Search NASA⌕ Search

SEARCH · Search NASA

Results for “Bayesian Neural Network”

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 343 records · Page 19

Towards a data-driven model of hadronization using normalizing flows

We introduce a model of hadronization based on invertible neural networks that faithfully reproduces a simplified version of the Lund string model for meson hadronization. Additionally, we introduce a new training method for normalizing flows, termed MAGIC, that improves the agreement between simulated and experimental distributions of high-level (macroscopic) observables by adjusting single-emission (microscopic) dynamics. Our results constitute an important step toward realizing a machine-learning based model of hadronization that utilizes experimental data during training. Finally, we demonstrate how a Bayesian extension to this normalizing-flow architecture can be used to provide analysis of statistical and modeling uncertainties on the generated observable distributions.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

Extending Parsimonious Bayesian Inference

Parsimonious Bayesian inference is a theoretical framework for efficient data assimilation that seeks to balance increased consistency between predictions and training data against corresponding increases in model complexity. Within this framework, over-training is understood as optimization that encodes excessive information within model parameters while only achieving small improvements between predictions and training data. This project aims to develop practical methods of limiting excess model information during optimization. One key observation is that practical heuristics for parsimonious learning in high-dimensions must balance expressivity, i.e. the ability of the model to capture diverse predictions with only a few non-zero parameters, against discoverability, i.e. the ability to train the model with gradient-based optimization and drive parameters to low information states. As such, we developed logical activation functions that are able to adaptively approximate arbitrary truth tables that define Boolean logic operations within a probabilistic framework. These functions have demonstrated the ability to learn exclusive disjunction (XOR) and conditioned disjunction (if [condition] then [result_if_true] else [result_if_false]) within a single layer of a neural network. To efficiently exploit these activation functions to drive parsimonious learning required several other advances within the domain of variational inference. The most efficient form of complexity suppression is structured sparsification, driving most model parameters to zero while achieving the structural coherence among nonzeros needed for bandwidth reduction. Such models are not only far more efficient at suppressing information-theoretic complexity, they also reduce the other forms of complexity (computations, communication, storage, and the number of dependencies needed to evaluate predictions). Aiming to support enhanced sparsification, this project examined new approaches to high-dimensional variational inference that allow us to calibrate and control parameter uncertainty during optimization. By identifying which parameters can sustain sparsifying perturbations with little impact on prediction quality, we can develop better pruning strategies by framing them as approximate Bayesian inference. These advances also open paths to mitigate concerns with deploying advanced learning methods in resource-constrained environments, such as running models on power-limited or communication-limited devices.

97 MATHEMATICS AND COMPUTING↗

Data Imbalance, Uncertainty Quantification, and Transfer Learning in Data‐Driven Parameterizations: Lessons From the Emulation of Gravity Wave Momentum Transport in WACCM

Abstract Neural networks (NNs) are increasingly used for data‐driven subgrid‐scale parameterizations in weather and climate models. While NNs are powerful tools for learning complex non‐linear relationships from data, there are several challenges in using them for parameterizations. Three of these challenges are (a) data imbalance related to learning rare, often large‐amplitude, samples; (b) uncertainty quantification (UQ) of the predictions to provide an accuracy indicator; and (c) generalization to other climates, for example, those with different radiative forcings. Here, we examine the performance of methods for addressing these challenges using NN‐based emulators of the Whole Atmosphere Community Climate Model (WACCM) physics‐based gravity wave (GW) parameterizations as a test case. WACCM has complex, state‐of‐the‐art parameterizations for orography‐, convection‐, and front‐driven GWs. Convection‐ and orography‐driven GWs have significant data imbalance due to the absence of convection or orography in most grid points. We address data imbalance using resampling and/or weighted loss functions, enabling the successful emulation of parameterizations for all three sources. We demonstrate that three UQ methods (Bayesian NNs, variational auto‐encoders, and dropouts) provide ensemble spreads that correspond to accuracy during testing, offering criteria for identifying when an NN gives inaccurate predictions. Finally, we show that the accuracy of these NNs decreases for a warmer climate (4 × CO 2 ). However, their performance is significantly improved by applying transfer learning, for example, re‐training only one layer using ∼1% new data from the warmer climate. The findings of this study offer insights for developing reliable and generalizable data‐driven parameterizations for various processes, including (but not limited to) GWs.

54 ENVIRONMENTAL SCIENCES↗

Explainable and Differentiable Reinforcement Learning for Multi-objective Optimization in Particle Accelerators

Operating particle accelerators involves optimizing multiple goals simultaneously, which can be challenging due to trade-offs among objectives. While evolutionary algorithms like the genetic algorithm (GA) have been used for various Multi-Objective Optimization (MOO) tasks, they are not inherently suited for complex control problems. This talk highlights two variations of Reinforcement Learning (RL) for concurrently optimizing heat load and trip rates at the Continuous Electron Beam Accelerator Facility (CEBAF). The problem involves strict constraints on individual states, actions, and overall energy requirements of the beam. First, this talk highlights how differentiability can be harnessed through a Deep Differentiable Reinforcement Learning (DDRL) approach to address MOO issues within particle accelerators. We examine the DDRL method alongside Model Free Reinforcement Learning (MFRL), GA, and Bayesian Optimization (BO). The performance of these methods is assessed by generating a Pareto-front for two objectives. Our findings indicate that DDRL excels in handling high-dimensional problems more effectively than MFRL, BO, and GA. Next, we will show integration of explainable physics-based constraints into RL algorithms to enhance trans- parency and trust in decision-making processes by enabling users to verify that agents adhere to established physical principles. This surrogate function can be modeled using neural networks or sparse dictionary mod- els. By examining the mathematical form of the learned constraint function, we are able to confirm the agent has learned to use the established physics of each environment provided but the surrogate model. In addi- tion, we find that the introduction of a mathematical functional dictionary based surrogate model enables our reinforcement learning algorithms to reliably converge for difficult high-dimensional accelerator controls environments.

Rajput, Kishansingh [Thomas Jefferson National Acc↗

Improving Subsurface Stress Characterization for Carbon Dioxide Storage Projects by Incorporating Machine Learning Techniques

The overall objective of this project is to develop a framework for reliable characterization and prediction of the state of stress in the overburden and underburden (including the basement) in CO 2 storage reservoirs using machine learning and integrated geomechanics and geophysical methods. Specifically, we propose to develop workflow encompassing of technologies and/or methods to predict stress and pressure changes due to CO 2 injection in an active tertiary recovery site and their impacts on subtle fault activation, fractures and occurrence of microseismic events and compare responses to field observations. In this project, we anticipate using dataset from the Farnsworth field Unit (FWU) which is operated by Purdure Petroleum. A novel elastic-waveform VSP inversion technique will be used to estimate high-resolution spatial and temporal changes of elastic moduli in CO 2 storage reservoirs, which will be combined with velocity-stress relationship derived from laboratory tests to obtain subsurface pressure and stress. Clustered microseismic data will be jointly inverted for improved focal mechanisms. Least-squares reverse-time migration of microseismic waveform data will be performed to directly image fracture/fault zones. Additionally, a deep neural network machine learning technique with convolutional and recurrent layers will be used for learning the spectro-temporal structures in microseismic waveforms. The results of this geotechnical data analysis will be integrated to develop a high-resolution 3D mechanical earth model extending from the overburden sealing formations to the underburden including the basement. Mechanical properties will be derived through integration of mechanical logs, tests, available results from chemo-mechanical laboratory tests, and elastic inversion of seismic data using a combination of Bayesian and stochastic methods as well as machine learning technique. Failure features (faults/fractures) will be represented and/or modeled based on seismic and core data analysis. A transient hydrodynamic-geomechanical model will be developed through coupling with the calibrated FWU reservoir simulation model. The full physics coupled model will be used to train a reduced order proxy model using machine learning algorithm for estimating stress which will then be used with appropriate constitutive relationships and forward seismological models to simulate pressure changes and induced microseismicity. An advanced optimization framework will be developed to perform a history match to minimize error between field observations and simulated. The history matched proxy model will be verified against the full-physics equivalent. The field observations that will be used in the coupled model calibration process include pressure/stress inverted from VSP, moment magnitude from microseismic analysis, real time downhole pressure measurements, production and injection data. Parameter sensitivity and uncertainty analysis will be performed to characterize the impact of model parameter uncertainty on stress estimates. The proposed project will have significant impact on future field implementation of the proposed technology. Because the project field site is an ongoing CO 2 EOR development, the value of the new technology will be demonstrated in an operational context and evaluated as a viable risk mitigation strategy. Cost/benefit will be evaluated together with the various commercial incentives for CO 2 sequestration available to oil and gas operators. The extensive available dataset and ongoing data acquisition under the SWP Phase III work plan provides flexibility for investigation of multiple approaches and reduces technical risk.

58 GEOSCIENCES↗

Training Spiking Neural Networks Using Combined Learning Approaches

Spiking neural networks (SNNs), the class of neural networks used in neuromorphic computing, are difficult to train using traditional back-propagation techniques. Spike timingdependent plasticity (STDP) is a biologically inspired learning mechanism that can be used to train SNNs. Evolutionary algorithms have also been demonstrated as a method for training SNNs. In this work, we explore the relationship between these two training methodologies. We evaluate STDP and evolutionary optimization as standalone methods for training networks, and also evaluate a combined approach where STDP weight updates are applied within an evolutionary algorithm. We also apply Bayesian hyperparameter optimization as a meta learner for each of the algorithms. We find that STDP by itself is not an ideal learning rule for randomly connected networks, while the inclusion of STDP within an evolutionary algorithm leads to similar performance, with a few interesting differences. This study suggests future work in understanding the relationship between network topology and learning rules.

Elbrecht, Daniel↗

Structure–Property Linkage in Alloys Using Graph Neural Network and Explainable Artificial Intelligence

Deep learning tools have recently shown significant potential for accelerating the prediction of microstructure–property linkage in materials. While deep neural networks like convolution neural networks (CNNs) can extract physics information from 3D microstructure images, they often require a large network architecture and substantial training time. In this research, we trained a graph neural network (GNN) using phase field generated microstructures of Ni-Al alloys to predict the evolution of mechanical properties. We found that a single GNN is capable of accurately predicting the strengthening of Ni-Al alloys with microstructures of varying sizes and dimensions, which cannot otherwise be done with a CNN. Additionally, GNN requires significantly less GPU utilization than CNN and offers more interpretable explanation of predictions using saliency analysis as features are manually defined in the graph. We also utilize explainable artificial intelligence tool Bayesian Inference to determine the coefficients in the power law equation that governs coarsening of precipitates. Overall, our work demonstrates the ability of the GNN to accurately and efficiently extract relevant information from material microstructures without having restrictions on microstructure size or dimension and offers an interpretable explanation.

Chemistry↗

Uncertainty-aware Continuous Implicit Neural Representations for Remote Sensing Object Counting

Many existing object counting methods rely on density map estimation (DME) of the discrete grid representation by decoding extracted image semantic features from designed convolutional neural networks (CNNs). Relying on discrete density maps not only leads to information loss dependent on the original image resolution, but also has a scalability issue when analyzing high-resolution images with cubically increasing memory complexity. Furthermore, none of the existing methods can offer reliable uncertainty quantification (UQ) for the derived count estimates. To overcome these limitations, we design UNcertainty-aware, hypernetwork-based Implicit neural representations for Counting (UNIC) to assign probabilities and the corresponding counting confidence over continuous spatial coordinates. We derive a sampling-based Bayesian counting loss function and develop the corresponding model training algorithm. UNIC outperforms existing methods on the Remote Sensing Object Counting (RSOC) dataset with reliable UQ and improved interpretability of the derived count estimates. Our code is available at https://github.com/SiyuanXu-tamu/UNIC.

97 MATHEMATICS AND COMPUTING↗

Gaussian processes with built-in dimensionality reduction: Applications to high-dimensional uncertainty propagation

Uncertainty quantification (UQ) tasks, such as model calibration, uncertainty propagation, and optimization under uncertainty, typically require several thousand evaluations of the underlying computer codes. To cope with the cost of simulations, one replaces the real response surface with a cheap surrogate based, e.g., on polynomial chaos expansions, neural networks, support vector machines, or Gaussian processes (GP). However, the number of simulations required to learn a generic multivariate response grows exponentially as the input dimension increases. This curse of dimensionality can only be addressed, if the response exhibits some special structure that can be discovered and exploited. A wide range of physical responses exhibit a special structure known as an active subspace (AS). An AS is a linear manifold of the stochastic space characterized by maximal response variation. The idea is that one should first identify this low dimensional manifold, project the high-dimensional input onto it, and then link the projection to the output. If the dimensionality of the AS is low enough, then learning the link function is a much easier problem than the original problem of learning a high-dimensional function. The classic approach to discovering the AS requires gradient information, a fact that severely limits its applicability. Furthermore, and partly because of its reliance to gradients, it is not able to handle noisy observations. The latter is an essential trait if one wants to be able to propagate uncertainty through stochastic simulators, e.g., through molecular dynamics codes. In this work, we develop a probabilistic version of AS which is gradient-free and robust to observational noise. Our approach relies on a novel Gaussian process regression with built-in dimensionality reduction. In particular, the AS is represented as an orthogonal projection matrix that serves as yet another covariance function hyper-parameter to be estimated from the data. To train the model, we design a two-step maximum likelihood optimization procedure that ensures the orthogonality of the projection matrix by exploiting recent results on the Stiefel manifold, i.e., the manifold of matrices with orthogonal columns. The additional benefit of our probabilistic formulation, is that it allows us to select the dimensionality of the AS via the Bayesian information criterion. We validate our approach by showing that it can discover the right AS in synthetic examples without gradient information using both noiseless and noisy observations. We demonstrate that our method is able to discover the same AS as the classical approach in a challenging one-hundred-dimensional problem involving an elliptic stochastic partial differential equation with random conductivity. Finally, we use our approach to study the effect of geometric and material uncertainties in the propagation of solitary waves in a one dimensional granular system.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

A Statistician’s Overview of Physics-Informed Neural Networks for Spatio-Temporal Data

The recent success of deep neural network models with physical constraints (so-called, Physics-Informed Neural Networks, PINNs) has led to renewed interest in the incorporation of mechanistic information in predictive models. Statisticians and others have long been interested in this problem, which has led to several practical and innovative solutions dating back decades. In this overview, we focus on the problem of data-driven prediction and inference of dynamic spatio-temporal processes that include mechanistic information, such as would be available from partial differential equations, with a strong focus on the quantification of uncertainty associated with data, process, and parameters. Here, we give a brief review of several paradigms and focus our attention on Bayesian implementations given they naturally accommodate uncertainty quantification. We then show that it is straight-forward to include the Bayesian PINN (B-PINN) within the Bayesian hierarchical model (BHM) framework that has long been considered for modeling dynamic spatio-temporal processes. Such a BHM-PINN is illustrated via a simulation study in which a latent nonlinear Burgers’ equation PDE governs the dynamics of Poisson distributed spatio-temporal data. Supplementary materials for this article are available online, including a standardized description of the materials available for reproducing the work.

Bayesian↗

A Deep Generative Deconvolutional Image Model

A deep generative model is developed for representation and analysis of images, based on a hierarchical convolutional dictionary-learning framework. Stochastic unpooling is employed to link consecutive layers in the model, yielding top-down image generation. A Bayesian support vector machine is linked to the top-layer features, yielding max-margin discrimination. Deep deconvolutional inference is employed when testing, to infer the latent features, and the top-layer features are connected with the max-margin classifier for discrimination tasks. The model is efficiently trained using a Monte Carlo expectation-maximization (MCEM) algorithm; the algorithm is implemented on graphical processor units (GPU) to enable large-scale learning, and fast testing. Excellent results are obtained on several benchmark datasets, including ImageNet, demonstrating that the proposed model achieves results that are highly competitive with similarly sized convolutional neural networks.

convolutional neural network↗

DeepHyper: Asynchronous Hyperparameter Search for Deep Neural Networks

Hyperparameters employed by deep learning (DL) methods play a substantial role in the performance and reliability of these methods in practice. Unfortunately, finding performance-optimizing hyperparameter settings is a notoriously difficult task. Hyperparameter search methods typically have limited production-strength implementations or do not target scalability within a highly parallel machine, portability across different machines, experimental comparison between different methods, and tighter integration with workflow systems. In this paper, we present DeepHyper, a Python package that provides a common interface for the implementation and study of scalable hyperparameter search methods. It adopts the Balsam workflow system to hide the complexities of running large numbers of hyperparameter configurations in parallel on high-performance computing (HPC) systems. We implement and study asynchronous model-based search methods that consist of sampling a small number of input hyperparameter configurations and progressively fitting surrogate models over the input-output space until exhausting a user-defined budget of evaluations. We evaluate the efficacy of these methods relative to approaches such as random search, genetic algorithms, Bayesian optimization, and hyperband on DL benchmarks on CPU- and GPU-based HPC systems.

Balaprakash, Prasanna↗

A Bayesian approach to strong lens finding in the era of wide-area surveys

ABSTRACT The arrival of the Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST), Euclid-Wide and Roman wide-area sensitive surveys will herald a new era in strong lens science in which the number of strong lenses known is expected to rise from $\mathcal {O}(10^3)$ to $\mathcal {O}(10^5)$. However, current lens-finding methods still require time-consuming follow-up visual inspection by strong lens experts to remove false positives which is only set to increase with these surveys. In this work, we demonstrate a range of methods to produce calibrated probabilities to help determine the veracity of any given lens candidate. To do this we use the classifications from citizen science and multiple neural networks for galaxies selected from the Hyper Suprime-Cam survey. Our methodology is not restricted to particular classifier types and could be applied to any strong lens classifier which produces quantitative scores. Using these calibrated probabilities, we generate an ensemble classifier, combining citizen science, and neural network lens finders. We find such an ensemble can provide improved classification over the individual classifiers. We find a false-positive rate of 10−3 can be achieved with a completeness of 46 per cent, compared to 34 per cent for the best individual classifier. Given the large number of galaxy–galaxy strong lenses anticipated in LSST, such improvement would still produce significant numbers of false positives, in which case using calibrated probabilities will be essential for population analysis of large populations of lenses and to help prioritize candidates for follow-up.

79 ASTRONOMY AND ASTROPHYSICS↗

BayesCLUMPY: BAYESIAN INFERENCE WITH CLUMPY DUSTY TORUS MODELS

Our aim is to present a fast and general Bayesian inference framework based on the synergy between machine learning techniques and standard sampling methods and apply it to infer the physical properties of clumpy dusty torus using infrared photometric high spatial resolution observations of active galactic nuclei. We make use of the Metropolis-Hastings Markov Chain Monte Carlo algorithm for sampling the posterior distribution function. Such distribution results from combining all a priori knowledge about the parameters of the model and the information introduced by the observations. The main difficulty resides in the fact that the model used to explain the observations is computationally demanding and the sampling is very time consuming. For this reason, we apply a set of artificial neural networks that are used to approximate and interpolate a database of models. As a consequence, models not present in the original database can be computed ensuring continuity. We focus on the application of this solution scheme to the recently developed public database of clumpy dusty torus models. The machine learning scheme used in this paper allows us to generate any model from the database using only a factor of 10{sup -4} of the original size of the database and a factor of 10{sup -3} in computing time. The posterior distribution obtained for each model parameter allows us to investigate how the observations constrain the parameters and which ones remain partially or completely undetermined, providing statistically relevant confidence intervals. As an example, the application to the nuclear region of Centaurus A shows that the optical depth of the clouds, the total number of clouds, and the radial extent of the cloud distribution zone are well constrained using only six filters. The code is freely available from the authors.

79 ASTRONOMY AND ASTROPHYSICS↗

Unsupervised machine learning discovery of structural units and transformation pathways from imaging data

We show that unsupervised machine learning can be used to learn chemical transformation pathways from observational Scanning Transmission Electron Microscopy (STEM) data. To enable this analysis, we assumed the existence of atoms, a discreteness of atomic classes, and the presence of an explicit relationship between the observed STEM contrast and the presence of atomic units. With only these postulates, we developed a machine learning method leveraging a rotationally invariant variational autoencoder (VAE) that can identify the existing molecular fragments observed within a material. The approach encodes the information contained in STEM image sequences using a small number of latent variables, allowing the exploration of chemical transformation pathways by tracing the evolution of atoms in the latent space of the system. The results suggest that atomically resolved STEM data can be used to derive fundamental physical and chemical mechanisms involved, by providing encodings of the observed structures that act as bottom-up equivalents of structural order parameters. The approach also demonstrates the potential of variational (i.e., Bayesian) methods in the physical sciences and will stimulate the development of more sophisticated ways to encode physical constraints in the encoder–decoder architectures and generative physical laws and causal relationships in the latent space of VAEs.

97 MATHEMATICS AND COMPUTING↗

DESI Strong Lens Foundry. I. HST Observations and Modeling with GIGA-Lens

We present the Dark Energy Spectroscopic Instrument (DESI) Strong Lens Foundry. We discovered ∼3500 new strong gravitational lens candidates in the DESI Legacy Imaging Surveys using residual neural networks (ResNet). We observed a subset (51) of our candidates using the Hubble Space Telescope (HST). Except for one ambiguous case, we have confirmed 50 of the 51 candidates to be strong lenses. We also briefly describe spectroscopic follow-up observations by DESI and Keck NIRES programs. From this very rich data set, a number of studies will be carried out, including evaluating the quality of the ResNet search candidates and lens modeling. In this paper, we present our initial effort in these directions. In particular, as a demonstration, we present the lens model for DESI-165.4754−06.0423, with imaging data from HST, and lens and source redshifts from DESI and Keck NIRES, respectively. In this effort, we have applied a fully forward-modeling Bayesian approach (GIGA-Lens), using multiple GPUs, to a strong lens with HST data, and achieved statistical convergence.

79 ASTRONOMY AND ASTROPHYSICS↗

Forward variable selection enables fast and accurate dynamic system identification with Karhunen-Loève decomposed Gaussian processes

A promising approach for scalable Gaussian processes (GPs) is the Karhunen-Loève (KL) decomposition, in which the GP kernel is represented by a set of basis functions which are the eigenfunctions of the kernel operator. Such decomposed kernels have the potential to be very fast, and do not depend on the selection of a reduced set of inducing points. However KL decompositions lead to high dimensionality, and variable selection thus becomes paramount. This paper reports a new method of forward variable selection, enabled by the ordered nature of the basis functions in the KL expansion of the Bayesian Smoothing Spline ANOVA kernel (BSS-ANOVA), coupled with fast Gibbs sampling in a fully Bayesian approach. It quickly and effectively limits the number of terms, yielding a method with competitive accuracies, training and inference times for tabular datasets of low feature set dimensionality. Theoretical computational complexities are O ( N P 2 ) in training and O ( P ) per point in inference, where N is the number of instances and P the number of expansion terms. The inference speed and accuracy makes the method especially useful for dynamic systems identification, by modeling the dynamics in the tangent space as a static problem, then integrating the learned dynamics using a high-order scheme. The methods are demonstrated on two dynamic datasets: a ‘Susceptible, Infected, Recovered’ (SIR) toy problem, along with the experimental ‘Cascaded Tanks’ benchmark dataset. Comparisons on the static prediction of time derivatives are made with a random forest (RF), a residual neural network (ResNet), and the Orthogonal Additive Kernel (OAK) inducing points scalable GP, while for the timeseries prediction comparisons are made with LSTM and GRU recurrent neural networks (RNNs) along with the SINDy package.

Hayes, Kyle↗

Occlusion-Perturbed Deep Learning for Probabilistic Solar Forecasting via Sky Images

Solar forecasting is shifting to the probabilistic paradigm due to the inherent uncertainty within the solar resource. Input uncertainty quantification is one of the widely used and best-performing ways to model solar uncertainty. However, compared to other sources of inputs, such as numerical weather prediction models, pure sky image-based probabilistic solar forecasting lags behind. In this research, an occlusion-perturbed convolutional neural network, named the PSolarNet, is developed. The PSolarNet provides very short-term deterministic forecasts, forecast scenarios, and probabilistic forecasts of the global horizontal irradiance from sky image sequences. Case studies based on 6 years of open-source data show that the developed PSolarNet is able to generate accurate 10-minute ahead deterministic forecasts with a 5.62% normalized root mean square error, realistic and diverse forecast scenarios with a 0.966 average correlation with the actual time series, and reliable and sharp probabilistic forecasts with a 2.77% normalized continuous ranked probability score.

Bayesian model averaging↗