Search NASA⌕ Search

SEARCH · Search NASA

Results for “Distributed training”

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 253 records · Page 14

Intelligent-based Structural Damage Detection Model

This paper presents the application of a novel Artificial Neural Network (ANN) model for the diagnosis of structural damage. The ANN model, denoted as the GRNNFA, is a hybrid model combining the General Regression Neural Network Model (GRNN) and the Fuzzy ART (FA) model. It not only retains the important features of the GRNN and FA models (i.e. fast and stable network training and incremental growth of network structure) but also facilitates the removal of the noise embedded in the training samples. Structural damage alters the stiffness distribution of the structure and so as to change the natural frequencies and mode shapes of the system. The measured modal parameter changes due to a particular damage are treated as patterns for that damage. The proposed GRNNFA model was trained to learn those patterns in order to detect the possible damage location of the structure. Simulated data is employed to verify and illustrate the procedures of the proposed ANN-based damage diagnosis methodology. The results of this study have demonstrated the feasibility of applying the GRNNFA model to structural damage diagnosis even when the training samples were noise contaminated.

97 MATHEMATICS AND COMPUTING↗

Generative models on phase space

Deep generative models such as diffusion and flow matching are powerful machine learning tools capable of learning and sampling from high-dimensional distributions. They are particularly useful when the training data appears to be concentrated on a submanifold of the data embedding space. For high-energy physics data, consisting of collections of relativistic energy-momentum 4-vectors, this submanifold can enforce extremely strong physically-motivated priors, such as energy and momentum conservation. If these constraints are learned only approximately, rather than exactly, this can inhibit the interpretability and reliability of such generative models. To remedy this deficiency, we introduce generative models which are, by construction, confined at every step of their sampling trajectory to the manifold of massless N-particle Lorentz-invariant phase space in the center-of-momentum frame. In the case of diffusion models, the "pure noise" forward process endpoint corresponds to the uniform distribution on phase space, which provides a clear starting point from which to identify how correlations among the particles emerge during the reverse (de-noising) process. We demonstrate that our models are able to learn both few-particle and many-particle distributions with various singularity structures, paving the way for future interpretability studies using generative models trained on simulated jet data.

Bogorad, Zachary [Fermilab]↗

A multi-dimensional parametric study of variability in multi-phase flow dynamics during geologic CO 2 sequestration accelerated with machine learning

Successful geologic CO 2 storage projects depend on numerical simulations to predict reservoir performance during site selection, injection verification, and post-injection monitoring phases of the project. These numerical simulations solve non-linear sets of coupled partial differential equations, while accounting for multi-phase fluid dynamics on the basis of constitutive equations that are embedded into the solution scheme. As a consequence, individual simulations often require tens to hundreds of hours to complete on high-performance computing clusters. Moreover, laboratory experiments reveal that parametric functions for capillary pressure and relative permeability exhibit substantial variability, even within the same rock type. This combination of computational expense and wide-ranging parametric variability means that there remains substantial uncertainty in the behavior of multi-phase CO 2 -water systems, particularly in the context of feedbacks between relative permeability and capillary pressure. To bridge this knowledge gap, here we develop a novel workflow that utilizes physics-based numerical simulation to train an artificial neural network (ANN) emulator for interrogating the multivariate parameter space that governs both capillary pressure and relative permeability. With this approach, the ANN is trained to emulate both fluid pressure distribution and CO 2 saturation, which are then interrogated quantitatively to generate parametric response surface mappings with high-fidelity resolution. Results from this study initially show that capillary entry pressure is the dominant control on both CO 2 plume geometry and fluid pressure propagation when considering the combined effects of capillary pressure and relative permeability, particularly when phase interference is low and residual CO 2 saturation is high. Moreover, the ANN emulator provides tremendous computational speed-up by computing 2691 individual simulations in several minutes; whereas, the same simulation ensemble would have required ~3 years of simulation time using only physics-based simulation methods (25,000 times speed up).

58 GEOSCIENCES↗

Physics-based hybrid machine learning for critical heat flux prediction with uncertainty quantification

Critical heat flux (CHF) is a key quantity in nuclear system modeling due to its impact on heat transfer, safety margins, and reactor performance. This study develops and validates an uncertainty-aware hybrid modeling approach that combines machine learning with physics-based models to predict CHF in cases of dryout. The Biasi and Bowring empirical correlations were paired with three ML uncertainty quantification (UQ) techniques: deep neural network (DNN) ensembles, Bayesian neural networks (BNNs), and deep Gaussian processes (DGPs). A pure ML model without a base model was evaluated for comparison. Model performance was assessed under plentiful (7,350 points) and limited (9 points) training data scenarios using parity, uncertainty distributions, and calibration curves. Results show that the Biasi hybrid DNN ensemble achieved the best overall performance, with a mean absolute relative error of 1.846%, and well-calibrated uncertainty estimates. The BNN-based hybrids showed slightly higher error (2.14%) but superior uncertainty calibration. DGP models underperformed, with over 6% error and poor uncertainty calibration. All hybrid models outperformed pure machine learning configurations, demonstrating resistance against data scarcity. These findings indicate that hybrid modeling significantly improves predictive accuracy, interpretability, and resilience to data scarcity. The integration of uncertainty awareness provides actionable confidence in CHF predictions, which is vital for safety-critical decisions in nuclear applications. This hybrid approach offers a viable pathway for deploying ML models in reactor analysis tools while preserving domain knowledge and physical consistency.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗

Generative learning of densities on manifolds

A generative modeling framework is proposed that combines diffusion models and manifold learning to efficiently sample data densities on manifolds. The approach utilizes Diffusion Maps to uncover possible low-dimensional underlying (latent) spaces in the high-dimensional data (ambient) space. Two approaches for sampling from the latent data density are described. The first is a score-based diffusion model, which is trained to map a standard normal distribution to the latent data distribution using a neural network. The second one involves solving an Itô stochastic differential equation in the latent space. Additional realizations of the data are generated by lifting the samples back to the ambient space using Double Diffusion Maps , a recently introduced technique typically employed in studying dynamical system reduction; here the focus lies in sampling densities rather than system dynamics. The proposed approaches enable sampling high dimensional data densities restricted to low-dimensional, a priori unknown manifolds. The efficacy of the proposed framework is demonstrated through a benchmark problem and a material with multiscale structure.

Double diffusion maps↗

A DETAILED STUDY OF PHOTOMETRIC REDSHIFTS FOR GOODS-SOUTH GALAXIES

We use the deepest and the most comprehensive photometric data currently available for GOODS-South (GOODS-S) galaxies to measure their photometric redshifts. The photometry includes VLT/VIMOS (U band), HST/ACS (F435W, F606W, F775W, and F850LP bands), VLT/ISAAC (J, H, and K{sub s} bands), and four Spitzer/IRAC channels (3.6, 4.5, 5.8, and 8.0 {mu}m). The catalog is selected in the z band (F850LP) and photometry in each band is carried out using the recently completed TFIT algorithm, which performs point-spread function (PSF) matched photometry uniformly across different instruments and filters, despite large variations in PSFs and pixel scales. Photometric redshifts are derived using the GOODZ code, which is based on the template fitting method using priors. The code also implements 'training' of the template spectral energy distribution (SED) set, using available spectroscopic redshifts in order to minimize systematic differences between the templates and the SEDs of the observed galaxies. Our final catalog covers an area of 153 arcmin{sup 2} and includes photometric redshifts for a total of 32,505 objects. The scatter between our estimated photometric and spectroscopic redshifts is {sigma} = 0.040 with 3.7% outliers to the full z-band depth of our catalog, decreasing to {sigma} = 0.039 and 2.1% outliers at a magnitude limit m{sub z} < 24.5. This is consistent with the best results previously published for GOODS-S galaxies, however, the present catalog is the deepest yet available and provides photometric redshifts for significantly more objects to deeper flux limits and higher redshifts than earlier works. Furthermore, we show that the photometric redshifts estimated here for galaxies selected as dropouts are consistent with those expected based on the Lyman break technique.

79 ASTRONOMY AND ASTROPHYSICS↗

MatPhase: Material phase prediction for Li-ion Battery Reconstruction using Hierarchical Curriculum Learning

Li-ion Batteries (LIB), one of the most efficient energy storage devices, are used extensively in many industrial applications. These batteries consist of electrodes that are put together with heterogeneous material compositions. Imaging data of these battery electrodes obtained from X-ray tomography can explain the distribution of material constituents and allow reconstructions to study electron transport pathways. Such reconstructions of material constituents help quantify various associated properties of electrodes (e.g., volume-specific surface area, porosity) which determine the performance of batteries. These images often suffer from low image contrast between multiple material constituents, hence making it difficult for humans to distinguish and characterize these constituents through visual inspection. A minor error in detecting distributions of the material constituents can lead to magnified errors in the calculated parameters of material properties (e.g., porosity). We present MatPhase, a novel hierarchical curriculum learning technique to address the complex task of estimating material constituent distribution in battery electrodes. MatPhase comprises three modules: (i) an uncertainty-aware global model trained to yield inferences conditioned upon global knowledge of material distribution, (ii) a local model to capture relatively more fine-grained (local) distributional signals, (iii) an aggregator model to appropriately fuse the local and global effects towards obtaining the final distribution. On average, MatPhase improves prediction up to 8.5% relative to other sophisticated modeling pipelines and state-of-the-arts (SOTA) object detection models employed in the performance comparison.

Tabassum, Anika↗

3D Gaussian Splatting for Volume Compression

This codebase uses machine learning to train a collection of 3D Gaussian distributions to approximate scientific volume data. Because this collection uses less memory than the original dataset, it can be used as a compressed model of the original data for applications such as visualization.

Dyken, Landon↗

How to understand limitations of generative networks

Well-trained classifiers and their complete weight distributions provide us with a well-motivated and practicable method to test generative networks in particle physics. We illustrate their benefits for distribution-shifted jets, calorimeter showers, and reconstruction-level events. In all cases, the classifier weights make for a powerful test of the generative network, identify potential problems in the density estimation, relate them to the underlying physics, and tie in with a comprehensive precision and uncertainty treatment for generative networks.

Physics↗

Integrated Design of Ultradurable, Low CO 2 Alternative Binder Systems via Machine Learning

This ARPA-E project developed a machine learning tool to use in formulation design of cementitious binders for concrete having 50% less embodied CO 2 and possessing twice the durability compared to concrete based on ordinary portland cement (OPC) binders. The technical focus was on limestone/calcined clay cement (LC3), the leading replacement for OPC. Here, hierarchical machine learning (HML) was used to model the flowability, set time, strength, and durability of LC3 concrete. This methodology identifies latent variables derived from domain knowledge and empirical models that develop an accurate model for a response surface from small datasets. For the flowability metric, particle packing was a dominant factor, while strength and durability were both strongly determined by the fraction of metakaolin and the water:solids ratio. Under constraints of water:binder ratio, material performance metrics, embodied CO 2 , and cost per tonne of OPC, multi-objective optimization was used to design binders parameterized by the mineral composition replacing OPC, particle size distributions, and water:solids ratio. The trained algorithm was able to predict multiple mixes met these performance criteria, and experimental testing validated the predictions. The model demonstrated here is relevant for North America, where pure kaolin deposits are found broadly. The approach is being taken forward into commercial application by Ansatz AI, a materials informatics company founded by PI Washburn and co-PI Poczos. Through collaborations with the cement and concrete industry, and funding from SBIR programs, a commercial software will be developed in future research.

36 MATERIALS SCIENCE↗

Four Design Demonstration Elments (DDEs) to be irradiated in the Advanced Test Reactor (ATR) and the Belgian Reactor 2 (BR2)

The U.S. High Performance Research Reactor Conversion fuel qualification team is focused on developing and qualifying plate type uranium-molybdenum (U-Mo) monolithic and uranium silicide (U3Si4) fuel. Several irradiations of mini-plates and full-size plate irradiations will demonstrate the fuel’s behavior. Along with qualifying fuel, larger tests are planned to support licensing fuel elements. These tests are referred to as DDEs. As a result of a scoping study performed in 2018, a decision was made to irradiate two DDEs in ATR at the Idaho National Laboratory (INL) and two other DDEs in BR2 in Belgium. The objective is to demonstrate successful irradiation of each DDE to support the licensing submittal of that specific reactor. Neutronics analysis is being performed to match the power density distribution of the represented elements. Test train hardware design are being done to ensure all the target conditions are being met for the two DDEs to be irradiated.

11 NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗

Do graph neural networks learn traditional jet substructure?

At the CERN LHC, the task of jet tagging, whose goal is to infer the origin of a jet given a set of final-state particles, is dominated by machine learning methods. Graph neural networks have been used to address this task by treating jets as point clouds with underlying, learnable, edge connections between the particles inside. We explore the decision-making process for one such state-of-the-art network, ParticleNet, by looking for relevant edge connections identified using the layerwise-relevance propagation technique. As the model is trained, we observe changes in the distribution of relevant edges connecting different intermediate clusters of particles, known as subjets. The resulting distribution of subjet connections is different for signal jets originating from top quarks, whose subjets typically correspond to its three decay products, and background jets originating from lighter quarks and gluons. This behavior indicates that the model is using traditional jet substructure observables, such as the number of prongs -- energetic particle clusters -- within a jet, when identifying jets.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

Active operator learning with predictive uncertainty quantification for partial differential equations

With the increased prevalence of neural operators being used to provide rapid solutions to partial differential equations (PDEs), understanding the accuracy of model predictions and the associated error levels is necessary for deploying reliable surrogate models in scientific applications. Existing uncertainty quantification (UQ) frameworks employ ensembles or Bayesian methods, which can incur substantial computational costs during both training and inference. Here, we propose a lightweight predictive UQ method tailored for Deep operator networks (DeepONets) that also generalizes to other operator networks. Numerical experiments on linear and nonlinear PDEs demonstrate that the framework’s uncertainty estimates are unbiased and provide accurate out-of-distribution uncertainty predictions with a sufficiently large training dataset. Our framework provides fast inference and uncertainty estimates that can efficiently drive outer-loop analyses that would be prohibitively expensive with conventional solvers. We demonstrate how predictive uncertainties can be used in the context of Bayesian optimization and active learning problems to yield improvements in accuracy and data-efficiency for outer-loop optimization procedures. In the active learning setup, we extend the framework to Fourier Neural Operators (FNO) and describe a generalized method for other operator networks. To enable real-time deployment, we introduce an inference strategy based on precomputed trunk outputs and a sparse placement matrix, reducing evaluation time by more than a factor of five. Our method provides a practical route to uncertainty-aware operator learning in time-sensitive settings.

97 MATHEMATICS AND COMPUTING↗

An aerosol activation metamodel of v1.2.0 of the pyrcel cloud parcel model: development and offline assessment for use in an aerosol–climate model

We describe an emulator of a detailed cloud parcel model which has been trained to assess droplet nucleation from a complex, multimodal aerosol size distribution simulated by a global aerosol–climate model. The emulator is constructed using a sensitivity analysis approach (polynomial chaos expansion) which reproduces the behavior of the targeted parcel model across the full range of aerosol properties and meteorology simulated by the parent climate model. An iterative technique using aerosol fields sampled from a global model is used to identify the critical aerosol size distribution parameters necessary for accurately predicting activation. Across the large parameter space used to train them, the emulators estimate cloud droplet number concentration (CDNC) with a mean relative error of 9.2% for aerosol populations without giant cloud condensation nuclei (CCN) and 6.9% when including them. Versus a parcel model driven by those same aerosol fields, the best-performing emulator has a mean relative error of 4.6%, which is comparable with two commonly used activation schemes also evaluated here (which have mean relative errors of 2.9 and 6.7%, respectively). We identify the potential for regional biases in modeled CDNC, particularly in oceanic regimes, where our best-performing emulator tends to overpredict by 7%, whereas the reference activation schemes range in mean relative error from -3 to 7%. The emulators which include the effects of giant CCN are more accurate in continental regimes (mean relative error of 0.3%) but strongly overestimate CDNC in oceanic regimes by up to 22%, particularly in the Southern Ocean. Finally, the biases in CDNC resulting from the subjective choice of activation scheme could potentially influence the magnitude of the indirect effect diagnosed from the model incorporating it.

54 ENVIRONMENTAL SCIENCES↗

Exploiting Multi-Domain Features for Detection of Unclassified Electromagnetic Signals

Deep Learning based classification techniques have shown excellent performance in static environments, where the training and testing samples are drawn from the same distribution. However, real world scenarios often present samples that do not belong to the known set of classes chosen during training. This is quite common for electromagnetic signals, where it is impractical to assume that all possible waveforms are known a-priori, specially in scenarios like warfare. To address this problem, we propose a deep learning based adversarial model where the generator learns to generate waveform features that can deceive the discriminator model as true samples. We introduce domain knowledge of wireless signals by decomposing the signal into a lower dimensional unique feature set, which is used for classifying known versus unknown signals. We further introduce multiple domain representations of the signal to extract features and combine them together to accurately classify new waveforms as an unknown class. Our results show that combined features from multiple domains outperform any single domain representation, especially at low SNR regimes with fewer number of samples to classify.

99 - GENERAL AND MISCELLANEOUS↗

Exploiting Multi-Domain Features for Detection of Unclassified Electromagnetic Signals (Presentation)

Deep Learning based classification techniques have shown excellent performance in static environments, where the training and testing samples are drawn from the same distribution. However, real world scenarios often present samples that do not belong to the known set of classes chosen during training. This is quite common for electromagnetic signals, where it is impractical to assume that all possible waveforms are known a-priori, specially in scenarios like warfare. To address this problem, we propose a deep learning based adversarial model where the generator learns to generate waveform features that can deceive the discriminator model as true samples. We introduce domain knowledge of wireless signals by decomposing the signal into a lower dimensional unique feature set, which is used for classifying known versus unknown signals. We further introduce multiple domain representations of the signal to extract features and combine them together to accurately classify new waveforms as an unknown class. Our results show that combined features from multiple domains outperform any single domain representation, especially at low SNR regimes with fewer number of samples to classify.

99 - GENERAL AND MISCELLANEOUS↗

Selecting XFEL single-particle snapshots by geometric machine learning

A promising new route for structural biology is single-particle imaging with an X-ray Free-Electron Laser (XFEL). This method has the advantage that the samples do not require crystallization and can be examined at room temperature. However, high-resolution structures can only be obtained from a sufficiently large number of diffraction patterns of individual molecules, so-called single particles. Here, we present a method that allows for efficient identification of single particles in very large XFEL datasets, operates at low signal levels, and is tolerant to background. This method uses supervised Geometric Machine Learning (GML) to extract low-dimensional feature vectors from a training dataset, fuse test datasets into the feature space of training datasets, and separate the data into binary distributions of “single particles” and “non-single particles.” As a proof of principle, we tested simulated and experimental datasets of the Coliphage PR772 virus. We created a training dataset and classified three types of test datasets: First, a noise-free simulated test dataset, which gave near perfect separation. Second, simulated test datasets that were modified to reflect different levels of photon counts and background noise. These modified datasets were used to quantify the predictive limits of our approach. Third, an experimental dataset collected at the Stanford Linear Accelerator Center. The single-particle identification for this experimental dataset was compared with previously published results and it was found that GML covers a wide photon-count range, outperforming other single-particle identification methods. Moreover, a major advantage of GML is its ability to retrieve single particles in the presence of structural variability.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

PHOTOMETRIC REDSHIFT PROBABILITY DISTRIBUTIONS FOR GALAXIES IN THE SDSS DR8

We present redshift probability distributions for galaxies in the Sloan Digital Sky Survey (SDSS) Data Release 8 imaging data. We used the nearest-neighbor weighting algorithm to derive the ensemble redshift distribution N(z), and individual redshift probability distributions P(z) for galaxies with r < 21.8 and u < 29.0. As part of this technique, we calculated weights for a set of training galaxies with known redshifts such that their density distribution in five-dimensional color-magnitude space was proportional to that of the photometry-only sample, producing a nearly fair sample in that space. We estimated the ensemble N(z) of the photometric sample by constructing a weighted histogram of the training-set redshifts. We derived P(z)'s for individual objects by using training-set objects from the local color-magnitude space around each photometric object. Using the P(z) for each galaxy can reduce the statistical error in measurements that depend on the redshifts of individual galaxies. The spectroscopic training sample is substantially larger than that used for the DR7 release. The newly added PRIMUS catalog is now the most important training set used in this analysis by a wide margin. We expect the primary sources of error in the N(z) reconstruction to be sample variance and spectroscopic failures: The training sets are drawn from relatively small volumes of space, and some samples have large incompleteness. Using simulations we estimated the uncertainty in N(z) due to sample variance at a given redshift to be {approx}10%-15%. The uncertainty on calculations incorporating N(z) or P(z) depends on how they are used; we discuss the case of weak lensing measurements. The P(z) catalog is publicly available from the SDSS Web site.

79 ASTRONOMY AND ASTROPHYSICS↗