.6 TeV collected in 2018 and 2022 with integrated luminosities of 60 and 35 fb -1 , respectively. Techniques to calibrate the performance of the τ h identification algorithm in simulation with respect to its measured performance in real data are presented, together with a subset of results among those measured for use in CMS physics analyses.
We introduce a variational Monte Carlo framework that combines neural-network quantum states with the Lorentz integral transform technique to compute the dynamical properties of self-bound quantum many-body systems in continuous Hilbert spaces. While broadly applicable to various quantum systems, including atoms and molecules, in this initial application we focus on the photoabsorption cross section of light nuclei, where benchmarks against numerically exact techniques are available. Our accurate theoretical predictions are complemented by robust uncertainty quantification, enabling meaningful comparisons with experiments. Here, we demonstrate that a relatively simple nuclear Hamiltonian—based on a leading-order pionless EFT expansion and known to accurately reproduce ground-state energies of nuclei with 𝐴 ≤ 40—also provides a reliable description of the photoabsorption cross section.
Convolutional Neural Networks (CNN) have recently demonstrated effective performance in many Natural Language Processing tasks. In this study, we explore a novel approach for pruning a CNN's convolution filters using our new data-driven utility score. We have applied this technique to an information extraction task of classifying a dataset of cancer pathology reports by cancer type, a highly imbalanced dataset. Compared to standard CNN training, our new algorithm resulted in a nearly .07 increase in the micro-averaged F1-score and a strong .22 increase in the macro-averaged F1-score using a model with nearly a third fewer network weights. We show how directly utilizing a network's interpretation of data can result in strong performance gains, particularly with severely imbalanced datasets.
The challenge of choosing an appropriate convolutional neural network (CNN) architecture for specific applications and different data sets is still poorly understood in the literature. This is problematic, since CNNs have shown strong promise for analyzing scientific data from many domains including particle imaging detectors. In this paper, we proposed a systematic language that is useful for comparison between different CNN's architectures before training time. This helps us predict whether a network can perform better than a certain threshold accuracy before training up to 70% accuracy using simple machine learning models. Additionally, we found a coefficient of determination of 0.966 for an Ordinary Least Squares model in a regression task to predict accuracy of a large population of networks.
The DeepDataProfiler is a methodology and framework for providing interpretability to trained neural networks. Its approach is to decompose a network into a weighted graph of neurons and synapses and link the components of the graph to human identifiable concepts. By identifying concepts important to the network and tracking the decision process employed by the network, the network becomes more transparent and less like a black box. Spurious decisions and poor generalization strategies can be identified and a measure of trustworthiness can be established.
We introduce a variational Monte Carlo framework that combines neural-network quantum states with the Lorentz integral transform technique to compute the dynamical properties of self-bound quantum many-body systems in continuous Hilbert spaces. While broadly applicable to various quantum systems, including atoms and molecules, in this initial application we focus on the photoabsorption cross section of light nuclei, where benchmarks against numerically exact techniques are available. Our accurate theoretical predictions are complemented by robust uncertainty quantification, enabling meaningful comparisons with experiments. We demonstrate that a simple nuclear Hamiltonian, based on a leading-order pionless effective field theory expansion and known to accurately reproduce the ground-state energies of nuclei with $A\leq 20$ nucleons also provides a reliable description of the photoabsorption cross section.
Patterns over (-1,0,1) define, by their outer products, partially connected neural networks, consisting of internally strongly connected, externally weakly connected subnetworks. The connectivity patterns may have highly organized structures, such as lattices and fractal trees or nests. Subpatterns over (-1,1) define the subcodes stored in the subnetwork, that agree in their common bits. It is first shown that the code words are locally stable stares of the network, provided that each of the subcodes consists of mutually orthogonal words or of, at most, two words. Then it is shown that if each of the subcodes consists of two orthogonal words, the code words are the unique ground states (absolute minima) of the Hamiltonian associated with the network. The regions of attraction associated with the code words are shown to grow with the number of subnetworks sharing each of the neurons. Depending on the particular network architecture, the code sizes of partially connected networks can be vastly greater than those of fully connected ones and their error correction capabilities can be significantly greater than those of the disconnected subnetworks. The codes associated with lattice-structured and hierarchical networks are discussed in some detail.
This note describes an attempt to use standard neural network tools to fashion a means of detecting eddy patterns in the wall region of a turbulent flow. The research was motivated by the desire to formulate a means to use only flow parameters that can be sensed on the wall to describe the passing eddy structure. If a simple formulation can be obtained, it could conceivably be utilized to control actuators embedded in the wall. Such actuators have been developed by Jacobson and Reynolds (1993a), Blackwelder and Liu (1994), Tung et al. (1995), and others. These actuators have the common characteristics that they are small and are typically flush with the wall when not deployed. When they are activated, it is assumed that they will be able to interact constructively with the turbulent eddies near their location to either decrease the wall shear stress, enhance or reduce the mixing, etc. At present, there is only a nascent understanding of the interaction dynamics between the actuators and the eddies in the flow. Nevertheless, for such interaction to succeed, methods to couple the actuators to the oncoming flow must be obtained. General methods must be found that will detect the space and temporal location of the desired structure. In particular, it will be necessary to know when the eddies will arrive at the location of the actuator. This research attempted to use the shear stress measurements on the wall in the vicinity of an actuator location to predict when a particular eddy pattern would arrive and/or occur at the designated location. In this work the eddy pattern to be detected was identified by its velocity signature only.
The analysis of vision-based deep neural networks (DNNs) is highly desirable but it is very challenging due to the difficulty of expressing formal specifications for vision tasks and the lack of efficient verification procedures. In this paper, we propose to leverage emerging multimodal, vision-language, foundation models (VLMs) as a lens through which we can reason about vision models. VLMs have been trained on a large body of images accompanied by their textual description, and are thus implicitly aware of high-level, human-understandable concepts describing the images. We describe a logical specification language Con spec designed to facilitate writing specifications in terms of these concepts. To define and formally check Con spec specifications, we build a map between the internal representations of a given vision model and a VLM, leading to an efficient verification procedure of natural-language properties for vision models. We demonstrate our techniques on a ResNet-based classifier trained on the RIVAL-10 dataset using CLIP as the multimodal model.
In this paper, leveraging the capabilities of neural networks for modeling the non-linearities that exist in the data, we propose several models that can project data into a low dimensional, discriminative, and smooth manifold. The proposed models can transfer knowledge from the domain of known classes to a new domain where the classes are unknown. A clustering algorithm is further applied in the new domain to find potentially new classes from the pool of unlabeled data. The research problem and data for this paper originated from the Gravity Spy project which is a side project of Advanced Laser Interferometer Gravitational-wave Observatory (LIGO). The LIGO project aims at detecting cosmic gravitational waves using huge detectors. However non-cosmic, non-Gaussian disturbances known as "glitches", show up in gravitational-wave data of LIGO. This is undesirable as it creates problems for the gravitational wave detection process. Gravity Spy aids in glitch identification with the purpose of understanding their origin. Since new types of glitches appear over time, one of the objective of Gravity Spy is to create new glitch classes. Towards this task, we offer a methodology in this paper to accomplish this.
An analogy is drawn among the irreversible evolution of a neural-network-based A.I., an information field associated to spacetime configurations and the behaviour of entities described by the Schrödinger equation.
Accurate chemical kinetics modeling is crucial for improving the efficiency of chemical processing and synthesis of ceramic matrix composites. Detailed kinetic models are computationally expensive due to the large number of transported chemical species, while the simplified physics-based models, such as single-step global mechanisms, are efficient but often overlook key chemical intermediates and pathways. Recent deep learning approaches promise accurate and cost-effective models. Yet, they require additional closures for the transported nonlinear latent variables, complicating integration with existing solvers. In this work, we develop a hybrid linear—nonlinear reduced model for silicon carbide deposition from methyltrichlorosilane precursor by combining principal component analysis (PCA) and autoencoder (AE) neural network (NN) approaches. PCA is used to identify a smaller set of linear transport variables, enabling direct reuse of conventional transport solvers. NNs then reconstruct the full chemical state from these reduced variables. We demonstrate the method on a chemical vapor deposition reactor—comprising a gas-phase pyrolysis plug flow reactor and a heterogeneous surface reactor—over a wide range of temperatures, pressures, and residence times. Our PCA–AE model achieves high accuracy with only five transported scalars, achieving an eightfold cost reduction compared to detailed mechanisms, in both a priori (using data from the test set only) and a posteriori (coupled with a differential equation solver). In conclusion, notable errors arise primarily near training domain boundaries and for long residence times, indicating the need for domain shift indicators and better long-horizon predictions in future reduced chemistry model development.
The classical Back-Propagation (BP) scheme with gradient-based optimization in training Artificial Neural Networks (ANNs) suffers from many drawbacks, such as the premature convergence, and the tendency of being trapped in local optimums. Therefore, as an alternative for the BP and gradient-based optimization schemes, various Evolutionary Algorithms (EAs), i.e., Particle Swarm Optimization (PSO), Genetic Algorithm (GA), Simulated Annealing (SA), and Differential Evolution (DE), have gained popularity in the field of ANN weight training. This study applied a new efficient and effective Shuffled Complex Evolutionary Global Optimization Algorithm with Principal Component Analysis – University of California Irvine (SP-UCI) to the weight training process of a three-layer feed-forward ANN. A large-scale numerical comparison is conducted among the SP-UCI-, PSO-, GA-, SA-, and DE-based ANNs on 17 benchmark, complex, and real-world datasets. Here the results show that SP-UCI-based ANN outperforms other EA-based ANNs in the context of convergence and generalization. Results suggest that the SP-UCI algorithm possesses good potential in support of the weight training of ANN in real-word problems. In addition, the suitability of different kinds of EAs on training ANN is discussed. The large-scale comparison experiments conducted in this paper are fundamental references for selecting proper ANN weight training algorithms in practice.
Machine learning methods and uncertainty quantification have been gaining interest throughout the last several years in low-energy nuclear physics. In particular, Gaussian processes and Bayesian neural networks have increasingly been applied to improve mass model predictions while providing well-quantified uncertainties. In this work, we use the probabilistic Mixture Density Network (MDN) to directly predict the mass excess of the 2016 Atomic Mass Evaluation within the range of measured data, and we extrapolate the inferred models beyond available experimental data. The MDN provides not only mean values but also full posterior distributions both within the training set and extrapolated testing set. We show that the addition of physical information to the feature space increases the accuracy of the match to the training data as well as provides for more physically meaningful extrapolations beyond the the limits of experimental data.
A machine learning model was created to predict the electron spectrum generated by a GeV-class laser wakefield accelerator. The model was constructed from variational convolutional neural networks, which mapped the results of secondary laser and plasma diagnostics to the generated electron spectrum. An ensemble of trained networks was used to predict the electron spectrum and to provide an estimation of the uncertainty of that prediction. It is anticipated that this approach will be useful for inferring the electron spectrum prior to undergoing any process that can alter or destroy the beam. In addition, the model provides insight into the scaling of electron beam properties due to stochastic fluctuations in the laser energy and plasma electron density.