SEARCH · Search NASA

Results for “default parameters”

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.

SourceResearch connectionFrom yearThrough year

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 163 records · Page 9

All sources NASA NTRS DOE OSTI

NASA NTRS2002

A Systems Model for Power Technology Assessment

A computer model is under continuing development at NASA Glenn Research Center that enables first-order assessments of space power technology. The model, an evolution of NASA Glenn's Array Design Assessment Model (ADAM), is an Excel workbook that consists of numerous spreadsheets containing power technology performance data and sizing algorithms. Underlying the model is a number of databases that contain default values for various power generation, energy storage and power management and distribution component parameters. These databases are actively maintained by a team of systems analysts so that they contain state-of-art data as well as the most recent technology performance projections. Sizing of the power subsystems can be accomplished either by using an assumed mass specific power (W/kg) or energy (Wh/kg) or by a bottoms-up calculation that accounts for individual component performance and masses. The power generation, energy storage and power management and distribution subsystems are sized for given mission requirements for a baseline case and up to three alternatives. This allows four different power systems to be sized and compared using consistent assumptions and sizing algorithms. The component sizing models contained in the workbook are modular so that they can be easily maintained and updated. All significant input values have default values loaded from the databases that can be over-written by the user. The default data and sizing algorithms for each of the power subsystems are described in some detail. The user interface and workbook navigational features are also discussed. Finally, an example study case that illustrates the model's capability is presented.

Hoffman, David J.↗

DOE OSTI2016

SU-F-T-681: Does the Biophysical Modeling for Immunological Aspects in Radiotherapy Precisely Predict Tumor and Normal Tissue Responses?

Purpose: Recent advances in immunotherapy make possible to combine with radiotherapy. The aim of this study was to assess the TCP/NTCP model with immunological aspects including stochastic distribution as intercellular uncertainties. Methods: In the clinical treatment planning system (Eclipse ver.11.0, Varian medical systems, US), biological parameters such as α/β, D50, γ, n, m, TD50 including repair parameters (bi-exponential repair) can be set as any given values to calculate the TCP/NTCP. Using a prostate cancer patient data with VMAT commissioned as a 6-MV photon beam of Novalis-Tx (BrainLab, US) in clinical use, the fraction schedule were hypothesized as 70–78Gy/35–39fr, 72–81Gy/40–45fr, 52.5–66Gy/16–22fr, 35–40Gy/5fr of 5–7 fractions in a week. By use of stochastic biological model applying for Gaussian distribution, the effects of the TCP/NTCP variation of repair parameters of the immune system as well as the intercellular uncertainty of tumor and normal tissues have been evaluated. Results: As respect to the difference of the α/β, the changes of the TCP/NTCP were increased in hypo-fraction regimens. The difference between the values of n and m affect the variation of the NTCP with the fraction schedules, independently. The elongation of repair half-time (long) increased the TCP/NTCP twice or much higher in the case of hypo-fraction scheme. For tumor, the repopulation parameters such as Tpot and Tstart, which is immunologically working to the tumor, improved TCP. Conclusion: Compared to default fixed value, which has affected by the probability of cell death and cure, hypo-fractionation schemes seemed to have advantages for the variations of the values of m. The possibility of an increase of the α/β or TD50 and repair parameters in tumor and normal tissue by immunological aspects were highly expected. For more precise prediction, treatment planning systems should be incorporated the complicated biological optimization in clinical practice combined with basic experiments data.

60 APPLIED LIFE SCIENCES↗

DOE OSTI2025

Improved Representations of Land‐Atmosphere Interactions Over the Continental U.S. Through Dynamic Root Modeling

Previous studies have identified the oversimplified root system representation as a key factor leading to inaccuracies in vegetation‐atmosphere feedbacks. In this study, a dynamic root water uptake scheme in the Noah‐MP land surface model has been coupled to the Weather Research and Forecasting (WRF) model to investigate its impact on the surface hydroclimate variables and land‐atmosphere interactions. To evaluate the impact of the dynamic root, two coupled simulations were conducted, one with the dynamic root water uptake scheme (DynRt) and one with the static root water uptake scheme (StcRt), which is based on the default root representation in Noah‐MP, with slight modifications, primarily in vegetation‐related parameters. Both DynRt and StcRt simulations were conducted with a small ensemble of three members to account for variations in physical parameterizations, initial and boundary forcing and model setup. When compared with reference data sets, the DynRt simulations show improved results than the StcRt simulations, reducing biases in the simulated leaf area index, surface energy fluxes, soil moisture and precipitation. Two different mechanisms through which roots affect land‐atmosphere coupling have been identified. Over the transitional climate zone between the dry and wet climate, the dynamic root scheme affects surface climate and land‐atmosphere coupling mainly through changes in soil moisture through hydraulic redistribution by plant root system. Over the energy‐limited mesic zone, the dynamic root affects regional land‐atmosphere coupling mainly through changes in carbon allocation. This work highlights the importance of dynamic root representation in improving vegetation‐atmosphere simulations by enhancing predictions of water, energy, and carbon fluxes.

54 ENVIRONMENTAL SCIENCES↗

NASA NTRS2015

Environment Modeling Using Runtime Values for JPF-Android

Software applications are developed to be executed in a specific environment. This environment includes external native libraries to add functionality to the application and drivers to fire the application execution. For testing and verification, the environment of an application is simplified abstracted using models or stubs. Empty stubs, returning default values, are simple to generate automatically, but they do not perform well when the application expects specific return values. Symbolic execution is used to find input parameters for drivers and return values for library stubs, but it struggles to detect the values of complex objects. In this work-in-progress paper, we explore an approach to generate drivers and stubs based on values collected during runtime instead of using default values. Entry-points and methods that need to be modeled are instrumented to log their parameters and return values. The instrumented applications are then executed using a driver and instrumented libraries. The values collected during runtime are used to generate driver and stub values on- the-fly that improve coverage during verification by enabling the execution of code that previously crashed or was missed. We are implementing this approach to improve the environment model of JPF-Android, our model checking and analysis tool for Android applications.

Verification↗

NASA NTRS2006

Intelligent Weather Agent

Method and system for automatically displaying, visually and/or audibly and/or by an audible alarm signal, relevant weather data for an identified aircraft pilot, when each of a selected subset of measured or estimated aviation situation parameters, corresponding to a given aviation situation, has a value lying in a selected range. Each range for a particular pilot may be a default range, may be entered by the pilot and/or may be automatically determined from experience and may be subsequently edited by the pilot to change a range and to add or delete parameters describing a situation for which a display should be provided. The pilot can also verbally activate an audible display or visual display of selected information by verbal entry of a first command or a second command, respectively, that specifies the information required.

Spirkovska, Liljana↗

DOE OSTI2013

Evaluation of a Modified Scheme for Shallow Convection: Implementation of CuP and Case Studies

A new treatment for shallow clouds has been introduced into the Weather Research and Forecasting (WRF) model. The new scheme, called the cumulus potential (CuP) scheme, replaces the ad-hoc trigger function used in the Kain-Fritsch cumulus parameterization with a trigger function related to the distribution of temperature and humidity in the convective boundary layer via probability density functions (PDFs). An additional modification to the default version of WRF is the computation of a cumulus cloud fraction based on the time scales relevant for shallow cumuli. Results from three case studies over the U.S. Department of Energy’s Atmospheric Radiation Measurement (ARM) site in north central Oklahoma are presented. These days were selected because of the presence of shallow cumuli over the ARM site. The modified version of WRF does a much better job predicting the cloud fraction and the downwelling shortwave irradiance thancontrol simulations utilizing the default Kain-Fritsch scheme. The modified scheme includes a number of additional free parameters, including the number and size of bins used to define the PDF, the minimum frequency of a bin within the PDF before that bin is considered for shallow clouds to form, and the critical cumulative frequency of bins required to trigger deep convection. A series of tests were undertaken to evaluate the sensitivity of the simulations to these parameters. Overall, the scheme was found to be relatively insensitive to each of the parameters.

Berg, Larry K↗

DOE OSTI2019

Impact of cross section uncertainties on NOvA oscillation analyses

The NOvA experiment is an off-axis long-baseline neutrino oscillation experiment measuring $\nu_{\mu}$ disappearance and $\nu_{e}$ appearance in a $\nu_{\mu}$ beam originating at Fermilab. Oscillations are observed in the far detector relative to measurements made in the near detector, with a full simulation of the beam line and detectors being used to perform extrapolation between them. The neutrino simulation uses the GENIE event generator, which contains implementations of numerous theoretical neutrino interaction models applied to a variety of nuclear targets. However, recent data, recent reanalysis of extant data, and continued development of theoretical models have brought to light deficiencies in the default GENIE cross section model, which informs the predicted spectra used to infer oscillation parameters. We explore how uncertainties in this model, together with modifications to GENIE version 2.12.2 based on external information and NOvA Near Detector (ND) data, affect NOvA's oscillation parameter inferences. We also discuss how these uncertainties can be mitigated by judicious analysis design, including usage of the ND data.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

DOE OSTI2014

A Sensitivity Study on Modeling Black Carbon in Snow and its Radiative Forcing over the Arctic and Northern China

Black carbon in snow (BCS) simulated in the Community Atmosphere Model (CAM5) is evaluated against measurements over Northern China and the Arctic, and its sensitivity to atmospheric deposition and two parameters that affect post-depositional enrichment is explored. The BCS concentration is overestimated (underestimated) by a factor of two in Northern China (Arctic) in the default model, but agreement with observations is good over both regions in the simulation with improvements in BC transport and deposition. Sensitivity studies indicate that uncertainty in the melt-water scavenging efficiency (MSE) parameter substantially affects BCS and its radiative forcing (by a factor of 2-7) in the Arctic through post-depositional enrichment. The MSE parameter has a relatively small effect on the magnitude of BCS seasonal cycle but can alter its phase in Northern China. The impact of the snow aging scaling factor (SAF) on BCS, partly through the post-depositional enrichment effect, shows more complex latitudinal and seasonal dependence. Similar to MSE, SAF affects more significantly the magnitude (phase) of BCS season cycle over the Arctic (Northern China). While uncertainty associated with the representation of BC transport and deposition processes in CAM5 is more important than that associated with the two snow model parameters in Northern China, the two uncertainties have comparable effect in the Arctic.

Artic↗

NASA NTRS2016

Implementing Connected Component Labeling as a User Defined Operator for SciDB

We have implemented a flexible User Defined Operator (UDO) for labeling connected components of a binary mask expressed as an array in SciDB, a parallel distributed database management system based on the array data model. This UDO is able to process very large multidimensional arrays by exploiting SciDB's memory management mechanism that efficiently manipulates arrays whose memory requirements far exceed available physical memory. The UDO takes as primary inputs a binary mask array and a binary stencil array that specifies the connectivity of a given cell to its neighbors. The UDO returns an array of the same shape as the input mask array with each foreground cell containing the label of the component it belongs to. By default, dimensions are treated as non-periodic, but the UDO also accepts optional input parameters to specify periodicity in any of the array dimensions. The UDO requires four stages to completely label connected components. In the first stage, labels are computed for each subarray or chunk of the mask array in parallel across SciDB instances using the weighted quick union (WQU) with half-path compression algorithm. In the second stage, labels around chunk boundaries from the first stage are stored in a temporary SciDB array that is then replicated across all SciDB instances. Equivalences are resolved by again applying the WQU algorithm to these boundary labels. In the third stage, relabeling is done for each chunk using the resolved equivalences. In the fourth stage, the resolved labels, which so far are "flattened" coordinates of the original binary mask array, are renamed with sequential integers for legibility. The UDO is demonstrated on a 3-D mask of O(1011) elements, with O(108) foreground cells and O(106) connected components. The operator completes in 19 minutes using 84 SciDB instances.

UDO↗

DOE OSTI2021

Investigating high-energy proton-induced reactions on spherical nuclei: Implications for the preequilibrium exciton model

A number of accelerator-based isotope production facilities utilize $100-200$ MeV proton beams due to the high production rates enabled by high-intensity beam capabilities and the greater diversity of isotope production brought on by the long range of high-energy protons. However, nuclear reaction modeling at these energies can be challenging because of the interplay between different reaction modes and a lack of existing guiding cross section data. A Tri-lab collaboration has been formed between the Lawrence Berkeley, Los Alamos, and Brookhaven National Laboratories to address these complexities by characterizing charged-particle nuclear reactions relevant to the production of established and novel radioisotopes. In the inaugural collaboration experiments, stacked-targets of niobium foils were irradiated at the Brookhaven Linac Isotope Producer ($E_p=200$ MeV) and the Los Alamos Isotope Production Facility ($E_p=100$ MeV) to measure $^{93}$Nb(p,x) cross sections between $50-200$ MeV. The results were compared with literature data as well as the default calculations of the nuclear model codes TALYS, CoH, EMPIRE, and ALICE. The default code predictions largely failed to reproduce the measurements. Therefore, we developed a standardized procedure, which determines the reaction model parameters that best reproduce the most prominent reaction channels in a physically justifiable manner. Overall, the primary focus of the procedure was to determine the best parameterization for the pre-equilibrium two-component exciton model. This modeling study revealed a trend towards a relative decrease for internal transition rates at intermediate proton energies ($E_p=20-60$ MeV) in the current exciton model as compared to the default values. The results of this work are instrumental for the planning, execution, and analysis essential to isotope production.

43 PARTICLE ACCELERATORS↗

DOE OSTI2015

Sparse maps—A systematic infrastructure for reduced-scaling electronic structure methods. I. An efficient and simple linear scaling local MP2 method that uses an intermediate basis of pair natural orbitals

In this work, a systematic infrastructure is described that formalizes concepts implicit in previous work and greatly simplifies computer implementation of reduced-scaling electronic structure methods. The key concept is sparse representation of tensors using chains of sparse maps between two index sets. Sparse map representation can be viewed as a generalization of compressed sparse row, a common representation of a sparse matrix, to tensor data. By combining few elementary operations on sparse maps (inversion, chaining, intersection, etc.), complex algorithms can be developed, illustrated here by a linear-scaling transformation of three-center Coulomb integrals based on our compact code library that implements sparse maps and operations on them. The sparsity of the three-center integrals arises from spatial locality of the basis functions and domain density fitting approximation. A novel feature of our approach is the use of differential overlap integrals computed in linear-scaling fashion for screening products of basis functions. Finally, a robust linear scaling domain based local pair natural orbital second-order Möller-Plesset (DLPNO-MP2) method is described based on the sparse map infrastructure that only depends on a minimal number of cutoff parameters that can be systematically tightened to approach 100% of the canonical MP2 correlation energy. With default truncation thresholds, DLPNO-MP2 recovers more than 99.9% of the canonical resolution of the identity MP2 (RI-MP2) energy while still showing a very early crossover with respect to the computational effort. Based on extensive benchmark calculations, relative energies are reproduced with an error of typically <0.2 kcal/mol. The efficiency of the local MP2 (LMP2) method can be drastically improved by carrying out the LMP2 iterations in a basis of pair natural orbitals. While the present work focuses on local electron correlation, it is of much broader applicability to computation with sparse tensors in quantum chemistry and beyond.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

NASA NTRS2008

SiC JFET Transistor Circuit Model for Extreme Temperature Range

A technique for simulating extreme-temperature operation of integrated circuits that incorporate silicon carbide (SiC) junction field-effect transistors (JFETs) has been developed. The technique involves modification of NGSPICE, which is an open-source version of the popular Simulation Program with Integrated Circuit Emphasis (SPICE) general-purpose analog-integrated-circuit-simulating software. NGSPICE in its unmodified form is used for simulating and designing circuits made from silicon-based transistors that operate at or near room temperature. Two rapid modifications of NGSPICE source code enable SiC JFETs to be simulated to 500 C using the well-known Level 1 model for silicon metal oxide semiconductor field-effect transistors (MOSFETs). First, the default value of the MOSFET surface potential must be changed. In the unmodified source code, this parameter has a value of 0.6, which corresponds to slightly more than half the bandgap of silicon. In NGSPICE modified to simulate SiC JFETs, this parameter is changed to a value of 1.6, corresponding to slightly more than half the bandgap of SiC. The second modification consists of changing the temperature dependence of MOSFET transconductance and saturation parameters. The unmodified NGSPICE source code implements a T(sup -1.5) temperature dependence for these parameters. In order to mimic the temperature behavior of experimental SiC JFETs, a T(sup -1.3) temperature dependence must be implemented in the NGSPICE source code. Following these two simple modifications, the Level 1 MOSFET model of the NGSPICE circuit simulation program reasonably approximates the measured high-temperature behavior of experimental SiC JFETs properly operated with zero or reverse bias applied to the gate terminal. Modification of additional silicon parameters in the NGSPICE source code was not necessary to model experimental SiC JFET current-voltage performance across the entire temperature range from 25 to 500 C.

Neudeck, Philip G.↗

DOE OSTI2023

Economic Evaluation of a Coupled Nuclear Power Plant and Hydrogen Production Facility: A Case Study

This study optimized the design sizes and operation of a power-to-hydrogen-to-power integrated energy system to allow a baseload power plant to operate flexibly in the energy market. In collaboration with a utility industry partner, the system, consisting of an electrolyzer, compressors, storage tank, and fuel cell, was optimized under conditions specific to the proposed project at the site of a nuclear power plant. The Design Integration and Synthesis Platform to Advance Tightly Coupled Hybrid Energy Systems (DISPATCHES) maximized net present value by optimizing sizing of components and dispatch decisions. Revenues included sale of electricity, capacity payments typical of the New York Independent System Operator, and the section 45V hydrogen production tax credit of the Inflation Reduction Act of 2022 (the tax credit was assumed to be available to legacy plants in the absence of clear guidance at present). Under default assumptions which excluded many capital expenditures, the base case optimized solution had a net present value of $\$$1.4 million over a 30 year lifetime, with a 0.365 MW fuel cell operating nearly continuously and 85% of revenues supplied by the hydrogen production tax credit (which was counted as a revenue regardless of profit, thus assuming credit monetization or offset of taxes within the larger firm was possible in all years). Beyond the base case, a sensitivity study elucidated drivers of the economics as capacity payment rate and hydrogen production tax credit rate vary. Additional sensitivity studies also extended results to variation of other, previously fixed parameters, including the fuel cell capital cost, and to imposition of further constraints. Optimization was also repeated for the default assumptions but recognizing tax credits upon use of hydrogen rather than upon its production, producing no change in the optimal solution. Most notably, capacity payments above $\$$15/kW-month drove optimal fuel cells multiple times larger than those with the default estimated capacity payment of $\$$2.5/kW-month (approaching 11 vs. 0.365 MW), and these larger fuel cells operated rarely (capacity factors of ~0.03). Furthermore, when the hydrogen production tax credit was provided for only 10 years, under the specific assumptions of this study (e.g., neither site preparation costs nor electrolyzer capital cost counted), the optimal solution avoided economic loss by ceasing system operation after the 10th year. Viewed broadly, this study demonstrated the capabilities of DISPATCHES, which can be user-adapted to serve other industrial case studies.

08 HYDROGEN↗

NASA NTRS2022

Constraining Clouds and Convective Parameterizations in A Climate Model Using Paleoclimate Data

Cloud and convective parameterizations strongly influence uncertainties in equilibrium climate sensitivity. We provide a proof-of-concept study to constrain these parameterizations in a perturbed parameter ensemble of the atmosphere-only version of the Goddard Institute for Space Studies Model E2.1 simulations by evaluating model biases in the present-day runs using multiple satellite climatologies and by comparing simulated δ 18 O of precipitation (δ 18 O p ), known to be sensitive to parameterization schemes, with a global database of speleothem δ 18 O records covering the Last Glacial Maximum (LGM), mid-Holocene (MH) and pre-industrial (PI) periods. Relative to modern interannual variability, paleoclimate simulations show greater sensitivity to parameter changes, allowing for an evaluation of model uncertainties over a broader range of climate forcing and the identification of parts of the world that are parameter sensitive. Certain simulations reproduced absolute δ 18 O p values across all time periods, along with LGM and MH δ 18 O p anomalies relative to the PI, better than the default parameterization. No single set of parameterizations worked well in all climate states, likely due to the non-stationarity of cloud feedbacks under varying boundary conditions. Future work that involves varying multiple parameter sets simultaneously with coupled ocean feedbacks will likely provide improved constraints on cloud and convective parameterizations.

cloud and convective parameterization↗

DOE OSTI2014

Precipitation characteristics of CAM5 physics at mesoscale resolution during MC3E and the impact of convective timescale choice

Abstract The physics suite of the Community Atmosphere Model version 5 (CAM5) has recently been implemented in the Weather Research and Forecasting (WRF) model to explore the behavior of the parameterization suite at high resolution and within the controlled setting of a limited area model. The initial paper documenting this capability characterized the behavior for a northern high‐latitude region. This paper characterizes the precipitation characteristics for continental, midlatitude, springtime conditions during the Midlatitude Continental Convective Clouds Experiment (MC3E) over the central United States. This period exhibited a range of convective conditions from those driven strongly by large‐scale synoptic regimes to more locally driven convection. The study focuses on the precipitation behavior at 32 km grid spacing to better anticipate how the physics will behave in a global model when used at similar grid spacing in the coming years. Importantly, one change to the Zhang‐McFarlane deep convective parameterization when implemented in WRF was to make the convective timescale parameter an explicit function of grid spacing. This study examines the sensitivity of the precipitation to the default value of the convective timescale in WRF, which is 600 s for 32 km grid spacing, to the value of 3600 s used for 2° grid spacing in CAM5. For comparison, a 1200 s and an infinite convective timescale are also used. The results show that the 600 s timescale gives the most accurate precipitation amount over the central United States. However, this setting has the worst precipitation diurnal cycle, with the convection too tightly linked to the daytime surface heating. Longer timescales greatly improve the diurnal cycle but result in less precipitation and produce a low bias. An analysis of rain rates shows the accurate precipitation amount with the shorter timescale is assembled from an over abundance of drizzle combined with too few heavy rain events. With longer timescales, one can improve the frequency distribution, particularly for the extreme rain rates. Ultimately, without changing other aspects of the physics, one must decide between accurate diurnal timing and rain amount when choosing an appropriate convective timescale.

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2023

CMS PYTHIA 8 colour reconnection tunes based on underlying-event data

New sets of parameter tunes for two of the colour reconnection models, quantum chromodynamics-inspired and gluon-move, implemented in the PYTHIA 8 event generator, are obtained based on the default CMS PYTHIA 8 underlying-event tune, CP5. Measurements sensitive to the underlying event performed by the CMS experiment at centre-of-mass energies $\sqrt{s}=$7 and 13 TeV, and by the CDF experiment at 1.96 TeV are used to constrain the parameters of colour reconnection models and multiple-parton interactions simultaneously. The new colour reconnection tunes are compared with various measurements at 1.96, 7, 8, and 13 TeV including measurements of the underlying-event, strange-particle multiplicities, jet substructure observables, jet shapes, and colour flow in top quark pair ($t\bar{t}$) events. The new tunes are also used to estimate the uncertainty related to colour reconnection modelling in the top quark mass measurement using the decay products of $t\bar{t}$ events in the semileptonic channel at 13 TeV .

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

DOE OSTI2013

Invited Article: Coherent imaging using seeded free-electron laser pulses with variable polarization: First results and research opportunities

FERMI-Elettra, the first vacuum ultraviolet and soft X-ray free-electron laser (FEL) using by default a 'seeded' scheme, became operational in 2011 and has been opened to users since December 2012. The parameters of the seeded FERMI FEL pulses and, in particular, the superior control of emitted radiation in terms of spectral purity and stability meet the stringent requirements for single-shot and resonant coherent diffraction imaging (CDI) experiments. The advantages of the intense seeded FERMI pulses with variable polarization have been demonstrated with the first experiments performed using the multipurpose experimental station operated at the diffraction and projection imaging (DiProI) beamline. The results reported here were obtained with fixed non-periodic targets during the commissioning period in 2012 using 20-32 nm wavelength range. They demonstrate that the performance of the FERMI FEL source and the experimental station meets the requirements of CDI, holography, and resonant magnetic scattering in both multi- and single-shot modes. Moreover, we present the first magnetic scattering experiments employing the fully circularly polarized FERMI pulses. The ongoing developments aim at pushing the lateral resolution by using shorter wavelengths provided by double-stage cascaded FERMI FEL-2 and probing ultrafast dynamic processes using different pump-probe schemes, including jitter-free seed laser pump or FEL-pump/FEL-probe with two color FEL pulses generated by the same electron bunch.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

DOE OSTI2017

Adaptive Neuron Apoptosis for Accelerating Deep Learning on Large Scale Systems

Machine Learning and Data Mining (MLDM) algorithms are becoming ubiquitous in {\em model learning} from the large volume of data generated using simulations, experiments and handheld devices. Deep Learning algorithms -- a class of MLDM algorithms -- are applied for automatic feature extraction, and learning non-linear models for unsupervised and supervised algorithms. Naturally, several libraries which support large scale Deep Learning -- such as TensorFlow and Caffe -- have become popular. In this paper, we present novel techniques to accelerate the convergence of Deep Learning algorithms by conducting low overhead removal of redundant neurons -- {\em apoptosis} of neurons -- which do not contribute to model learning, during the training phase itself. We provide in-depth theoretical underpinnings of our heuristics (bounding accuracy loss and handling apoptosis of several neuron types), and present the methods to conduct adaptive neuron apoptosis. We implement our proposed heuristics with the recently introduced TensorFlow and using its recently proposed extension with MPI. Our performance evaluation on two difference clusters -- one connected with Intel Haswell multi-core systems, and other with nVIDIA GPUs -- using InfiniBand, indicates the efficacy of the proposed heuristics and implementations. Specifically, we are able to improve the training time for several datasets by 2-3x, while reducing the number of parameters by 30x (4-5x on average) on datasets such as ImageNet classification. For the Higgs Boson dataset, our implementation improves the accuracy (measured by Area Under Curve (AUC)) for classification from 0.88/1 to 0.94/1, while reducing the number of parameters by 3x in comparison to existing literature, while achieving a 2.44x speedup in comparison to the default (no apoptosis) algorithm.

Siegel, Charles M.↗