Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “regularized regression”

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 73 records · Page 4
All sourcesNASA NTRSDOE OSTI
NASA NTRSDate not supplied

Improvements of the Load Schedule for the Machine Calibration of a Strain-Gage Balance

The load schedule for the calibration of a six-component force balance in a calibration machine was improved. Now, single-component loads are repeated in regular intervals during the calibration. This approach has several advantages. First, the number of single-component loads increases to about twenty-two percent of all loads and load combinations. Consequently, more accurate numerical estimates of the primary bridge sensitivities can be obtained if global regression is used for the analysis of the calibration data. In addition, single-component repeats make it possible to track the stability of the applied loads during the calibration process. Finally, interactions of single-component repeats can be compared with interactions that are observed during the application of manual loads to the balance. Machine calibration and manual data sets of two force balances are used to illustrate benefits of the new load schedule. It is shown in the examples how differences between the observed interactions of machine calibration and manual data can be quantified. The suggested improvements can also be implemented in the load schedule for the machine calibration of a moment or direct-read balance as long as single-component loads are included that are described in the design load format of the balance.

wind tunnel test↗
NASA NTRS2009

MODIS Tree Cover Validation for the Circumpolar Taiga-Tundra Transition Zone

A validation of the 2005 500m MODIS vegetation continuous fields (VCF) tree cover product in the circumpolar taiga-tundra ecotone was performed using high resolution Quickbird imagery. Assessing the VCF's performance near the northern limits of the boreal forest can help quantify the accuracy of the product within this vegetation transition area. The circumpolar region was divided into longitudinal zones and validation sites were selected in areas of varying tree cover where Quickbird imagery is available in Google Earth. Each site was linked to the corresponding VCF pixel and overlaid with a regular dot grid within the VCF pixel's boundary to estimate percent tree crown cover in the area. Percent tree crown cover was estimated using Quickbird imagery for 396 sites throughout the circumpolar region and related to the VCF's estimates of canopy cover for 2000-2005. Regression results of VCF inter-annual comparisons (2000-2005) and VCF-Quickbird image-interpreted estimates indicate that: (1) Pixel-level, inter-annual comparisons of VCF estimates of percent canopy cover were linearly related (mean R(sup 2) = 0.77) and exhibited an average root mean square error (RMSE) of 10.1 % and an average root mean square difference (RMSD) of 7.3%. (2) A comparison of image-interpreted percent tree crown cover estimates based on dot counts on Quickbird color images by two different interpreters were more variable (R(sup 2) = 0.73, RMSE = 14.8%, RMSD = 18.7%) than VCF inter-annual comparisons. (3) Across the circumpolar boreal region, 2005 VCF-Quickbird comparisons were linearly related, with an R(sup 2) = 0.57, a RMSE = 13.4% and a RMSD = 21.3%, with a tendency to over-estimate areas of low percent tree cover and anomalous VCF results in Scandinavia. The relationship of the VCF estimates and ground reference indicate to potential users that the VCF's tree cover values for individual pixels, particularly those below 20% tree cover, may not be precise enough to monitor 500m pixel-level tree cover in the taiga-tundra transition zone.

Montesano, P. M.↗
DOE OSTI2025

An adaptive and stability-promoting layerwise training approach for sparse deep neural network architecture

This work presents a two-stage adaptive framework for progressively developing deep neural network (DNN) architectures that generalize well for a given training data set. In the first stage, a layerwise training approach is adopted where a new layer is added each time and trained independently by freezing parameters in the previous layers. We impose desirable structures on the DNN by employing manifold regularization, sparsity regularization, and physics-informed terms. We introduce a ε – δ – stability-promoting concept as a desirable property for a learning algorithm and show that employing manifold regularization yields a ε – δ stability-promoting algorithm. Further, we also derive the necessary conditions for the trainability of a newly added layer and investigate the training saturation problem. In the second stage of the algorithm (post-processing), a sequence of shallow networks is employed to extract information from the residual produced in the first stage, thereby improving the prediction accuracy. Numerical investigations on prototype regression and classification problems demonstrate that the proposed approach can outperform fully connected DNNs of the same size. Moreover, by equipping the physics-informed neural network (PINN) with the proposed adaptive architecture strategy to solve partial differential equations, we numerically show that adaptive PINNs not only are superior to standard PINNs but also produce interpretable hidden layers with provable stability. As a result, we also apply our architecture design strategy to solve inverse problems governed by elliptic partial differential equations.

42 ENGINEERING↗
DOE OSTI2025

Neural network-based classification and regression of magnetohydrodynamic modes in tokamaks

We present a machine learning-based magnetohydrodynamic (MHD) classifier and regressor that utilizes real or complex-valued 3D magnetic sensor array data to determine neoclassical tearing mode (NTM) onset times in tokamaks with millisecond accuracy. The input dataset consists of poloidal profiles of complex Fourier amplitudes with an n = 1 toroidal mode number from 144 human-labeled ITER Baseline Scenario discharges in the DIII-D tokamak, spanning both tearing-dominated and sawtooth-dominated regimes. Since m, n = 2,1 NTMs frequently emerge alongside sawteeth at the same frequency in this scenario, the focus is on isolating the m = 1 and m = 2 components of the n = 1 MHD mode near the tearing onset. To improve model regularization and prediction stability, singular value decomposition was applied to balance the sawtooth and tearing datasets. The enriched datasets facilitated training neural networks that learn the key distinguishing features of sawtooth and tearing modes in the poloidal profiles of their magnetic amplitude and phase. When the modes occur independently, the networks achieve perfect classification due to the modes’ distinct characteristics and low measurement noise. In the more experimentally relevant case where both modes coexist, the networks maintain exceptional performance across key metrics. Tests on synthetic data with known ground truth demonstrate the superior accuracy of the neural network trained on complex-valued input compared to models using real amplitude, phase, or pseudo-complex data, achieving both a mean time delay and standard deviation below 1 ms. Notably, standard linear regression methods fitting the dominant singular modes to the data closely match the neural network’s performance. Applying these methods across a broad range of H-mode scenarios will enable future studies to systematically identify dominant NTM triggers as scenario-specific variables, paving the way for more effective tearing mode avoidance strategies in future fusion reactor designs.

machine learning↗
DOE OSTI2023

Generalizing the Gurson model using symbolic regression and transfer learning to relax inherent assumptions

Abstract To generate material models with fewer limiting assumptions while maintaining closed-form, interpretable solutions, we propose using genetic programming based symbolic regression (GPSR), a machine learning (ML) approach that describes data using free-form symbolic expressions. To maximize interpretability, we start from an analytical, derived material model, the Gurson model for porous ductile metals, and systematically relax inherent assumptions made in its derivation to understand each assumption’s contribution to the GPSR model forms. We incorporate transfer learning methods into the GPSR training process to increase GPSR efficiency and generate models that abide by known mechanics of the system. The results show that regularizing the GPSR fitness function is critical for generating physically valid models and illustrate how GPSR allows a high level of interpretability compared with other ML approaches. The method of systematic assumption relaxation allows the generation of models that address limiting assumptions found in the Gurson model, and the symbolic forms allow conjecture of decreased material strength due to void interaction and non-symmetric void shapes.

36 MATERIALS SCIENCE↗
DOE OSTI2025

A kinetic-based regularization method for data science applications

We propose a physics-based regularization technique for function learning, inspired by statistical mechanics. By drawing an analogy between optimizing the parameters of an interpolator and minimizing the energy of a system, we introduce corrections that impose constraints on the lower-order moments of the data distribution. This minimizes the discrepancy between the discrete and continuum representations of the data, in turn allowing to access more favorable energy landscapes, thus improving the accuracy of the interpolator. Our approach improves performance in both interpolation and regression tasks, even in high-dimensional spaces. Unlike traditional methods, it does not require empirical parameter tuning, making it particularly effective for handling noisy data. We also show that thanks to its local nature, the method offers computational and memory efficiency advantages over Radial Basis Function interpolators, especially for large datasets.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

Probabilistic Day-Ahead Forecasting Using an Analog Ensemble Approach for Wind Farm Grid Services

Wind resource assessment and wind power forecasting are used in research and industry to anticipate future power output at scales ranging from individual wind turbines to entire wind farms. Probabilistic day-ahead wind forecasting is useful for anticipating how a wind farm could potentially participate in the day-ahead market by providing upper and lower bounds for expected power generation, thus informing grid operators of its uncertainty. Understanding this uncertainty is part of a larger project focused on building a platform that combines efforts in weather forecasting, aerodynamic and economic modeling to create maximum value of a wind plant to better provide services to the grid. This effort is also known as the Atmosphere to Electrons to Grid (A2E2G) project. One method for producing a probabilistic forecast is through the analog ensemble approach (Delle Monache et al., 2011). This method leverages historical forecasts and their corresponding observations as a training data set from which future forecasts can be made. For some future forecast, the most similar historical forecasts (analogs) are identified on a regular time basis such as once per a 3-hour window. The most similar analogs, based on a metric such as root mean square error (RMSE), are recorded and their corresponding verifying observations are used as an ensemble member for this future forecast. Prior work in this area demonstrates improvements over raw Numerical Weather Prediction (NWP) forecasts and shows skill similar to techniques such as logistic regression and machine learning (Delle Monache et al., 2013; Alessandrini et al., 2015). Here, we take the High-Resolution Rapid Refresh model (HRRR) day-ahead forecast (0-36 hours) to create a probabilistic day-ahead forecast using an analog ensemble approach. The HRRR has an hourly temporal resolution, with a spatial resolution of 3 km. The 12 UTC HRRR model run is downloaded every day for one year from August 2019 - July 2020, with the first 11 months serving as a bank of analogs from which the forecasting algorithm can create a probabilistic forecast. Once downloaded, the original HRRR forecast is temporally interpolated to 5-minutes, aligning with both the temporal resolution of the observations as well as the timescale relevant for day-ahead power forecasts. The forecast is validated at the M2 tower at the Flatirons Campus of the National Renewable Energy Laboratory (NREL) at a typical wind turbine height of 80 m. Variables such as wind speed, wind direction, and turbulence intensity are incorporated into the probabilistic forecast model and weighted according to their relative importance to the forecast. Based on metrics such as mean bias error (MBE), mean absolute error (MAE), and root mean square error, the analog ensemble forecast outperforms the raw HRRR forecast during the testing period of July 2020. Figure 1 illustrates an example day-ahead forecast compared against the verifying observations. The general variability and ramps are captured throughout the day, with potential to further improve the analog ensemble model through machine learning techniques.

numerical weather prediction↗
DOE OSTI2021

Using heterogeneous data to identify signatures of dengue outbreaks at fine spatio-temporal scales across Brazil

Dengue virus remains a significant public health challenge in Brazil, and seasonal preparation efforts are hindered by variable intra- and interseasonal dynamics. Here, we present a framework for characterizing weekly dengue activity at the Brazilian mesoregion level from 2010–2016 as time series properties that are relevant to forecasting efforts, focusing on outbreak shape, seasonal timing, and pairwise correlations in magnitude and onset. In addition, we use a combination of 18 satellite remote sensing imagery, weather, clinical, mobility, and census data streams and regression methods to identify a parsimonious set of covariates that explain each time series property. The models explained 54% of the variation in outbreak shape, 38% of seasonal onset, 34% of pairwise correlation in outbreak timing, and 11% of pairwise correlation in outbreak magnitude. Regions that have experienced longer periods of drought sensitivity, as captured by the “normalized burn ratio,” experienced less intense outbreaks, while regions with regular fluctuations in relative humidity had less regular seasonal outbreaks. Both the pairwise correlations in outbreak timing and outbreak trend between mesoresgions were best predicted by distance. Our analysis also revealed the presence of distinct geographic clusters where dengue properties tend to be spatially correlated. Forecasting models aimed at predicting the dynamics of dengue activity need to identify the most salient variables capable of contributing to accurate predictions. Our findings show that successful models may need to leverage distinct variables in different locations and be catered to a specific task, such as predicting outbreak magnitude or timing characteristics, to be useful. This advocates in favor of “adaptive models” rather than “one-size-fits-all” models. The results of this study can be applied to improving spatial hierarchical or target-focused forecasting models of dengue activity across Brazil.

59 BASIC BIOLOGICAL SCIENCES↗
NASA NTRS2009

In-Situ Measurement of Hall Thruster Erosion Using a Fiber Optic Regression Probe

One potential life-limiting mechanism in a Hall thruster is the erosion of the ceramic material comprising the discharge channel. This is especially true for missions that require long thrusting periods and can be problematic for lifetime qualification, especially when attempting to qualify a thruster by analysis rather than a test lasting the full duration of the mission. In addition to lifetime, several analytical and numerical models include electrode erosion as a mechanism contributing to enhanced transport properties. However, there is still a great deal of dispute over the importance of erosion to transport in Hall thrusters. The capability to perform an in-situ measurement of discharge channel erosion is useful in addressing both the lifetime and transport concerns. An in-situ measurement would allow for real-time data regarding the erosion rates at different operating points, providing a quick method for empirically anchoring any analysis geared towards lifetime qualification. Erosion rate data over a thruster s operating envelope would also be useful in the modeling of the detailed physics inside the discharge chamber. There are many different sensors and techniques that have been employed to quantify discharge channel erosion in Hall thrusters. Snapshots of the wear pattern can be obtained at regular shutdown intervals using laser profilometry. Many non-intrusive techniques of varying complexity and sensitivity have been employed to detect the time-varying presence of erosion products in the thruster plume. These include the use quartz crystal microbalances, emission spectroscopy, laser induced flourescence, and cavity ring-down spectroscopy. While these techniques can provide a very accurate picture of the level of eroded material in the thruster plume, it is more difficult to use them to determine the location from which the material was eroded. Furthermore, none of the methods cited provide a true in-situ measure of erosion at the channel surface while the thruster is in operation (i.e. none yield a continuous channel erosion measurement). A recent fundamental sensor development effort has led to a novel regression, erosion, and ablation sensor technology (REAST). The REAST sensor allows for measurement of real-time surface erosion rates at a discrete surface location. The sensor was tested using a linear Hall thruster geometry (see Fig. 1), which served as a means of producing plasma erosion of a ceramic discharge chamber. The mass flow rate, discharge voltage, and applied magnetic field strength could be varied, allowing for erosion measurements over a broad thruster operating envelope. Results are presented demonstrating the ability of the REAST sensor to capture not only the insulator erosion rates but also changes in these rates as a function of the discharge parameters.

Polzin, Kurt↗
DOE OSTI2023

An interlaboratory comparison of mid-infrared spectra acquisition: Instruments and procedures matter

Diffuse reflectance spectroscopy has been extensively employed to deliver timely and cost-effective predictions of a number of soil properties. However, although several soil spectral laboratories have been established worldwide, the distinct characteristics of instruments and operations still hamper further integration and interoperability across mid-infrared (MIR) soil spectral libraries. In this study, we conducted a large-scale ring trial experiment to understand the lab-to-lab variability of multiple MIR instruments. By developing a systematic evaluation of different mathematical treatments with modeling algorithms, including regular preprocessing and spectral standardization, we quantified and evaluated instruments' dissimilarity and how this impacts internal and shared model performance. We found that all instruments delivered good predictions when calibrated internally using the same instruments' characteristics and standard operating procedures by solely relying on regular spectral preprocessing that accounts for light scattering and multiplicative/additive effects, e.g., using standard normal variate (SNV). When performing model transfer from a large public library (the USDA NSSCKSSL MIR library) to secondary instruments, good performance was also achieved by regular preprocessing (e. g., SNV) if both instruments shared the same manufacturer. However, significant differences between the KSSL MIR library and contrasting ring trial instruments responses were evident and confirmed by a semi-unsupervised spectral clustering. For heavily contrasting setups, spectral standardization was necessary before transferring prediction models. Non-linear model types like Cubist and memory-based learning delivered more precise estimates because they seemed to be less sensitive to spectral variations than global partial least square regression. In summary, the results from this study can assist new laboratories in building spectroscopy capacity utilizing existing MIR spectral libraries and support the recent global efforts to make soil spectroscopy universally accessible with centralized or shared operating procedures.

58 GEOSCIENCES↗
NASA NTRS2004

A Method for Calculating the Probability of Successfully Completing a Rocket Propulsion Ground Test

Propulsion ground test facilities face the daily challenges of scheduling multiple customers into limited facility space and successfully completing their propulsion test projects. Due to budgetary and schedule constraints, NASA and industry customers are pushing to test more components, for less money, in a shorter period of time. As these new rocket engine component test programs are undertaken, the lack of technology maturity in the test articles, combined with pushing the test facilities capabilities to their limits, tends to lead to an increase in facility breakdowns and unsuccessful tests. Over the last five years Stennis Space Center's propulsion test facilities have performed hundreds of tests, collected thousands of seconds of test data, and broken numerous test facility and test article parts. While various initiatives have been implemented to provide better propulsion test techniques and improve the quality, reliability, and maintainability of goods and parts used in the propulsion test facilities, unexpected failures during testing still occur quite regularly due to the harsh environment in which the propulsion test facilities operate. Previous attempts at modeling the lifecycle of a propulsion component test project have met with little success. Each of the attempts suffered form incomplete or inconsistent data on which to base the models. By focusing on the actual test phase of the tests project rather than the formulation, design or construction phases of the test project, the quality and quantity of available data increases dramatically. A logistic regression model has been developed form the data collected over the last five years, allowing the probability of successfully completing a rocket propulsion component test to be calculated. A logistic regression model is a mathematical modeling approach that can be used to describe the relationship of several independent predictor variables X(sub 1), X(sub 2),..,X(sub k) to a binary or dichotomous dependent variable Y, where Y can only be one of two possible outcomes, in this case Success or Failure. Logistic regression has primarily been used in the fields of epidemiology and biomedical research, but lends itself to many other applications. As indicated the use of logistic regression is not new, however, modeling propulsion ground test facilities using logistic regression is both a new and unique application of the statistical technique. Results from the models provide project managers with insight and confidence into the affectivity of rocket engine component ground test projects. The initial success in modeling rocket propulsion ground test projects clears the way for more complex models to be developed in this area.

Messer, Bradley P.↗
NASA NTRS1978

Ride quality evaluation. IV - Models of subjective reaction to aircraft motion

The paper examines models of human reaction to the motions typically experienced on short-haul aircraft flights. Data are taken on the regularly scheduled flights of four commercial airlines - three airplanes and one helicopter. The data base consists of: (1) a series of motion recordings distributed over each flight, each including all six degrees of freedom of motion; temperature, pressure, and noise are also recorded; (2) ratings of perceived comfort and satisfaction from the passengers on each flight; (3) moment-by-moment comfort ratings from a test subject assigned to each airplane; and (4) overall comfort ratings for each flight from the test subjects. Regression models are obtained for prediction of rated comfort from rms values for six degrees of freedom of motion. It is shown that the model C = 2.1 + 17.1 T + 17.2 V (T = transverse acceleration, V = vertical acceleration) gives a good fit to the airplane data but is less acceptable for the helicopter data.

Jacobson, I. D.↗
DOE OSTI2023

Development of a Supercharged Octane Number and a Supercharged Octane Index

Gasoline knock resistance is characterized by the Research and Motor Octane Number (RON and MON), which are rated on the CFR octane rating engine at naturally aspirated conditions. However, modern automotive downsized boosted spark ignition (SI) engines generally operate at higher cylinder pressures and lower temperatures relative to the RON and MON tests. Using the naturally aspirated RON and MON ratings, the octane index (OI) characterizes the knock resistance of gasolines under boosted operation by linearly extrapolating into boosted “beyond RON” conditions via RON, MON, and a linear regression K factor. Using OI solely based on naturally aspirated RON and MON tests to extrapolate into boosted conditions can lead to significant errors in predicting boosted knock resistance between gasolines due to non-linear changes in autoignition and knocking characteristics with increasing pressure conditions. Here, a new “Supercharged Octane Number” (SON) method was developed on the CFR engine at increased intake pressures, which improved the correlation to boosted knock-limited automotive SI engine data over RON for several surrogate fuels and gasolines, including five “Co-Optima” RON 98 fuels and an E10 regular grade gasoline. Furthermore, the conventional OI was extended to a newly introduced Supercharged Octane Index (OI S ) based on SON and RON, which significantly improved the correlation to fuel knock resistance measurements from modern boosted SI engine knock-limited spark advance tests. This demonstrated the first proof of concept of a SON and OI S to better characterize a fuel’s knock resistance in modern boosted SI engines.

42 ENGINEERING↗
NASA NTRS2019

Evaluating Combinations of Sentinel-2 Data and Machine-Learning Algorithms for Mangrove Mapping in West Africa

Creating a national baseline for natural resources, such as mangrove forests, and monitoring them regularly often requires a consistent and robust methodology. With freely available satellite data archives and cloud computing resources, it is now more accessible to conduct such large-scale monitoring and assessment. Yet, few studies examine the reproducibility of such mangrove monitoring frameworks, especially in terms of generating consistent spatial extent. Our objective was to evaluate a combination of image processing approaches to classify mangrove forests along the coast of Senegal and The Gambia. We used freely available global satellite data (Sentinel-2), and cloud computing platform (Google Earth Engine) to run two machine learning algorithms, random forest (RF), and classification and regression trees (CART). We calibrated and validated the algorithms using 800 reference points collected using high-resolution images. We further re-ran 10 iterations for each algorithm, utilizing unique subsets of the initial training data. While all iterations resulted in thematic mangrove maps with over 90% accuracy, the mangrove extent ranges between 827-2807 km2 for Senegal and 245-1271 km2 for The Gambia with one outlier for each country. We further report "Places of Agreement" (PoA) to identify areas where all iterations for both methods agree (506.6 km2 and 129.6 km2 for Senegal and The Gambia, respectively), thus have a high confidence in predicting mangrove extent. While we acknowledge the time- and cost-effectiveness of such methods for the landscape managers, we recommend utilizing them with utmost caution, as well as post-classification on-the-ground checks, especially for decision making.

Mondal, Pinki↗
DOE OSTI2022

Prediction of DIII-D Pedestal Structure from Externally Controllable Parameters

The sharp increase of pressure at the edge of a high confinement mode (H-mode) plasma, the pedestal, strongly impacts overall plasma performance. Predicting the pedestal is a necessity to control and optimize tokamak operations. An experimental data-driven machine learning (ML) approach is presented that predicts the pedestal heights and widths of electron density (ne) and electron temperature (Te) profiles as well as the separatrix ne from externally controllable parameters such as the plasma shape, heating method and power, and gas puff rate and integrated gas puff. The OMFIT framework was used with DIII-D data to efficiently, robustly, and automatically build a database of pedestal parameters to train machine learning models. Database creation was enabled by the search engine tool for DIII-D data, TokSearch, which parallelizes data fetching, enabling fast searches through basic signals of thousands of DIII-D shots and selection of relevant time intervals. Principal Component Analysis (PCA) separated the database into three clusters that represent classes of plasma shapes that are regularly used in DIII-D. The most important parameters for setting the pedestal structure were plasma current (Ip), toroidal magnetic field (Bφ), neutral beam heating power (PNBI) and shaping quantities. The Deep Jointly Informed Neural Networks (DJINN) algorithm was applied to identify suitable neural network (NN) architectures that appropriately capture the features of the pedestal database. Separate NNs were implemented for each pedestal parameter, and ensembling methods were used to improve the prediction accuracy and allowed estimation of the prediction uncertainty. The pedestal predictions of the test dataset lie within the measurement uncertainties of the pedestal parameters. The NN outperformed simple Linear Regression (LR) analysis, indicating non-linear dependencies in the pedestal structure. The presented achievements illustrate a promising path for future research, using feature extraction to infer experimental trends and thereby improve pedestal models as well as deploying NN for a fast pedestal prediction in DIII-D scenario development.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2024

Open Call LDRD: Physically Informed Autoencoders for Galactic Redshift Regression

Physical constraints have been suggested to make neural network models more generalizable, act scientifically plausible, and be more data-efficient over unconstrained baselines. In this report, we present preliminary work on evaluating the effects of adding soft physical constraints to computer vision neural networks trained to estimate the conditional density of redshift on input galaxy images for the Sloan Digital Sky Survey. We introduce physically motivated soft constraint terms that are not implemented with differential or integral operators. We frame this work as a simple ablation study where the effect of including soft physical constraints is compared to an unconstrained baseline. We compare networks using standard point estimate metrics for photometric redshift estimation, as well as metrics to evaluate how faithful our conditional density estimate represents the probability over the ensemble of our test dataset. We find no evidence that the implemented soft physical constraints are more effective regularizers than augmentation.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

Prediction of DIII-D Pedestal Structure From Externally Controllable Parameters

The sharp increase of pressure at the edge of a high confinement mode (H-mode) plasma, the pedestal, strongly impacts overall plasma performance. Predicting the pedestal is a necessity to control and optimize tokamak operations. Here, an experimental data-driven machine learning (ML) approach is presented that predicts the pedestal heights and widths of electron density (n e ) and electron temperature (T e ) profiles as well as the separatrix ne from externally controllable parameters such as the plasma shape, heating method and power, and gas puff rate and integrated gas puff. The OMFIT framework was used with DIII-D data to efficiently, robustly, and automatically build a database of pedestal parameters to train machine learning models. Database creation was enabled by the search engine tool for DIII-D data, TokSearch, which parallelizes data fetching, enabling fast searches through basic signals of thousands of DIII-D shots and selection of relevant time intervals. Principal Component Analysis (PCA) separated the database into three clusters that represent classes of plasma shapes that are regularly used in DIII-D. The most important parameters for setting the pedestal structure were plasma current (I p ), toroidal magnetic field (B Φ ), neutral beam heating power (P NBI ) and shaping quantities. The Deep Jointly Informed Neural Networks (DJINN) algorithm was applied to identify suitable neural network (NN) architectures that appropriately capture the features of the pedestal database. Separate NNs were implemented for each pedestal parameter, and ensembling methods were used to improve the prediction accuracy and allowed estimation of the prediction uncertainty. The pedestal predictions of the test dataset lie within the measurement uncertainties of the pedestal parameters. The NN outperformed simple Linear Regression (LR) analysis, indicating non-linear dependencies in the pedestal structure. The presented achievements illustrate a promising path for future research, using feature extraction to infer experimental trends and thereby improve pedestal models as well as deploying NN for a fast pedestal prediction in DIII-D scenario development.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
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↗