Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “variational analysis”

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 577 records · Page 32
All sourcesNASA NTRSDOE OSTI
DOE OSTI2021

Quantifying sensitivity in numerical weather prediction-modeled offshore wind speeds through an ensemble modeling approach

A decade of research has shown that numerical weather prediction (NWP)-modeled wind speeds can be highly sensitive to the inputs and setups within the NWP model. For wind resource characterization applications, this sensitivity is often addressed by constructing a range of setups and selecting the one that best validates against observations. However, this approach is not possible in areas that lack high-quality hub height observations, especially offshore wind areas. In such cases, techniques to quantify and disseminate confidence in NWP-modeled wind speeds in the absence of observations are needed. We address this need in the present study and propose best practices for quantifying the spread in NWP-modeled wind speeds. We implement an ensemble approach in which we consider 24 different setups to the Weather Research and Forecasting (WRF) model. We construct the ensemble by considering variations in WRF version, WRF namelist, atmospheric forcing, and sea surface temperature (SST) forcing. Our analysis finds that the standard deviation produces more consistent estimates compared to the interquartile range and tends to be the more conservative estimator for ensemble variability. We further find that model spread increases closer to the surface and on shorter time scales. In conclusion, we explore methods to attribute total ensemble variability to the different ensemble components (e.g., atmospheric forcing and SST product) and find that contributions by components also vary depending on time scale. We anticipate that the methods and results presented in this paper will provide a reasonable foundation for further research into ensemble-based wind resource data sets.

17 WIND ENERGY↗
DOE OSTI2023

Cost‐benefit assessment framework for robotics‐driven inspection of floating offshore wind farms

Abstract Operations and maintenance (O&M) of floating offshore wind farms (FOWFs) poses various challenges in terms of greater distances from the shore, harsher weather conditions, and restricted mobility options. Robotic systems have the potential to automate some parts of the O&M leading to continuous feature‐rich data acquisition, operational efficiency, along with health and safety improvements. There remains a gap in assessing the techno‐economic feasibility of robotics in the FOWF sector. This paper investigates the costs and benefits of incorporating robotics into the O&M of a FOWF. A bottom‐up cost model is used to estimate the costs for a proposed multi‐robot platform (MRP). The MRP houses unmanned aerial vehicle (UAV) and remotely operated vehicle (ROV) to conduct the inspection of specific FOWF components. Emphasis is laid on the most conducive O&M activities for robotization and the associated technical and cost aspects. The simulation is conducted in Windfarm Operations and Maintenance cost‐Benefit Analysis Tool (WOMBAT), where the metrics of incurred operational expenditure (OPEX) and the inspection time are calculated and compared with those of a baseline case consisting of crew transfer vessels, rope‐access technicians, and divers. Results show that the MRP can reduce the inspection time incurred, but this reduction has dependency on the efficacy of the robotic system and the associated parameterization e.g., cost elements and the inspection rates. Conversely, the increased MRP day rate results in a higher annualized OPEX. Residual risk is calculated to assess the net benefit of incorporating the MRP. Furthermore, sensitivity analysis is conducted to find the key parameters influencing the OPEX and the inspection time variation. A key output of this work is a robust and realistic framework which can be used for the cost‐benefit assessment of future MRP systems for specific FOWF activities.

17 WIND ENERGY↗
DOE OSTI2025

Impact of U-10Mo HALEU Fuel Element Tolerances on the Massachusetts Institute of Technology Reactor safety and operational performance – Neutronics

The U.S. is coordinating efforts for the conversion of six U.S. high performance research reactors (USHPRR), including one critical assembly from highly enriched uranium (HEU) to low-enriched uranium (LEU). In order to continue the mission of these reactors, including the Massachusetts Institute of Technology Reactor (MITR), and achieve similar performance, high assay low-enriched uranium (HALEU) with a high-density metallic alloy of uranium with 10 wt% molybdenum (U-10Mo) is being considered. Following the preliminary design of the proposed MITR LEU fuel elements using the U-10Mo monolithic alloy, the impact of the fabrication specification was assessed. This work focuses on the analysis of select neutronics characteristics of the MITR LEU core as a function of the variation of the relevant fuel specification parameters (e.g., U-10Mo composition, fuel plate thickness, etc.). A separate article submitted to this journal addresses the impact on the thermal hydraulic performance. The analyses in these works are performed based on an all-LEU conversion management plan identified in previous work, in which only the proposed elements are utilized for achieving the conversion of MITR. The variations of two main neutronics characteristics are assessed as a function of the variability of the specification parameters resulting from the fabrication process: the MITR LEU core reactivity and the fuel cycle length. The main findings of this work show that the MITR core can meet the operational requirements during the LEU transition plan under the limiting fabrication parameter combinations considered. In addition, the analyses show that the dependency of the core neutronics characteristics on the specification parameters is highly linear within the specification tolerances. The rates of variation are reported in detail for each parameter and can serve as a powerful tool for future MITR fuel management in cases such as when HALEU supply is established that may allow additional cycle length or other operational benefits.

Conversion↗
DOE OSTI2026

Pore-Scale Transport Effects in Electrochemical CO 2 Reduction on Gold via Coupled Microkinetic-Transport Modeling

A pore-resolved modeling framework is developed to quantify how pore-scale transport affects the intrinsic microkinetics of CO 2 -to-CO on Au. A DFT-informed microkinetic model is coupled self-consistently to a Generalized-Modified Poisson–Nernst–Planck (GMPNP) transport description in a single, electrolyte-filled cylindrical pore, allowing local concentrations and potential to feed back into site-specific reaction rates. FIB-SEM is used to determine pore sizes within realistic electrode materials. Across pore diameters, d p = 10–6000 nm, the surface-averaged CO 2 reduction rate is systematically reduced relative to the ideal microkinetic baseline where mass transport is not accounted for; the effectiveness factor 𝜂 𝑠,CO 2 , which quantifies this ratio, decreases rapidly at more negative potentials and is about 1% near −1.0 V vs SHE due to reactant depletion. Spatial maps reveal pore-bulk alkalization that emerges at higher cathodic bias, with a small, near-wall pH dip due to electrostatic repulsion of hydroxide at the cathode interface. For a fixed aspect ratio L p /d p , narrower pores exhibit larger 𝜂 𝑠,CO 2 by shortening diffusion paths, whereas variations in the aspect ratio L p /d p play a secondary role. A dimensionless analysis (surface/bulk Damköhler numbers) delineates operating regimes. In conclusion, this work offers a concept for incorporating microkinetic models into homogenized porous-electrode models through effectiveness factors and pore-size distribution.

Au-catalyst↗
DOE OSTI2021

A study of real-world micrograph data quality and machine learning model robustness

Abstract Machine-learning (ML) techniques hold the potential of enabling efficient quantitative micrograph analysis, but the robustness of ML models with respect to real-world micrograph quality variations has not been carefully evaluated. We collected thousands of scanning electron microscopy (SEM) micrographs for molecular solid materials, in which image pixel intensities vary due to both the microstructure content and microscope instrument conditions. We then built ML models to predict the ultimate compressive strength (UCS) of consolidated molecular solids, by encoding micrographs with different image feature descriptors and training a random forest regressor, and by training an end-to-end deep-learning (DL) model. Results show that instrument-induced pixel intensity signals can affect ML model predictions in a consistently negative way. As a remedy, we explored intensity normalization techniques. It is seen that intensity normalization helps to improve micrograph data quality and ML model robustness, but microscope-induced intensity variations can be difficult to eliminate.

36 MATERIALS SCIENCE↗
DOE OSTI2026

Large Language Model for Validation, Optical Calibration, and Learning (VOCAL) Distributed Temperature Sensing Interface

Distributed temperature sensing (DTS) using fiber optic sensors (FOS) offers a promising method for temperature measurements in advanced reactors, such as sodium fast reactors and molten salt cooled reactors. To support the calibration and validation of DTS measurements, Argonne National Laboratory developed the Validation, Optical Calibration, and Learning (VOCAL) software package. This report describes the integration of a local large language model (LLM) with a retrieval-augmented generation (RAG) system into the VOCAL interface to serve as an interactive user assistant. The LLM framework enhances the VOCAL platform’s accessibility to users by explaining interface components, clarifying inputs and outputs, and answering user queries dynamically in real-time. The accuracy of the LLM assistant performance was evaluated with 20 queries regarding the interface and its parameters using experimental data from the Thermal Hydraulic Experimental Test Article (THETA) facility. Results demonstrate that the LLM achieved a 95% accuracy rate, with a BERTScore of 0.8816 and SBERT value of 0.7417. Furthermore, validation of the RAG system within the LLM framework showed optimal accuracy with k-values between 1 and 2 using the k-refinement convergence test. The prompt perturbation analysis demonstrated good initial consistency for the RAG system, exhibiting the highest accuracy under punctuation variations and the greatest sensitivity under query reordering. Notably, the model’s errors were limited to data retrieval failures rather than factual hallucinations, reinforcing its baseline reliability. The integration of LLM provides a highly accurate, userfriendly enhancement to the VOCAL platform without disrupting its core computational capabilities for FOS calibration and validation.

Hong, Evan↗
DOE OSTI2023

Can strong substorm-associated MeV electron injections be an important cause of large radiation belt enhancements?

It has become well-established that strong outer radiation belt enhancements are due to wave-driven electron energization by whistler-mode chorus waves. However, in this study, we examine strong MeV electron injections on 10 July 2019 and find substantial evidence that such injections may be a crucial contributor to outer radiation belt enhancement events. For such an examination, it is essential to precisely separate temporal flux changes from spatial variations observed as Van Allen Probes move along their orbits. Employing a new “hourly snapshot” analysis approach, we discover unprecedented details of electron flux evolutions that suggest that for this event, the outer belt enhancement was not continuous but instead intermittent, mostly composed of 4 large discrete injection-driven flux increases. The injections appear as sharp flux increases when observed near apogee. Otherwise, by comparing hourly snapshots for different times, we infer injections and infer temporally stable fluxes between injections, despite strong and continuous chorus emission. The fast and intermittent electron flux growth successively extending earthwards implies cumulative outer belt enhancement via a series of repetitive inward transport associated with injection-induced electric fields.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2024

The Function of Horn Ridges for Impact Damping

This study explores the damping effects of ram horn ridges on mechanical impacts resulting from ramming. We measured the amplitudes and frequencies of ridges along the axial (pitch) direction of the ridges of ram horns obtained from eight specimens across six different species. While the horns shared a similar spiral-shaped pattern with surface ridges, our findings show variations among the horns, including ridge spacing and growth trends. Additionally, we employed finite element analysis (FEA) to compare a ridged horn model with a non-ridged counterpart to provide an understanding of the damping characteristics of the surface ridges. Our FEA results reveal that the ridged horn decreased the initial ramming pressure by 20.7%, increased the shear stress by 66.9%, and decreased the axial strain by 27.3%, the radial strain by 16.7%, and the shear strain by 14.3% at a 50 ms impact duration compared to those of the non-ridged horn. The damping ratio was increased by 7.9% because of the ridges. This study elucidates three primary functions of the different species of ram horns’ spirals and ridges: (1) to transfer longitudinal waves into shear waves, (2) to filter shear waves, and (3) to stabilize the structure by mitigating excessive strain.

bio-inspired design↗
DOE OSTI2020

In Situ and Post-Mortem Characterizations of Ultrasonic Spot Welded AZ31B and Coated Dual Phase 590 Steel Joints

Ultrasonic spot welding using different welding conditions was applied to join dissimilar metals of galvanized DP590 steel and AZ31B magnesium sheets. In situ high-speed imaging, digital image correlation, and infrared thermography were utilized to quantitatively study the interfacial relative motion, surface indentation, and heat generation across the joint faying interface and the sheet/sonotrode interfaces under the welding condition of moderate welding power and short welding time. For welds made with high power and long welding time, lap shear tensile tests as well as fatigue tests were carried out. Different fracture modes were observed after the lap shear tensile tests and fatigue tests performed under different peak loads. Post-weld cross-sectional analysis with scanning electron microscopy coupled with energy dispersive X-Ray spectroscopy revealed the variation of morphology and chemical composition at the joint interface for welds made with different welding conditions.

36 MATERIALS SCIENCE↗
DOE OSTI2022

Environmental effects on aerosol–cloud interaction in non-precipitating marine boundary layer (MBL) clouds over the eastern North Atlantic

Abstract. Over the eastern North Atlantic (ENA) ocean, a total of 20 non-precipitating single-layer marine boundary layer (MBL) stratus and stratocumulus cloud cases are selected to investigate the impacts of the environmental variables on the aerosol–cloud interaction (ACIr) using the ground-based measurements from the Department of Energy Atmospheric Radiation Measurement (ARM) facility at the ENA site during 2016–2018. The ACIr represents the relative change in cloud droplet effective radius re with respect to the relative change in cloud condensation nuclei (CCN) number concentration at 0.2 % supersaturation (NCCN,0.2 %) in the stratified water vapor environment. The ACIr values vary from −0.01 to 0.22 with increasing sub-cloud boundary layer precipitable water vapor (PWVBL) conditions, indicating that re is more sensitive to the CCN loading under sufficient water vapor supply, owing to the combined effect of enhanced condensational growth and coalescence processes associated with higher Nc and PWVBL. The principal component analysis shows that the most pronounced pattern during the selected cases is the co-variations in the MBL conditions characterized by the vertical component of turbulence kinetic energy (TKEw), the decoupling index (Di), and PWVBL. The environmental effects on ACIr emerge after the data are stratified into different TKEw regimes. The ACIr values, under both lower and higher PWVBL conditions, more than double from the low-TKEw to high-TKEw regime. This can be explained by the fact that stronger boundary layer turbulence maintains a well-mixed MBL, strengthening the connection between cloud microphysical properties and the below-cloud CCN and moisture sources. With sufficient water vapor and low CCN loading, the active coalescence process broadens the cloud droplet size spectra and consequently results in an enlargement of re. The enhanced activation of CCN and the cloud droplet condensational growth induced by the higher below-cloud CCN loading can effectively decrease re, which jointly presents as the increased ACIr. This study examines the importance of environmental effects on the ACIr assessments and provides observational constraints to future model evaluations of aerosol–cloud interactions.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2021

Machine learning deciphers CO 2 sequestration and subsurface flowpaths from stream chemistry

Endmember mixing analysis (EMMA) is often used by hydrogeochemists to interpret the sources of stream solutes, but variations in stream concentrations and discharges remain difficult to explain. We discovered that machine learning can be used to highlight patterns in stream chemistry that reveal information about sources of solutes and subsurface groundwater flowpaths. The investigation has implications, in turn, for the balance of CO 2 in the atmosphere. For example, CO 2 -driven weathering of silicate minerals removes carbon from the atmosphere over ~106-year timescales. Weathering of another common mineral, pyrite, releases sulfuric acid that in turn causes dissolution of carbonates. In that process, however, CO 2 is released instead of sequestered from the atmosphere. Thus, understanding long-term global CO 2 sequestration by weathering requires quantification of CO 2 - versus H 2 SO 4 -driven reactions. Most researchers estimate such weathering fluxes from stream chemistry, but interpreting the reactant minerals and acids dissolved in streams has been fraught with difficulty. We apply a machine-learning technique to EMMA in three watersheds to determine the extent of mineral dissolution by each acid, without pre-defining the endmembers. The results show that the watersheds continuously or intermittently sequester CO 2 , but the extent of CO 2 drawdown is diminished in areas heavily affected by acid rain. Prior to applying the new algorithm, CO 2 drawdown was overestimated. The new technique, which elucidates the importance of different subsurface flowpaths and long-timescale changes in the watersheds, should have utility as a new EMMA for investigating water resources worldwide.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2021

Search for CPT and Lorentz invariance violation in the muon $g -2$ experiment at Fermilab

The Muon g-2 experiment at Fermilab (E989) aims to measure the anomalous magnetic moment of the muon, $a_{\mu}= (g-2)/2$, to a groundbreaking precision of $140$ ppb, obtaining a near four-fold increase in precision over the previous experiment, E821, at the Brookhaven National Laboratory (BNL). The value of $a_{\mu}$ from BNL currently differs from the Standard Model prediction by $\sim 3.7$ standard deviations, suggesting the potential for new physics and therefore, motivating a new experiment.Because the theory predicts this number with high precision, testing the g-factor through experiment provides a stringent test of the SM and can suggest physics beyond the Standard Model. The goal of the Fermilab Muon $g-2$ experiment is to increase the statistical precision by more than a factor of 20 and reduce systematic errors by a factor of 3. By measuring muon precession rate ($\omega_a$) in an external magnetic field, the anomalous magnetic moment will be calculated. This is an incredibly challenging experiment with a unique opportunity to provide new insight into nature. \\ The $g-2$ data also provides a great opportunity for setting the most stringent limits on some of the Standard Model Extension CPT Lorentz violating (LV) parameters in the muon sector. One of the CPT and Lorentz violating signatures that we can look for using $g-2$ data is a sidereal variation of $\omega_a(t)$. Extensive simulation studies confirm that the sensitivity regarding the sidereal varation roughly scales with $\omega_a$ uncertainty. Hence, the $g-2$ experiment at FNAL should be able to reach limits of $\sim 5\times10^{-25}$ GeV. Because the CPT and LV analyses are essentially studies of variations in $\omega_a$ as a function of time and charge, performing an $\omega_a$ analysis sets the stage for the CPT and LV measurement. This dissertation focuses on the methodology of a fully functioning framework and analyzing the Fermilab Muon $g - 2$ Run 2 data containing $\sim 11$ billion events above an energy threshold of $1.7$~GeV.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2026

Image Distinguishability Analysis Testing Through Principal Components and Its Application to Hot Spot Scale Invariance

Hot spots are spatial regions of intense energy localization that govern initiation of secondary high explosives. Studies that characterize or compare simulated hot spots are frequently either qualitatively descriptive or resort to quantitative distribution functions that neglect stochastic variations and spatial correlations—effects that are also neglected in common comparison tests like the Kolmogorov–Smirnov test. To this end, we develop an image distinguishability analysis (IDA) test based on principal component (PC) analysis that makes pixel-by-pixel comparisons between small, for example, O(<10), image data sets. The IDA test makes comparisons through a generalized distance metric in the PC space and a test statistic that is derived to calculate mathematical equation-values. Here, we derive a statistical distribution and criticality criterion to determine whether images are distinguishable from established baselines. We apply the IDA test on images generated from molecular dynamics simulations of hot spots from pore collapse in TATB to assess scale invariance in the complex patterns of hot spots that form in a representative high explosive crystal. The IDA test shows that TATB hot spot spatial temperature fields and their derived temperature histograms exhibit scale-invariant features over specific intervals of shock orientation, strength, and initial pore diameter. However, the IDA test also shows that qualitatively different conclusions regarding invariance can be reached depending on whether the hot spot is treated as a spatially correlated field as opposed to a distribution function that lacks spatial information.

organic↗
DOE OSTI2021

Variations in pharmacokinetic-pharmacodynamic target values across MICs and their potential impact on determination of susceptibility test interpretive criteria

Abstract Background An antibacterial drug’s susceptibility test interpretive criteria (STIC) are determined by integrating clinical, microbiological and pharmacokinetic-pharmacodynamic (PK-PD) data. PTA analysis plays a pivotal or supportive role in STIC determination and is heavily dependent on the PK-PD target values determined from animal PK-PD studies. Therefore, variations in PK-PD target values may impact STIC determination. Factors contributing to variation in the PK-PD target values include the number of and MICs for bacterial isolates used in animal PK-PD studies. Objectives To analyse the relationship between PK-PD target values and MICs, describe the variations in PK-PD target values of isolates and evaluate whether the proposed/target STICs were within the ranges of the MICs for isolates used in animal PK-PD studies. Methods A database was compiled for this research by screening animal PK-PD study reports submitted to the FDA from 10 new drug applications (NDAs). Results A relationship evaluation between PK-PD target values and MICs for tested isolates for seven drugs (that used AUC/MIC ratio as the PK-PD index) showed that, generally, the AUC/MIC values decreased with an increase in MIC. These target values were highly variable, with the percentage coefficient of variation ranging between 1% and 132% for isolates having the same MIC. For 16/27 (59%) drug/bacteria combinations from all 10 drugs, the proposed/target STICs were higher than the highest MIC for bacteria isolates evaluated, while 6/27 (22.5%) were lower. Conclusions This research suggests that careful considerations related to selection of bacterial isolates for animal PK-PD studies could strengthen the STIC determination process.

Waack, Ursula↗
DOE OSTI2020

Aging Effect Analysis of PV Inverter Semiconductors for Ancillary Services Support

PV inverters can provide reactive power while generating active power. An ongoing microgrid implementation at Duke Energy actively engages non-utility PVs to generate/absorb reactive power in support of ancillary services to increase microgrid resiliency during extreme events. PV systems are requested to provide reactive power support: 1) in response to grid voltage variation to better regulate the local voltage; or 2) in response to utility incentives, such as following Transactive Energy System (TES) incentives. However, providing ancillary services might shorten the lifetime expectation of PV inverter semiconductors. This paper summarizes the potential impacts on a PV inverter semiconductor's lifetime when providing ancillary services. The analysis presented in this research work shows that providing reactive power support will increase the mean junction temperature and the junction temperature variation of the inverter diodes. This increased junction temperature will eventually lead to shorter diode lifetime. The lifetime estimation of semiconductors is briefly reviewed. The power losses of PV inverter semiconductors are derived as a support analysis to the junction temperature calculation. In addition, the impact of the filtering inductor on the semiconductor current distribution is discussed. The theoretical analysis presented in this research work is supported by simulation results.

14 SOLAR ENERGY↗
DOE OSTI2026

A trait syndrome ties cell morphology to glycolysis across the yeast subphylum

Co-variation of traits provides fundamental insights into constraints governing their evolution. An inverse correlation between glucose uptake rates (GURs) and cell surface area-to-volume (SA:V) ratios across 11 yeast species was recently reported. Here, we expand the analysis to 282 species to assess the generalizability of this correlation across Saccharomycotina yeasts and the contribution of shared evolutionary history to the co-variation of these traits. Using phylogenetic regression models, we found extracellular acidification rates (ECARs, used as a proxy for GURs) were weakly, but significantly, correlated with SA:V across Saccharomycotina. ECARs were also correlated with genome sizes and growth rates. Our findings support the reported correlation between GURs and SA:V ratios, but suggest other associated traits, including genome sizes. Specifically, yeasts that consume glucose faster tend to have lower SA:V, faster growth rates, and larger genomes, suggesting a trait syndrome governing several metabolic, genomic, and morphological traits across the yeast subphylum.

biological sciences↗
DOE OSTI2025

Impact of 12-nm FinFET Technology Variations on TID Effects: A Comparative Study of GF 12LP and 12LP+ at the Transistor Level

Here, this article presents a comparative analysis of total ionizing dose (TID) response in GlobalFoundries’ (GF) 12 low-power (LP) and 12LP+12-nm bulk fin field effect transistor (FinFET) technologies using 10-keV X-rays. Our findings show that 12LP+ n-type transistors demonstrate higher sensitivity to TID degradation of the off-state leakage drain current compared to 12LP. Data indicate that for both 12LP and 12LP+, transistors with higher threshold voltages (VTs) exhibit lower off-state drain-source leakage postirradiation compared to transistors with lower VTs. Data consistently show that transistors with fewer fins per transistor show superior TID tolerance, in both 12LP and 12LP+ technologies. Lower VT transistors in both technologies display similar preirradiation leakage currents. On the other hand, higher VT transistors in 12LP+ show lower preirradiation leakage currents than those in 12LP, highlighting that the front-end-of-line of 12LP+ technology has been modified compared to 12LP. p-type devices in 12LP+ presented negligible degradation. Larger TID sensitivity in 12LP+ might be attributed to the implementation of dual-metal gate work functions, reduced halo doping, deeper source/drain (S/D) doping profiles, and/or 12LP+ having narrower fins compared to 12LP.

Dual-metal gate work functions↗
DOE OSTI2025

Jensen–Shannon divergence based novel loss functions for Bayesian neural networks

Bayesian neural networks (BNNs) are state-of-the-art machine learning methods that can naturally regularize and systematically quantify uncertainties using their stochastic parameters. Kullback–Leibler (KL) divergence-based variational inference used in BNNs suffer from unstable optimization and challenges in approximating light-tailed posteriors due to the unbounded nature of the KL divergence. To resolve these issues, we formulate a novel loss function for BNNs based on a new modification to the generalized Jensen–Shannon (JS) divergence, which is bounded. In addition, we propose a Geometric JS divergence-based loss, which is computationally efficient since it can be evaluated analytically. We found that the JS divergence-based variational inference is intractable, and hence employed a constrained optimization framework to formulate these losses. Our theoretical analysis and empirical experiments on multiple regression and classification data sets suggest that the proposed losses perform better than the KL divergence-based loss, especially when the data sets are noisy or biased. Specifically, there are approximately 5% and 8% improvements in accuracy for a noise-added CIFAR-10 dataset and a regression dataset, respectively. There is about 13% reduction in false negative predictions of a biased histopathology dataset. Additionally, we quantify and compare the uncertainty metrics for the regression and classification tasks.

97 MATHEMATICS AND COMPUTING↗