Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Average Accuracy”

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 1,441 records · Page 80
All sourcesNASA NTRSDOE OSTI
DOE OSTI2023

Thermodynamic investigation of the NaCl-KCl salt system from 25 to 950 °C

NaCl-KCl molten salt system has been proposed as a primary component of several promising heat transfer eutectics and fuel host for Generation IV molten salt reactors (MSR). In this work, several key thermodynamic parameters, including high temperature enthalpies, isobaric heat capacities (C p ) and molar enthalpies of mixing (ΔH mix ), of the NaCl -KCl system (NaCl, 75mol% NaCl – 25mol% KCl, 51mol% NaCl – 49mol% KCl, 25mol% NaCl – 75mol% KCl, and KCl) were measured by high temperature drop calorimetry (HTDC) utilizing laser sealed nickel and aluminum crucibles. Salts were checked for phase purity by X-ray diffraction (XRD), and water content by Karl Fischer coulometric titrimetry and thermogravimetric analysis (TGA). Thermo-mechanical analysis (TMA) utilizing custom-built boron nitride (BN) crucibles was employed to determine temperatures of phase transitions (solid $\rightarrow$ liquid) for the salts and plotted against the pseudobinary phase diagram for the NaCl-KCl system. Small deviations (3-10 %) were found among C p values obtained from HTDC, differential scanning calorimetry (DSC), and NIST, comparable to the average uncertainty based on two standard deviations of measured data that are ~6% for DSC and ~3% for HTDC. Additionally, the measured enthalpies and C p values of mixed NaCl-KCl salts (e.g., 51mol% NaCl – 49mol% KCl eutectic) were shown to behave as statistical mixtures of its pure endmembers, consistent with ΔH mix = ~ 0 kJ/mol at 800 °C, and a regular interaction parameter of Ω = 1.20 ± 0.76 kJ/mol. In conclusion, this study represents a step towards improved accuracy and precision in determinations of thermodynamic parameters of molten salts by using DSC, TMA, and HTDC, in completion of the Molten Salt Thermal Properties Database (MSTDB).

11 NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗
DOE OSTI2022

Differentiation and classification of bacterial endotoxins based on surface enhanced Raman scattering and advanced machine learning

Bacterial endotoxin, a major component of the Gram-negative bacterial outer membrane leaflet, is a lipopolysaccharide shed from bacteria during their growth and infection and can be utilized as a biomarker for bacterial detection. Here, the surface enhanced Raman scattering (SERS) spectra of eleven bacterial endotoxins with an average detection amount of 8.75 pg per measurement have been obtained based on silver nanorod array substrates, and the characteristic SERS peaks have been identified. With appropriate spectral pre-processing procedures, different classical machine learning algorithms, including support vector machine, k-nearest neighbor, random forest, etc., and a modified deep learning algorithm, RamanNet, have been applied to differentiate and classify these endotoxins. It has been found that most conventional machine learning algorithms can attain a differentiation accuracy of >99%, while RamanNet can achieve 100% accuracy. Such an approach has the potential for precise classification of endotoxins and could be used for rapid medical diagnoses and therapeutic decisions for pathogenic infections.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2024

Reinforcement Learning as a Parsimonious Alternative to Prediction Cascades: A Case Study on Image Segmentation

Deep learning architectures have achieved state-of-the-art (SOTA) performance on computer vision tasks such as object detection and image segmentation. This may be attributed to the use of over-parameterized, monolithic deep learning architectures executed on large datasets. Although such large architectures lead to increased accuracy, this is usually accompanied by a larger increase in computation and memory requirements during inference. While this is a non-issue in traditional machine learning (ML) pipelines, the recent confluence of machine learning and fields like the Internet of Things (IoT) has rendered such large architectures infeasible for execution in low-resource settings. For some datasets, large monolithic pipelines may be overkill for simpler inputs. To address this problem, previous efforts have proposed decision cascades where inputs are passed through models of increasing complexity until the desired performance is achieved. However, we argue that cascaded prediction leads to sub-optimal throughput and increased computational cost due to wasteful intermediate computations. To address this, we propose PaSeR (Parsimonious Segmentation with Reinforcement Learning) a non-cascading, cost-aware learning pipeline as an efficient alternative to cascaded decision architectures. Through experimental evaluation on both real-world and standard datasets, we demonstrate that PaSeR achieves better accuracy while minimizing computational cost relative to cascaded models. Further, we introduce a new metric IoU/GigaFlop to evaluate the balance between cost and performance. On the real-world task of battery material phase segmentation, PaSeR yields 179% improvement over SOTA MatPhase model and a 196% improvement over IDK Cascades under the IoU/GigaFlop metric. We also demonstrate PaSeR’s adaptability to complementary models trained on a noisy MNIST dataset, where it outperforms all baselines on IoU/GigaFlop by an average of 44%.

Srikshan, Bharat↗
NASA NTRS1972

The geocentric orientation vector from limited Astro-geodetic data

Test computations using the comprehensive sets of both astro-geodetic and gravimetric data indicate that the accuracy of the geocentric orientation vector determined from incomplete representations of the earth's gravitational fields, is dependent on the overall extent of the datum. The lack of an astro-geodetic determination of the geoid due to an inadequate density of astro-geodetic stations over certain parts of the datum does not materially affect these determinations. The effect is only marginal if the station density of the area referred to does not markedly differ from the average for the region.

Mather, R. S.↗
NASA NTRS1979

Gravity anomaly and geoid undulation results in local areas from GEOS-3 altimeter data

The adjusted GEOS-3 altimeter data, taken as averages within a data frame, have been used to construct free air anomaly and geoid undulation profiles and maps in areas of geophysical interest. Profiles were constructed across the Philippine Trench (at a latitude of 6 deg) and across the Bonin Trench (at a latitude of 28 deg). In the latter case an anomaly variation of 443 mgals in 143 km was derived from the altimeter data. These variations agreed reasonably with terrestrial estimates, considering the predicted point accuracy was about + or - 27 mgals. An area over the Patton Sea mounts was also investigated with the altimeter anomaly field agreeing well with the terrestrial data except for the point directly over the top of the sea mount. It is concluded that the GEOS-3 altimeter data is valuable not only for determining 5 deg and 1 deg x 1 deg mean anomalies, but also can be used to describe more local anomaly variations.

Rapp, R. H.↗
NASA NTRS1987

Enhancements to Program LAURA for computation of three-dimensional hypersonic flow

Changes to Program Laura (Langley Aerothermodynamic Upwind Relaxation Algorithm) are presented which enhance both stability and accuracy of the algorithm. A discussion of iteration/sweeping strategies and their relation to computer architectures is included to best exploit the capabilities of serial, vector, and parallel processor machines. Test cases for Mach 10 perfect gas flow and Mach 32 real gas flow in chemical nonequilibrium over a blunt, raked elliptic cone using the thin-layer Navier-Stokes equations are presented in order to demonstrate the current improved capabilities. Algorithm changes include the use of volume averaging, application of a symmetric total variation diminishing (TVD) scheme, and stronger interaction between the grid/shock alignment routine and the relaxation algorithm. Good comparisons with heat transfer and pitching moment data at three different angles of attack for the Mach 10 tests serve to further validate the present algorithm. Parameters are defined which control the coupling of the specie continuity equations with the solution of the mixture conservation equations. A discussion of the consequences involved in the choice of strong versus weak coupling is presented, and a sample nonequilibrium calculation on a fine grid over a full scale model of the Aeroassist Flight Experiment (AFE) demonstrates current capabilities.

Gnoffo, Peter A.↗
NASA NTRS2002

Calibration Methodology for the Lightweight Rainfall Radiometer STAR Aircraft Sensor

A proposed methodology for the in-flight calibration of a Synthetic Thinned Aperture Radiometer (STAR) airborne sensor with the potential application to a space flight version. The application of the spaceflight version of this instrument will address several pressing issues related to the Global Precipitation Measurement Mission (GPM). The X-Band Lightweight Rainfall Radiometer using STAR technology (LRR-X) is an aircraft sensor that is jointly developed by the NASA Goddard Space Flight Center and the University of Michigan. This paper will describe the theory of calibration as well as the hardware design specifications used by the method. The on-board hardware uses individual uncorrelated warm loads on each receiver as well as to a single noise diode providing a correlated noise source to each receiver. A procedure for maintaining onboard calibration with an optimum running average using correlated bursts of thermal noise interleaved with scene data will be exercised during the maiden flight of the LRR-X instrument during the spring of 2003. The final component of calibration of a synthetic aperture radiometer is the image reconstruction algorithm that uses the measured correlations to produce the temperature brightness (TB) images. An overview of system-level testing, both on the ground and in-flight, will be presented to validate the absolute accuracy of the image reconstruction algorithm.

Principe, Caleb↗
DOE OSTI2023

Characterizing post-fire delayed tree mortality with remote sensing: sizing up the elephant in the room

Abstract Background Despite recent advances in understanding the drivers of tree-level delayed mortality, we lack a method for mapping delayed mortality at landscape and regional scales. Consequently, the extent, magnitude, and effects of delayed mortality on post-fire landscape patterns of burn severity are unknown. We introduce a remote sensing approach for mapping delayed mortality based on post-fire decline in the normalized burn ratio (NBR). NBR decline is defined as the change in NBR between the first post-fire measurement and the minimum NBR value up to 5 years post-fire for each pixel. We validate the method with high-resolution aerial photography from six wildfires in California, Oregon, and Washington, USA, and then compare the extent, magnitude, and effects of delayed mortality on landscape patterns of burn severity among fires and forest types. Results NBR decline was significantly correlated with post-fire canopy mortality (r 2 = 0.50) and predicted the presence of delayed mortality with 83% accuracy based on a threshold of 105 NBR decline. Plots with NBR decline greater than 105 were 23 times more likely to experience delayed mortality than those below the threshold (p < 0.001). Delayed mortality occurred across 6–38% of fire perimeters not affected by stand-replacing fire, generally affecting more areas in cold (22–41%) and wet (30%) forest types than in dry (1.7–19%) types. The total area initially mapped as unburned/very low-severity declined an average of 38.1% and generally persisted in smaller, more fragmented patches when considering delayed mortality. The total area initially mapped as high-severity increased an average of 16.2% and shifted towards larger, more contiguous patches. Conclusions Differences between 1- and 5-year post-fire burn severity maps depict dynamic post-fire mosaics resulting from delayed mortality, with variability among fires reflecting a range of potential drivers. We demonstrate that tree-level delayed mortality scales up to alter higher-level landscape patterns of burn severity with important implications for forest resilience and a range of fire-driven ecological outcomes. Our method can complement existing tree-level studies on drivers of delayed mortality, refine mapping of fire refugia, inform estimates of habitat and carbon losses, and provide a more comprehensive assessment of landscape and regional scale fire effects and trends.

Environmental Sciences & Ecology↗
DOE OSTI2021

A Novel Laser-Aided Machining and Polishing Process for Additive Manufacturing Materials with Multiple Endmill Emulating Scan Patterns

In additive manufacturing (AM), the surface roughness of the deposited parts remains significantly higher than the admissible range for most applications. Additionally, the surface topography of AM parts exhibits waviness profiles between tracks and layers. Therefore, post-processing is indispensable to improve surface quality. Laser-aided machining and polishing can be effective surface improvement processes that can be used due to their availability as the primary energy sources in many metal AM processes. While the initial roughness and waviness of the surface of most AM parts are very high, to achieve dimensional accuracy and minimize roughness, a high input energy density is required during machining and polishing processes although such high energy density may induce process defects and escalate the phenomenon of wavelength asperities. In this paper, we propose a systematic approach to eliminate waviness and reduce surface roughness with the combination of laser-aided machining, macro-polishing, and micro-polishing processes. While machining reduces the initial waviness, low energy density during polishing can minimize this further. The average roughness (Ra=1.11μm) achieved in this study with optimized process parameters for both machining and polishing demonstrates a greater than 97% reduction in roughness when compared to the as-built part.

42 ENGINEERING↗
NASA NTRS1982

Low-level water vapor fields from the VISSR atmospheric sounder (VAS) split window channels at 11 and 12 microns

A series of high-resolution water vapor fields were derived from the 11 and 12 micron channels of the VISSR Atmospheric Sounder (VAS) on GOES-5. The low-level tropospheric moisture content was separated from the surface and atmospheric radiances by using the differential adsorption across the 'split window' along with the average air temperature from imbedded radiosondes. Fields of precipitable water are presented in a time sequence of five false color images taken over the United States at 3-hour intervals. Vivid subsynoptic and mesoscale patterns evolve at 15 km horizontal resolution over the 12-hour observing period. Convective cloud formations develop from several areas of enhanced low-level water vapor, especially where the vertical water vapor gradient relatively strong. Independent verification at radiosonde sites indicates fairly good absolute accuracy, and the spatial and temporal continuity of the water vapor features indicates very good relative accuracy. Residual errors are dominated by radiometer noise and unresolved clouds.

Chesters, D.↗
NASA NTRS2010

Modified Phasemeter for a Heterodyne Laser Interferometer

Modifications have been made in the design of instruments of the type described in "Digital Averaging Phasemeter for Heterodyne Interferometry". A phasemeter of this type measures the difference between the phases of the unknown and reference heterodyne signals in a heterodyne laser interferometer. The phasemeter design lacked immunity to drift of the heterodyne frequency, was bandwidth-limited by computer bus architectures then in use, and was resolution-limited by the nature of field-programmable gate arrays (FPGAs) then available. The modifications have overcome these limitations and have afforded additional improvements in accuracy, speed, and modularity. The modifications are summarized.

Loya, Frank M.↗
NASA NTRS2014

Calibrating Historical IR Sensors Using GEO, and AVHRR Infrared Tropical Mean Calibration Models

Long-term, remote-sensing-based climate data records (CDRs) are highly dependent on having consistent, wellcalibrated satellite instrument measurements of the Earth's radiant energy. Therefore, by making historical satellite calibrations consistent with those of today's imagers, the Earth-observing community can benefit from a CDR that spans a minimum of 30 years. Most operational meteorological satellites rely on an onboard blackbody and space looks to provide on-orbit IR calibration, but neither target is traceable to absolute standards. The IR channels can also be affected by ice on the detector window, angle dependency of the scan mirror emissivity, stray-light, and detector-to-detector striping. Being able to quantify and correct such degradations would mean IR data from any satellite imager could contribute to a CDR. Recent efforts have focused on utilizing well-calibrated modern hyper-spectral sensors to intercalibrate concurrent operational IR imagers to a single reference. In order to consistently calibrate both historical and current IR imagers to the same reference, however, another strategy is needed. Large, well-characterized tropical-domain Earth targets have the potential of providing an Earth-view reference accuracy of within 0.5 K. To that effort, NASA Langley is developing an IR tropical mean calibration model in order to calibrate historical Advanced Very High Resolution Radiometer (AVHRR) instruments. Using Meteosat-9 (Met-9) as a reference, empirical models are built based on spatially/temporally binned Met-9 and AVHRR tropical IR brightness temperatures. By demonstrating the stability of the Met-9 tropical models, NOAA-18 AVHRR can be calibrated to Met-9 by matching the AVHRR monthly histogram averages with the Met-9 model. This method is validated with ray-matched AVHRR and Met-9 biasdifference time series. Establishing the validity of this empirical model will allow for the calibration of historical AVHRR sensors to within 0.5 K, and thereby establish a climate-quality IR data record.

Scarino, Benjamin↗
NASA NTRS2019

Characteristics of Total Column CO2 Retrievals from the OCO Missions: Biases, Information Content and Implications for Flux Inversions

The Orbiting Carbon Observatory-2 and Orbiting Carbon Observatory-3, launched in 2015 and 2019, respectively, are intended to collect and deliver high-resolution observations of CO2 with unprecedented space and time coverage. Observations of CO2 from these remote-sensing missions (also known as XCO2, or column-based average, dry air mole fraction of CO2) are then used by the global carbon cycle community to answer a wide range of science questions, from the distribution and quantification of global and regional CO2 source-sink patterns to quantification of anthropogenic sources at urban scales. Even though we have had the OCO-2 mission flying for a few years now, the retrieval algorithms are continuously evolving and improving to deliver XCO2 retrievals with very high precision and high accuracy (or low biases). In this presentation, we will discuss a simple yet effective quantitative framework that has been developed by the OCO-2 flux team to evaluate the information content of these XCO2 retrievals as soon as they are released, i.e., with lower latency than full-scale flux inversions. This framework serves as a precursor to advanced inverse modeling frameworks and is intended to provide an early but accurate assessment of the signal present in the satellite retrievals, the robustness of that signal, and the ability of these retrievals to resolve patterns in CO2 surface fluxes that cannot be resolved by our current network of surface sites. Specific results will tackle a tiered set of questions that are being addressed using this framework: (a) what are the distribution of retrievals in the different modes of operation and how do they vary in space and time? (b) what is the information that is being given to the inverse modeling frameworks from the space-based data, information above and beyond what is provided by the in-situ data? and (c) how do these factors influence our choices for doing flux inversions with the satellite retrievals? While the primary focus of the results will be on application of this technique to mature OCO-2 retrievals, we will show early results for a couple of months of OCO-3 retrievals. For the time-period that the retrievals from the two missions overlap, we will highlight how this framework allows us to effortlessly put the information from OCO-3 and OCO-2 on an equal footing, thus enabling easy comparison between the two pioneering missions.

Chatterjee, Abhishek↗
NASA NTRS2022

Solar Temperature Variations Computed from SORCE SIM Irradiances Observed During 2003-2020

NASA’s Solar Radiation and Climate Experiment (SORCE) Spectral Irradiance Monitor (SIM) instrument produced about 17 years of daily average Spectral Solar Irradiance ( SSI ) data for wavelengths 240 nm – 2416 nm. We choose a day of minimal solar activity, 2008-08-24, during the 2008 − 2009 minimum between cycles 23 and 24, and compute the brightness temperature (𝑇 o ) from that day’s solar spectral irradiance (𝑆𝑆𝐼 o ). We consider small variations of T and SSI about these reference values, and derive linear and quadratic analytic approximations by Taylor expansion about the reference day values. To determine approximation accuracy, we compare to exact brightness temperatures T computed from the Planck spectrum, by solving analytically for T , or equivalent root-finding in Wolfram Mathematica. We find that the linear analytic approximation overestimates, while the quadratic underestimates the exact result. This motivates search for statistical “fit” models “in between” the two analytic models, with minimum root-mean-square-error RMSE. We make this search using open-source statistical R software, determine coefficients for linear and quadratic fit models, and compare statistical with analytic RMSE’s. When only linear analytic and fit models are compared, the fit model is superior at ultraviolet, visible, and near infrared wavelengths. This again holds true when comparing only quadratic models. Quadratic is superior to linear for both analytic and statistical models, and statistical fits give smallest RMSE’s. Lastly, we use linear analytic and fit models to find an interpolating function in wavelength, useful in case the SIM results need adjustment to another choices of wavelengths, to compare or extend to any other instrument.

SORCE↗
DOE OSTI2021

A comparative study of machine learning models for predicting the state of reactive mixing

Mixing phenomena are important mechanisms controlling flow, species transport, and reaction processes in fluids and porous media. Accurate predictions of reactive mixing are critical for many Earth and environmental science problems such as contaminant fate and remediation, macroalgae growth, and plankton biomass evolution. Here, to investigate the evolution of mixing dynamics under different scenarios (e.g., anisotropy, fluctuating velocity fields), a finite-element-based numerical model was built to solve the fast, irreversible bimolecular reaction-diffusion equations to simulate a range of reactive-mixing scenarios. A total of 2,315 simulations were performed using different sets of model input parameters comprising various spatial scales of vortex structures in the velocity field, time-scales associated with velocity oscillations, the perturbation parameter for the vortex-based velocity, anisotropic dispersion contrast (i.e., ratio of longitudinal-to-transverse dispersion), and molecular diffusion. The outputs comprised concentration profiles of reactants and products. The inputs to and outputs from these simulations were concatenated into feature and label matrices, respectively, to train 20 different machine learning (ML) models intended to emulate system behavior. These 20 ML emulators, based on linear methods, Bayesian methods, ensemble learning methods, and multilayer perceptrons (MLPs), were trained to classify the state of mixing and predict three quantities of interest (QoIs) characterizing species production, decay (i.e., average concentration, square of average concentration), and degree of mixing (i.e., variances of species concentration). Unsurprisingly, linear classifiers and regressors failed to reproduce the QoIs; however, ensemble methods (classifiers and regressors) and the MLP model accurately classified the state of reactive mixing and the QoIs. Among ensemble methods, random forest and decision-tree-based AdaBoost faithfully predicted the QoIs. At run time, trained ML emulators produced results times faster than the finite-element simulations. Due to their low computational expense and high accuracy, ensemble and MLP models are excellent emulators for these numerical simulations and great utilities in uncertainty quantification exercises, which can require 1,000s of forward model runs.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

Improving scalability of parallel CNN training by adaptively adjusting parameter update frequency

Synchronous SGD with data parallelism, the most popular parallelization strategy for CNN training, suffers from the expensive communication cost of averaging gradients among all workers. The iterative parameter updates of SGD cause frequent communications and it becomes the performance bottleneck. In this paper, we propose a lazy parameter update algorithm that adaptively adjusts the parameter update frequency to address the expensive communication cost issue. Our algorithm accumulates the gradients if the difference of the accumulated gradients and the latest gradients is sufficiently small. Here, the less frequent parameter updates reduce the per-iteration communication cost while maintaining the model accuracy. Our experimental results demonstrate that the lazy update method remarkably improves the scalability while maintaining the model accuracy. For ResNet50 training on ImageNet, the proposed algorithm achieves a significantly higher speedup (739.6 on 2048 Cori KNL nodes) as compared to the vanilla synchronous SGD (276.6) while the model accuracy is almost not affected (<0.2% difference).

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2020

Accuracy of HVAC Load Predictions: Validation of EnergyPlus and DOE-2 using FLEXLAB Measurements

The aim of the project reported here was to better understand the level of accuracy of three building energy simulation (BES) engines (‘engines’) — EnergyPlus™, DOE-2.1e, and DOE-2.2 — by identifying and investigating significant deviations between the performance predicted by these engines and actual performance as measured in the FLEXLAB® test facility at Lawrence Berkeley National Laboratory (LBNL). The specific test conditions included some of those prescribed in ANSI/ASHRAE Standard 140 - Standard Method of Test for the Evaluation of Building Energy Analysis Computer Programs. Detailed measurements of FLEXLAB performance, including indoor temperatures and heat fluxes and air-flow and water flow rates and temperatures in the Heating, Ventilating and Air Conditioning (HVAC) system, together with hourly weather data, were recorded and used in analyzing the simulation results from EnergyPlus v8.8, DOE-2.2 v3.65 and DOE-2.1e v127. These engines are commonly used in the United States for building energy code compliance, federal, state, and utility incentives programs, as well as energy efficient design of new buildings and energy retrofit of existing buildings. Seven conventional overhead mixing ventilation scenarios were tested and each engine was found to have a similar level of agreement with the measurements of space-level heating and sensible cooling loads. These results provide useful information regarding the accuracy of these engines in predicting the cooling and heating load elements of whole building energy performance. This information is intended for practitioners who are concerned about transitioning between simulation tools with different engines and for managers of utility programs leveraging these tools for evaluating and/or projecting measure savings to be incentivized under their programs. The results of the comparisons of simulated and measured performance indicate that the predictions from all three engines are not significantly different. The 24-hour average value of the absolute mean bias indicates the likely magnitude of the error in any particular case. The average mean bias is reduced by cancelation of overprediction in one case by underprediction in another. The daytime absolute mean biases, which may be more important for both energy performance and occupant comfort, are ~6%, presumably because of the greater complexity involved in simulating in the presence of solar radiation. EnergyPlus typically overpredicts the cooling load and/or underpredicts the heating load by ~1.5% and the DOE-2 engines typically underpredict the cooling load by approximately the same amount. The Root Mean Square Error is relatively more sensitive to shorter term variations in the difference between predicted and measured loads; the three engines have similar values, ~10%, suggesting that the uncertainties in their predictions of peak loads may also be similar in magnitude. The implication of these results is that users, both designers and program analysts, can use EnergyPlus, DOE-2.1e, or DOE-2.2 to model conventional commercial buildings equipped with overhead mixing ventilation with a similar level of confidence. Further work is required to better understand the variability in the level of agreement between the engine predictions and FLEXLAB measurements, where a particular engine will agree well with FLEXLAB in some cases and not so well in others and another engine will agree or disagree in different cases. As the sources of this variability are identified and eliminated or reduced significantly, it is recommended that the experimental capabilities and methods developed in the study reported here should be applied to validating heating and cooling load calculations for spaces with different types of furniture and miscellaneous loads. These methods should then be applied to low energy space conditioning systems in EnergyPlus including, in particular, radiant slab and radiant ceiling panel cooling and heating systems and ‘mixed mode’ systems that combine mechanical cooling and natural ventilation systems, focusing on controls, including control of thermal mass. The work reported here addresses the conventional method of heating and cooling occupied spaces; other methods, such as the use of radiant heating and cooling systems have the potential to provide equivalent occupant comfort, or better, with lower energy consumption. These systems are addressed more explicitly in EnergyPlus but there is a need for empirical validation to give users the same level of confidence in modeling these systems that they have, or should have, in modeling conventional systems, based on the results presented here.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2020

Improving radiograph analysis throughput through transfer learning and object detection

SIGN Fracture Care International partners with surgeons in low-resource hospitals worldwide to provide access to effective orthopedic care by donating educational materials and innovatively designed surgical implants. Over two decades, SIGN’s Online Surgical Database (SOSD) has grown to contain over 500,000 medical images, with radiographs holding the majority share. One challenge in working with hospitals worldwide is that both the radiographs uploaded to the SOSD and the data entry accompanying the uploads vary in quality. To improve the accuracy of data in the SOSD, we trained a model to detect surgical implants in radiographs. We first developed a tool to automatically detect radiographs, then trained an object detection model to determine the number and placement of surgical implants visible in the radiograph. Active learning was used to generate a training set containing 2,510 radiographs with screws, nails, and plates labeled by bounding boxes. Training a model to simultaneously recognize all three classes of implants gave a low average precision (AP) for the plate class, likely due to the low number of plate instances in our training set and the large variety of surgical plates used by SIGN-partnered surgeons. Applying standard image augmentation techniques to increase the plate count in our training set did not appreciably increase the AP of plate detection. To improve plate detection, we redrew the bounding boxes to account for correlations between the screw and plate classes. Training one model to detect nails and screws and a separate model to detect plates increased the AP of plate detection by 78.8 percentage points. The AP of each class was 80.7% for screws, 93.6% for nails, and 92.6% for plates; meanwhile, the sensitivity was 92% for screws, 86% for nails, and 81% for plates. We show that object detection methods can be used to detect surgical implants in radiographs of varying quality; however, the detection ability is dependent on the type of implant, and some implants, in our case plates, must be treated differently than others. Such tools can improve the throughput of radiograph analysis, assisting physicians and surgeons with the treatment of bone fractures.

60 APPLIED LIFE SCIENCES↗