Search NASA⌕ Search

SEARCH · Search NASA

Results for “Bayesian Methods”

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 91 records · Page 5

Sequential Bayesian inference of transition rates in the hidden Markov model for multi-state system degradation

The more easily available system performance data and advances in data analytics have provided us with opportunities to optimize maintenance programs for engineered systems, for example nuclear power plants. One key task in maintenance optimization is to obtain an accurate model for system degradation. Here, in this research, we propose a Bayesian method to address this problem. Noting that systems usually exhibit multiple states and that the actual state of a system usually is not directly observable, in the method we first model the system degradation process and the observation process based on a hidden Markov model. Then we develop a sequential Bayesian inference algorithm based on importance sampling and the forward algorithm to infer the posterior distributions of the transition rates in the hidden Markov model based on available observations. The proposed Bayesian method allows us to take advantage of evidence from multiple sources, and also allows us to perform Bayesian inference sequentially, without the need to use the entire history of observations every time new observations are collected. We demonstrate the proposed method using both synthetic data for a nuclear power plant feedwater pump and realistic data for a nuclear power plant chemistry analytical device.

42 ENGINEERING↗

Linking big models to big data: efficient ecosystem model calibration through Bayesian model emulation

Data-model integration plays a critical role in assessing and improving our capacity to predict ecosystem dynamics. Similarly, the ability to attach quantitative statements of uncertainty around model forecasts is crucial for model assessment and interpretation and for setting field research priorities. Bayesian methods provide a rigorous data assimilation framework for these applications, especially for problems with multiple data constraints. However, the Markov chain Monte Carlo (MCMC) techniques underlying most Bayesian calibration can be prohibitive for computationally demanding models and large datasets. We employ an alternative method, Bayesian model emulation of sufficient statistics, that can approximate the full joint posterior density, is more amenable to parallelization, and provides an estimate of parameter sensitivity. Analysis involved informative priors constructed from a meta-analysis of the primary literature and specification of both model and data uncertainties, and it introduced novel approaches to autocorrelation corrections on multiple data streams and emulating the sufficient statistics surface. We report the integration of this method within an ecological workflow management software, Predictive Ecosystem Analyzer (PEcAn), and its application and validation with two process-based terrestrial ecosystem models: SIPNET and ED2. In a test against a synthetic dataset, the emulator was able to retrieve the true parameter values. A comparison of the emulator approach to standard brute-force MCMC involving multiple data constraints showed that the emulator method was able to constrain the faster and simpler SIPNET model's parameters with comparable performance to the brute-force approach but reduced computation time by more than 2 orders of magnitude. The emulator was then applied to calibration of the ED2 model, whose complexity precludes standard (brute-force) Bayesian data assimilation techniques. Both models are constrained after assimilation of the observational data with the emulator method, reducing the uncertainty around their predictions. Performance metrics showed increased agreement between model predictions and data. Our study furthers efforts toward reducing model uncertainties, showing that the emulator method makes it possible to efficiently calibrate complex models.

58 GEOSCIENCES↗

A Data-Fusion Method using Bayesian Approach to Enhance Raw Data Accuracy of Position and Distance Measurements for Connected Vehicles

Accurate positioning of vehicles is a critical element of autonomous and connected vehicle systems. Most of other studies heavily focused on enhancing simultaneous localization and mapping (SLAM) methods, i.e., constructing or updating a map of an unknown environment and tracking an object within the map. This paper provides a method that can, in addition to existing SLAM or relevant methods, enhance the raw measurements of position and distance. The basic idea of this study is to identify and update the error distribution of each data source by combining all available information. A Bayesian approach was incorporated to estimate and update the error distribution of individual data sources or sensors. The proposed method can be conducted in real-time environments, and a self-learning scheme determines whether enough data has been collected to further improve the accuracy of such measurements. The simulated experiments show that the proposed model noticeably improves the accuracy of position and distance measurements. Especially, the estimated biases of position coordinates and distance measures are very close to the biases of true error distributions, with the R-squared over 0.98. A similar approach can also be utilized to enhance accuracy of other sensors or measurements in connected vehicle or relevant systems, where multi-data sources are available.

Lim, Hyeonsup↗

An Adaptive Sparse-Grid High-Order Stochastic Collocation Method for Bayesian Inference in Groundwater Reactive Transport Modeling

Although Bayesian analysis has become vital to the quantification of prediction uncertainty in groundwater modeling, its application has been hindered due to the computational cost associated with numerous model executions needed for exploring the posterior probability density function (PPDF) of model parameters. This is particularly the case when the PPDF is estimated using Markov Chain Monte Carlo (MCMC) sampling. In this study, we develop a new approach that improves computational efficiency of Bayesian inference by constructing a surrogate system based on an adaptive sparse-grid high-order stochastic collocation (aSG-hSC) method. Unlike previous works using first-order hierarchical basis, we utilize a compactly supported higher-order hierar- chical basis to construct the surrogate system, resulting in a significant reduction in the number of computational simulations required. In addition, we use hierarchical surplus as an error indi- cator to determine adaptive sparse grids. This allows local refinement in the uncertain domain and/or anisotropic detection with respect to the random model parameters, which further improves computational efficiency. Finally, we incorporate a global optimization technique and propose an iterative algorithm for building the surrogate system for the PPDF with multiple significant modes. Once the surrogate system is determined, the PPDF can be evaluated by sampling the surrogate system directly with very little computational cost. The developed method is evaluated first using a simple analytical density function with multiple modes and then using two synthetic groundwater reactive transport models. The groundwater models represent different levels of complexity; the first example involves coupled linear reactions and the second example simulates nonlinear ura- nium surface complexation. The results show that the aSG-hSC is an effective and efficient tool for Bayesian inference in groundwater modeling in comparison with conventional MCMC sim- ulations. The computational efficiency is expected to be more beneficial to more computational expensive groundwater problems.

Zhang, Guannan↗

Physicochemical and Performance Characterization of Six Commercial Organic Solvent Nanofiltration Membranes

This work introduces a novel, gradient-free metamaterial design method based on Gaussian process regression to represent the density field of a unit cell. The dimension of the design space is determined by the covariance matrix dimension in the Gaussian process regression. We propose compressing this matrix using an autoencoder, enabling the decoder to generate the density field and effectively reduce the originally large design space to a lower-dimensional subspace. In this compressed space, we employ an active learning method, Bayesian Adaptive Direct Search (BADS), for efficient exploration of the design space. We demonstrate that for simple 2D designs aimed at maximizing unit cell stiffness, our method yields results comparable to those of standard topology optimization. Furthermore, we extend our approach to various mechanical problems, from linear elasticity to hyperelastic large deformation and elasto-plasticity under finite deformation, to 3D metamaterial design. This illustrates the method’s versatility and effectiveness across a range of applications.

Wu, Haoran↗

Probing Non-Standard Interactions in NOvA with Bayesian MCMC Methods

We present a search for non-standard neutrino interactions (NSI) using NOvA's joint $\nu_e$ appearance and $\nu_\mu$ disappearance samples in both neutrino and antineutrino beam modes. A Bayesian Markov Chain Monte Carlo approach is used to map the posterior distribution over NSI parameters $\varepsilon_{e\mu}$, $\varepsilon_{e\tau}$, and $\varepsilon_{\mu\tau}$, simultaneously with standard oscillation parameters. We investigate the effect of prior choice for the complex NSI parameters and present projected sensitivity to off-diagonal NSI.

Huang, Xiaoyan [Mississippi U.]↗

Radiation Source Mapping with Bayesian Inverse Methods

In this work, we present a method to map the spectral and spatial distributions of radioactive sources using a limited number of detectors. Locating and identifying radioactive materials is important for border monitoring, in accounting for special nuclear material in processing facilities, and in cleanup operations following a radioactive material spill. Most methods to analyze these types of problems make restrictive assumptions about the distribution of the source. In contrast, the source mapping method presented here allows an arbitrary three-dimensional distribution in space and a gamma peak distribution in energy. To apply the method, the problem is cast as an inverse problem where the system’s geometry and material composition are known and fixed, while the radiation source distribution is sought. A probabilistic Bayesian approach is used to solve the resulting inverse problem since the system of equations is ill-posed. The posterior is maximized with a Newton optimization method. The probabilistic approach also provides estimates of the confidence in the final source map prediction. A set of adjoint, discrete ordinates flux solutions, obtained in this work by the Denovo code, is required to efficiently compute detector responses from a candidate source distribution. These adjoint fluxes form the linear mapping from the state space to the response space. The test of the method’s success is simultaneously locating a set of 137 Cs and 60 Co gamma sources in a room. This test problem is solved using experimental measurements that we collected for this purpose. Because of the weak sources available for use in the experiment, some of the expected photopeaks were not distinguishable from the Compton continuum. However, by supplanting 14 flawed measurements (out of a total of 69) with synthetic responses computed by MCNP, the proof-of-principle source mapping was successful. The locations of the sources were predicted within 25 cm for two of the sources and 90 cm for the third, in a room with an ~4-x 4-m floor plan. Finally, the predicted source intensities were within a factor of ten of their true value.

07 ISOTOPE AND RADIATION SOURCES↗

Stopping Power of Different Ions in Si Measured with a Bulk Sample Method and Bayesian Inference Data Analysis

The accuracy of ion beam analysis experiments depends critically on the stopping power values available. While for H and He ions accuracies normally better than 5% are achieved by usual interpolative schemes such as SRIM, for heavier ions the accuracy is worse. One of the main reasons is that the experimental data bases are very sparse, even for important materials such as Si. New measurements are therefore needed. Measurement of stopping power is often made with transmission in thin films, with the usual problems of film thickness homogeneity. We have previously developed an alternative method based on measuring bulk spectra, and fitting the yield by treating the stopping power as a fit parameter in a Bayesian inference Markov chain Monte Carlo procedure included in the standard IBA code NDF. We report on improvements of the method and on its application to the determination of the stopping power of {sup 7}Li in Si. To validate the method, we also apply it to the stopping of {sup 4}He in Si, which is known with 2% accuracy.

36 MATERIALS SCIENCE↗

Measurement of the electron antineutrino mass in tritium beta decay in the Troitsk nu-mass experiment

The results obtained in the Troitsk nu-mass experiment by measuring the electron-antineutrino mass in tritium beta decay are presented. The facility used consists of a gaseous windowless tritium source and an electrostatic electron spectrometer involving an adiabatic magnetic collimation. Runs in which measurement conditions were reliably established were thoroughly selected in analyzing data obtained from 1994 to 2004. All known systematic effects were taken into account. For the square of the electron-antineutrino mass, the treatment of measured spectra yielded the following result: m{sub {nu}}{sup 2} = -0.67 {+-} 1.89{sub stat.} {+-} 1.68{sub syst.} eV{sup 2}. The use of the Bayesian method and the Feldman-Cousins unified approach made it possible to obtain the following upper limits on the mass: m{sub {nu}} < 2.12 eV (at a 95% C.L.; Bayesian method) and m{sub {nu}} < 2.05 eV (at a 95% C.L., Feldman-Cousins method). At the same time, an estimation of the sensitivity limit without allowance for negative values of the square of the mass leads to m{sub {nu}} < 2.2 eV (at a 95% C.L.). Measured spectra were analyzed for the possible existence of an additional structure (step) in the electron spectrum near the boundary energy. The conclusion drawn from this analysis was that, within the existing statistical errors, there are no reasons for introducing such a feature.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

Constraining transfer cross sections using Bayes' theorem

Background: Being able to rigorously quantify the uncertainties in reaction models is crucial to moving this field forward. Even though Bayesian methods are becoming increasingly popular in nuclear theory, they have yet to be implemented and applied in reaction theory problems. Purpose: The purpose of this work is to investigate, using Bayesian methods, the uncertainties in the optical potentials generated from fits to elastic-scattering data and the subsequent uncertainties in the transfer predictions. We also study the differences in two reaction models where the parameters are constrained in a similar manner, as well as the impact of reducing the experimental error bars on the data used to constrain the parameters. Method: We use Bayes' theorem combined with a Markov chain Monte Carlo to determine posterior distributions for the parameters of the optical model, constrained by neutron-, proton-, and/or deuteron-target elastic scattering. These potentials are then used to predict transfer cross sections within the adiabatic wave approximation or the distorted-wave Born approximation. Results: We study a number of reactions involving deuteron projectiles with energies in the range 10–25 MeV/nucleon on targets with mass A = 48–208. The case of 48 Ca(d,p) 49 Ca transfer to the ground state is described in detail. A comparative study of the effect of the size of experimental errors is also performed. Five transfer reactions are studied, and their results are then compiled in order to systematically identify trends.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗

A Scalable Gradient Free Method for Bayesian Experimental Design with Implicit Models

Bayesian experimental design (BED) is to answer the question that how to choose designs that maximize the information gathering. For implicit models, where the likelihood is intractable but sampling is possible, conventional BED methods have difficulties in efficiently estimating the posterior distribution and maximizing the mutual information (MI) between data and parameters. Recent work proposed the use of gradient ascent to maximize a lower bound on MI to deal with these issues. However, the approach requires a sampling path to compute the pathwise gradient of the MI lower bound with respect to the design variables, and such a pathwise gradient is usually inaccessible for implicit models. In this paper, we propose a novel approach that leverages recent advances in stochastic approximate gradient ascent incorporated with a smoothed variational MI estimator for efficient and robust BED. Without the necessity of pathwise gradients, our approach allows the design process to be achieved through a unified procedure with an approximate gradient for implicit models. Several experiments show that our approach outperforms baseline methods, and significantly improves the scalability of BED in high-dimensional problems.

Zhang, Jiaxin↗

BAYESIAN TECHNIQUES FOR COMPARING TIME-DEPENDENT GRMHD SIMULATIONS TO VARIABLE EVENT HORIZON TELESCOPE OBSERVATIONS

The Event Horizon Telescope (EHT) is a millimeter-wavelength, very-long-baseline interferometry (VLBI) experiment that is capable of observing black holes with horizon-scale resolution. Early observations have revealed variable horizon-scale emission in the Galactic Center black hole, Sagittarius A* (Sgr A*). Comparing such observations to time-dependent general relativistic magnetohydrodynamic (GRMHD) simulations requires statistical tools that explicitly consider the variability in both the data and the models. We develop here a Bayesian method to compare time-resolved simulation images to variable VLBI data, in order to infer model parameters and perform model comparisons. We use mock EHT data based on GRMHD simulations to explore the robustness of this Bayesian method and contrast it to approaches that do not consider the effects of variability. We find that time-independent models lead to offset values of the inferred parameters with artificially reduced uncertainties. Moreover, neglecting the variability in the data and the models often leads to erroneous model selections. We finally apply our method to the early EHT data on Sgr A*.

79 ASTRONOMY AND ASTROPHYSICS↗

A Sparse Grid Method for Bayesian Uncertainty Quantification with Application to Large Eddy Simulation Turbulence Models

There is wide agreement that the accuracy of turbulence models suffer from their sensitivity with respect to physical input data, the uncertainties of user-elected parameters, as well as the model inadequacy. However, the application of Bayesian inference to systematically quantify the uncertainties in parameters, by means of exploring posterior probability density functions (PPDFs), has been hindered by the prohibitively daunting computational cost associated with the large number of model executions, in addition to daunting computation time per one turbulence simulation. In this effort, we perform in this paper an adaptive hierarchical sparse grid surrogate modeling approach to Bayesian inference of large eddy simulation (LES). First, an adaptive hierarchical sparse grid surrogate for the output of forward models is constructed using a relatively small number of model executions. Using such surrogate, the likelihood function can be rapidly evaluated at any point in the parameter space without simulating the computationally expensive LES model. This method is essentially similar to those developed in Zhang et al. (Water Resour Res 49:6871–6892, 2013) for geophysical and groundwater models, but is adjusted and applied here for a much more challenging problem of uncertainty quantification of turbulence models. Through a numerical demonstration of the Smagorinsky model of two-dimensional flow around a cylinder at sub-critical Reynolds number, our approach is proven to significantly reduce the number of costly LES executions without losing much accuracy in the posterior probability estimation. Here, the model parameters are calibrated against synthetic data related to the mean flow velocity and Reynolds stresses at different locations in the flow wake. The influence of the user-elected LES parameters on the quality of output data will be discussed.

97 MATHEMATICS AND COMPUTING↗

A Bayesian approach for parameter estimation and prediction using a computationally intensive model

Bayesian methods have been successful in quantifying uncertainty in physics-based problems in parameter estimation and prediction. In these cases, physical measurements y are modeled as the best fit of a physics-based model eta(theta), where theta denotes the uncertain, best input setting. Hence the statistical model is of the form y = eta(theta) + c, where epsilon accounts for measurement, and possibly other, error sources. When nonlinearity is present in eta(center dot), the resulting posterior distribution for the unknown parameters in the Bayesian formulation is typically complex and nonstandard, requiring computationally demanding computational approaches such as Markov chain Monte Carlo (MCMC) to produce multivariate draws from the posterior. Although generally applicable, MCMC requires thousands (or even millions) of evaluations of the physics model eta(center dot). This requirement is problematic if the model takes hours or days to evaluate. To overcome this computational bottleneck, we present an approach adapted from Bayesian model calibration. This approach combines output from an ensemble of computational model runs with physical measurements, within a statistical formulation, to carry out inference. A key component of this approach is a statistical response surface, or emulator, estimated from the ensemble of model runs. We demonstrate this approach with a case study in estimating parameters for a density functional theory model, using experimental mass/binding energy measurements from a collection of atomic nuclei. We also demonstrate how this approach produces uncertainties in predictions for recent mass measurements obtained at Argonne National Laboratory.

Higdon, Dave↗

Bayesian Redshift Classification of Emission-line Galaxies with Photometric Equivalent Widths

We present a Bayesian approach to the redshift classification of emission-line galaxies when only a single emission line is detected spectroscopically. We consider the case of surveys for high-redshift Lyα-emitting galaxies (LAEs), which have traditionally been classified via an inferred rest-frame equivalent width (EW; W{sub Lyα}) greater than 20 Å. Our Bayesian method relies on known prior probabilities in measured emission-line luminosity functions and EW distributions for the galaxy populations, and returns the probability that an object in question is an LAE given the characteristics observed. This approach will be directly relevant for the Hobby–Eberly Telescope Dark Energy Experiment (HETDEX), which seeks to classify ∼10{sup 6} emission-line galaxies into LAEs and low-redshift [O II] emitters. For a simulated HETDEX catalog with realistic measurement noise, our Bayesian method recovers 86% of LAEs missed by the traditional W{sub Lyα} > 20 Å cutoff over 2 < z < 3, outperforming the EW cut in both contamination and incompleteness. This is due to the method’s ability to trade off between the two types of binary classification error by adjusting the stringency of the probability requirement for classifying an observed object as an LAE. In our simulations of HETDEX, this method reduces the uncertainty in cosmological distance measurements by 14% with respect to the EW cut, equivalent to recovering 29% more cosmological information. Rather than using binary object labels, this method enables the use of classification probabilities in large-scale structure analyses. It can be applied to narrowband emission-line surveys as well as upcoming large spectroscopic surveys including Euclid and WFIRST.

79 ASTRONOMY AND ASTROPHYSICS↗

3D Sedimentological and geophysical studies of clastic reservoir analogs: Facies architecture, reservoir properties, and flow behavior within delta front facies elements of the Cretaceous Wall Creek Member, Frontier Formation, Wyoming

Significant volumes of oil and gas occur in reservoirs formed by ancient river deltas. This has implications for the spatial distribution of rock types and the variation of transport properties. A between mudstones and sandstones may form baffles that influence productivity and recovery efficiency. Diagenetic processes such as compaction, dissolution, and cementation can also alter flow properties. A better understanding of these properties and improved methods will allow improved reservoir development planning and increased recovery of oil and gas from deltaic reservoirs. Surface exposures of ancient deltaic rocks provide a high-resolution view of variability. Insights gleaned from these exposures can be used to model analogous reservoirs, for which data is sparser. The Frontier Formation in central Wyoming provides an opportunity for high-resolution models. The same rocks exposed in the Tisdale anticline are productive in nearby oil fields. Kilometers of exposure are accessible, and bedding-plane exposures allow use of high-resolution ground-penetrating radar. This study combined geologic interpretations, maps, vertical sections, core data, and ground-penetrating radar to construct geostatistical and flow models. Strata-conforming grids were use to reproduce the observed geometries. A new Bayesian method integrates outcrop, core, and radar amplitude and phase data. The proposed method propagates measurement uncertainty and yields an ensemble of plausible models for calcite concretions. These concretions affect flow significantly. Models which integrate more have different flow responses from simpler models, as demonstrated an exhaustive two-dimensional reference image and in three dimensions. This method is simple to implement within widely available geostatistics packages. Significant volumes of oil and gas occur in reservoirs that are inferred to have been formed by ancient river deltas. This geologic setting has implications for the spatial distribution of rock types (\Eg sandstones and mudstones) and the variation of transport properties (\Eg permeability and porosity) within bodies of a particular rock type. Both basin-wide processes such as sea-level change and the autocyclicity of deltaic processes commonly cause deltaic reservoirs to have large variability in rock properties; in particular, alternations between mudstones and sandstones may form baffles and trends in rock body permeability can influence productivity and recovery efficiency. In addition, diagenetic processes such as compaction, dissolution, and cementation can alter the spatial pattern of flow properties. A better understanding of these properties, and improved methods to model the properties and their effects, will allow improved reservoir development planning and increased recovery of oil and gas from deltaic reservoirs. Surface exposures of ancient deltaic rocks provide a high resolution, low uncertainty view of subsurface variability. Patterns and insights gleaned from these exposures can be used to model analogous reservoirs, for which data is much sparser. This approach is particularly attractive when reservoir formations are exposed at the surface. The Frontier Formation in central Wyoming provides an opportunity for high resolution characterization. The same rocks exposed in the vicinity of the Tisdale anticline are productive in nearby oil fields, including Salt Creek. Many kilometers of good-quality exposure are accessible, and the common bedding-plane exposures allow use of shallow-penetration, high-resolution electromagnetic methods known as ground-penetrating radar. This study combined geologic interpretations, maps, vertical sections, core data, and ground-penetrating radar to construct high-resolution geostatistical and flow models for the Wall Creek Member of the Frontier Formation. Stratal-conforming grids were use to reproduce the progradational and aggradational geometries observed in outcrop and radar data. A new, Bayesian method integrates outcrop--derived statistics, core observations of concretions, and radar amplitude and phase data. The proposed method consistently propagates measurement uncertainty through the model-building process, and yields an ensemble of plausible models for diagenetic calcite concretions. These concretions have a statistically significant on flow. Furthermore, neither geostatistical data from the outcrops nor geophysical data from radar is sufficient: models which integrate these data have significantly different flow responses. This was demonstrated both for an exhaustive two-dimensional reference image and in three dimensions, using flow simulations. This project wholly supported one PhD student and part of the education of an additional MS and PhD student. It helped to sponsor 6 refereed articles and 8 conference or similar presentations.

02 PETROLEUM↗

Estimation of the Dose and Dose Rate Effectiveness Factor

Current models to estimate radiation risk use the Life Span Study (LSS) cohort that received high doses and high dose rates of radiation. Transferring risks from these high dose rates to the low doses and dose rates received by astronauts in space is a source of uncertainty in our risk calculations. The solid cancer models recommended by BEIR VII [1], UNSCEAR [2], and Preston et al [3] is fitted adequately by a linear dose response model, which implies that low doses and dose rates would be estimated the same as high doses and dose rates. However animal and cell experiments imply there should be curvature in the dose response curve for tumor induction. Furthermore animal experiments that directly compare acute to chronic exposures show lower increases in tumor induction than acute exposures. A dose and dose rate effectiveness factor (DDREF) has been estimated and applied to transfer risks from the high doses and dose rates of the LSS cohort to low doses and dose rates such as from missions in space. The BEIR VII committee [1] combined DDREF estimates using the LSS cohort and animal experiments using Bayesian methods for their recommendation for a DDREF value of 1.5 with uncertainty. We reexamined the animal data considered by BEIR VII and included more animal data and human chromosome aberration data to improve the estimate for DDREF. Several experiments chosen by BEIR VII were deemed inappropriate for application to human risk models of solid cancer risk. Animal tumor experiments performed by Ullrich et al [4], Alpen et al [5], and Grahn et al [6] were analyzed to estimate the DDREF. Human chromosome aberration experiments performed on a sample of astronauts within NASA were also available to estimate the DDREF. The LSS cohort results reported by BEIR VII were combined with the new radiobiology results using Bayesian methods.

Chappell, L.↗

Convolutional neural networks on the HEALPix sphere: a pixel-based algorithm and its application to CMB data analysis

We describe a novel method for the application of convolutional neural networks (CNNs) to fields defined on the sphere, using the Hierarchical Equal Area Latitude Pixelization scheme (HEALPix). Specifically, we have developed a pixel-based approach to implement convolutional and pooling layers on the spherical surface, similarly to what is commonly done for CNNs applied to Euclidean space. The main advantage of our algorithm is to be fully integrable with existing, highly optimized libraries for NNs (e.g., PyTorch, TensorFlow, etc.). We present two applications of our method: (i) recognition of handwritten digits projected on the sphere; (ii) estimation of cosmological parameter from simulated maps of the cosmic microwave background (CMB). The latter represents the main target of this exploratory work, whose goal is to show the applicability of our CNN to CMB parameter estimation. We have built a simple NN architecture, consisting of four convolutional and pooling layers, and we have used it for all the applications explored herein. Concerning the recognition of handwritten digits, our CNN reaches an accuracy of ~95%, comparable with other existing spherical CNNs, and this is true regardless of the position and orientation of the image on the sphere. For CMB-related applications, we tested the CNN on the estimation of a mock cosmological parameter, defining the angular scale at which the power spectrum of a Gaussian field projected on the sphere peaks. We estimated the value of this parameter directly from simulated maps, in several cases: temperature and polarization maps, presence of white noise, and partially covered maps. For temperature maps, the NN performances are comparable with those from standard spectrum-based Bayesian methods. For polarization, CNNs perform about a factor four worse than standard algorithms. Nonetheless, our results demonstrate, for the first time, that CNNs are able to extract information from polarization fields, both in full-sky and masked maps, and to distinguish between E and B -modes in pixel space. Lastly, we have applied our CNN to the estimation of the Thomson scattering optical depth at reionization ( τ ) from simulated CMB maps. Even without any specific optimization of the NN architecture, we reach an accuracy comparable with standard Bayesian methods. This work represents a first step towards the exploitation of NNs in CMB parameter estimation and demonstrates the feasibility of our approach.

79 ASTRONOMY AND ASTROPHYSICS↗