Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Training Time”

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 1,351 records · Page 75
All sourcesNASA NTRSDOE OSTI
DOE OSTI2023

The role of stiffness in training and generalization of ResNets

Neural ordinary differential equations (NODEs) have recently regained popularity as large-depth limits of a large class of neural networks. In particular, residual neural networks (ResNets) are equivalent to an explicit Euler discretization of an underlying NODE, where the transition from one layer to the next is one time step of the discretization. The relationship between continuous and discrete neural networks has been of particular interest. Notably, analysis from the ordinary differential equation viewpoint can potentially lead to new insights for understanding the behavior of neural networks in general. In this work, we take inspiration from differential equations to define the concept of stiffness for a ResNet via the interpretation of a ResNet as the discretization of a NODE. Here, we then examine the effects of stiffness on the ability of a ResNet to generalize, via computational studies on example problems coming from climate and chemistry models. We find that penalizing stiffness does have a unique regularizing effect, but we see no benefit to penalizing stiffness over L 2 regularization (penalization of network parameter norms) in terms of predictive performance.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2020

Learning compact physics‐aware delayed photocurrent models using dynamic mode decomposition

Abstract Radiation‐induced photocurrent in semiconductor devices can be simulated using complex physics‐based models, which are accurate, but computationally expensive. This presents a challenge for implementing device characteristics in high‐level circuit simulations where it is computationally infeasible to evaluate detailed models for multiple individual circuit elements. In this work we demonstrate a procedure for learning compact delayed photocurrent models that are efficient enough to implement in large‐scale circuit simulations, but remain faithful to the underlying physics. Our approach utilizes dynamic mode decomposition (DMD), a system identification technique for learning reduced‐order discrete‐time dynamical systems from time series data based on singular value decomposition. To obtain physics‐aware device models, we simulate the excess carrier density induced by radiation pulses by solving numerically the ambipolar diffusion equation, then use the simulated internal state as training data for the DMD algorithm. Our results show that the significantly reduced‐order delayed photocurrent models obtained via this method accurately approximate the dynamics of the internal excess carrier density—which can be used to calculate the induced current at the device boundaries—while remaining compact enough to incorporate into larger circuit simulations.

Hanson, Joshua↗
DOE OSTI2026

Developing Scenario‐Based Strategies for Health, Climate, and Environmental Preparedness: The One Health, One Earth Approach

Climate change amplifies many threats to human health. Despite advances in understanding climate change dynamics and impacts, there remains a critical gap in translating scientific knowledge into equitable, and community-driven health interventions. The inaugural One Earth, One Health workshop sought to explore this gap through human-centered design exercises involving interdisciplinary researchers from climate and Earth sciences, engineering, epidemiology, microbiology, and environmental health. Although participants did not co-develop solutions with affected communities, they used stakeholder role-playing to guide ideation and lay groundwork for actionable plans. Through these methods, participants identified community needs and proposed prototype solutions to alleviate health threats exacerbated by global environmental change. Prototypes were organized around infectious diseases, extreme weather, and air quality, as illustrative themes rather than an exhaustive set of risks. Key solutions included strategies for anticipatory systems and early warning (e.g., integrating environmental signals with health data), inclusive communication and infrastructure needs for responding to extreme weather events, and integrated platforms visualizing air quality trends to support tailored, context-aware guidance beyond one-size-fits-all alerts. The workshop highlighted opportunities such as leveraging machine learning, Earth observation, and real-time surveillance to protect communities, but also noted barriers including data quality, technological redundancy, privacy, and governance challenges. Additionally, participants emphasized the need for interdisciplinary teams capable of collaborating across sectors, breaking down silos and addressing gaps in training and education. Overall, the workshop illustrates how process-driven, human-centered approaches can help surface user needs and generate testable prototype concepts, while underscoring the importance of direct community partnership for implementation.

Abadi, Azar M. [University of Alabama, Birmingham,↗
DOE OSTI2022

Efficient Loads Surrogates for Waked Turbines in an Array

Accurately and efficiently predicting wind turbine structural loading is a crucial step in wind farm design. Without considering structural loading, wind farm optimization could negatively impact turbine fatigue and ultimate loads, especially for waked and partially waked turbines, which could result in higher maintenance costs and reduced turbine lifetime. However, predicting turbine loads throughout an array is a costly step, as these quantities require time-accurate results across long time histories, which is often intractable for large array optimization. Therefore, surrogate models that link array spacing to load outputs are often used, but the surrogates are then unique to the inflow conditions and array configurations in the training library. This work develops surrogate models for many wind turbine load outputs based solely on rotor plane velocity measurements, with no required input about array configuration or freestream inflow parameters. Surrogate models were constructed for many turbine quantities of interest (QoI), considering mean, standard deviation, ultimate, and fatigue loads. In general, most QoI statistics were accurately captured, as measured by predicted vs. actual correlation coefficient, confirming the suitability of the approach. Temporal mean values of the QoI required only temporal mean measurements of the rotor plane inflow velocity. However, accurate prediction of temporal standard deviation, ultimate, and fatigue values of QoI also required temporal standard deviations of the rotor plane velocity field. Poor surrogate performance was observed when too many correlated inputs were used, such as multiple velocity components. If the fewest inflow parameters are used to construct the surrogates, the average correlation coefficient value for all output QoI statistics is 0.89. Surrogates for standard deviations and damage equivalent loads (DELs) of turbine QoIs generally had lower accuracy and tower-base and shaft load channels posed the most difficult to capture accurately. The results suggest that these surrogates could be easily paired with analytic wake models, which are frequently used for pre-construction wind farm array optimization, to account for turbine loading in addition to power production. By including the optimal inflow conditions, the surrogate accuracy can improve to an average correlation coefficient value for all output QoI statistics of 0.92. This work has established the ability to build accurate surrogates for mean, standard deviation, ultimate load, and DEL turbine QoI values based on the rotor plane inflow velocity, and identified which inflow conditions lead to greater surrogate accuracy.

17 WIND ENERGY↗
DOE OSTI2026

Generative Thermodynamic Computing

Here, we introduce a generative modeling framework for thermodynamic computing, in which structured data are synthesized from noise by the natural time evolution of a physical system governed by Langevin dynamics. While conventional diffusion models use neural networks to perform denoising, here the information needed to generate structure from noise is encoded by the dynamics of a thermodynamic system. Training proceeds by maximizing the probability with which the computer generates the reverse of a noising trajectory, which ensures that the computer generates data with minimal heat emission. We demonstrate this framework within a digital simulation of a thermodynamic computer. If realized in analog hardware, such a system would function as a generative model that produces structured samples without the need for artificially injected noise or active control of denoising.

Whitelam, Stephen [Lawrence Berkeley National Labo↗
DOE OSTI2021

Machine learning to alleviate Hubbard-model sign problems

Lattice Monte Carlo calculations of interacting systems on nonbipartite lattices exhibit an oscillatory imaginary phase known as the phase or sign problem, even at zero chemical potential. One method to alleviate the sign problem is to analytically continue the integration region of the state variables into the complex plane via holomorphic flow equations. For asymptotically large flow times, the state variables approach manifolds of constant imaginary phase known as Lefschetz thimbles. Furthermore, flowing such variables and calculating the ensuing Jacobian is a computationally demanding procedure. In this paper, we demonstrate that neural networks can be trained to parametrize suitable manifolds for this class of sign problem and drastically reduce the computational cost for different severely afflicted small volume systems. In particular, we apply our method to the Hubbard model on the triangle and tetrahedron, both of which are nonbipartite. At strong interaction strengths and modest temperatures, the tetrahedron suffers from a severe sign problem that cannot be overcome with standard reweighting techniques, while it quickly yields to our method. We benchmark our results with exact calculations and comment on future directions of this work.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗
DOE OSTI2023

Robust neural network-enhanced estimation of local primordial non-Gaussianity

When applied to the nonlinear matter distribution of the universe, neural networks have been shown to be very statistically sensitive probes of cosmological parameters, such as the linear perturbation amplitude σ 8 . However, when used as a “black box,” neural networks are not robust to baryonic uncertainty. We propose a robust architecture for constraining primordial non-Gaussianity f NL , by training a neural network to locally estimate σ 8 , and correlating these local estimates with the large-scale density field. We apply our method to N-body simulations, and show that σ⁡(f NL ) is 3.5 times better than the constraint obtained from a standard halo-based approach. Finally, we show that our method has the same robustness property as large-scale halo bias: baryonic physics can change the normalization of the estimated f NL , but cannot change whether f NL is detected.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2020

Correction for Neutral Pressure-Driven Signal in Radiated Power Measurements on Proto-MPEX

A pinhole camera using four-channel resistive bolometers was designed and implemented on the Prototype Material Plasma Exposure eXperiment (Proto-MPEX) to measure the plasma radiated power. Quantification of the radiated power is required for comprehensive power balance assessments but faces unique challenges on linear plasma devices due to weak absolute signal, compared to typical use in toroidal fusion plasmas, and high neutral pressure changes. Around 1 mTorr, the pressure-driven signal contribution was shown to overwhelm the radiated power signal in Proto-MPEX, found to be ~3 μW . Since the pressure-driven signals could not be predicted from first principle arguments, the signals were corrected using empirical models that take two experimental pressure measurements and the raw bolometer signal to reveal the radiated power signals. Four “gas-only” Proto-MPEX shots were taken, without the radio frequency (rf) plasma discharge, to train, optimize, and validate empirical models. The data included two pressure measurements at different locations along with the bolometer signal. Multiple models were tested; partial least-squares regression with a cross correlation method to account for the time offset in the pressure measurements provided the best combination of accuracy and stability. The model was validated to have a modeling error of 0.0036 V, which was lower than 10% of the expected radiated power signal. The model was tested on a different “gas-only” shot with a complete time-series data. For plasma operation shots, the corrected bolometer measurements were compared with an absolute extreme ultraviolet photodiode measurement and similarities and differences are discussed.

Lee, Seungsup↗
DOE OSTI2022

Deep Learning Denoiser

SAND2022-14744 O The Deep Learning Denoiser uses a trained deep convolutional neural network (CNN) model to decompose an input waveform into a signal and noise waveform. In the approach, the CNN provides a signal mask and a noise mask for an input signal. The Short-Time Fourier Transform (STFT) of the estimated signal is obtained by multiplying the signal mask with the STFT of the input signal. Sandia National Laboratories is a multimission laboratory managed and operated by National Technology & Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International Inc., for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-NA0003525.

Koch, Clinton↗
DOE OSTI2022

Improved Throughput and Analysis of Scratch Test Results via Automation and Machine Learning

A data analysis automation interface that incorporates machine learning (ML) has been developed to improve productivity, efficiency, and consistency in identifying and defining critical load values (or other values associated with optically identifiable characteristics) of a coating when a scratch test is performed. In this specific program, the machine learning component of the program has been trained to identify the Critical Load 2 (L C2 ) value by analyzing images of the scratch tracks created in each test. An optical examination of the scratch by a human operator is currently used to determine where this value occurs. However, the vagueness of the standard has led to varying interpretations and nonuniform usage by different operators at different laboratories where the test is implemented, resulting in multiple definitions of the desired parameter. Using a standard set of training and validation images to create the dataset, the critical load can be identified consistently amongst different laboratories using the automation interface without requiring the training of human operators. When the model was used in conjunction with an instrument manufacturer's scratch test software, the model produced accurate and repeatable results and defined L C2 values in as little as half of the time compared to a human operator. When combined with a program that automates other aspects of the scratch testing process usually conducted by a human operator, scratch testing and analysis can occur with little to no intervention from a human beyond initial setup and frees them to complete other work in the lab.

36 MATERIALS SCIENCE↗
DOE OSTI2022

Emissions mitigation technology for advanced water-lean solvent-based CO 2 capture processes

This technical final report submitted to DOE/NETL presents all the research activities performed during the entirety of DE-FE0031660 project-Emissions Mitigation Technology for Advanced Water-Lean Solvent-Based CO 2 Capture Processes which spans from October 2018 through March 2022. RTI International has been conducting studies from fundamental and operational aspects to reduce the overall amine emissions from the advanced Water-Lean Solvent (WLS) systems, specifically RTI’s Non-Aqueous Solvent (NAS). This technical final report will highlight the key findings from project which align closely to the project objectives which are: Identify the contribution of vapor loss, entrainment, and aerosols to the overall emissions of water-lean systems; Determine the significance of CO 2 capture system operating parameters to the amine emissions; Develop an emissions model based on critical operating parameters; Evaluate the effectiveness of emissions mitigation devices to reduce the amine emissions to <1 ppm under flue coal-fired flue gas; and, Determine the contribution of the ECTs to the overall CO 2 capture cost. The following are the key findings based on numerous tests using both lab-scale setups and parametric testing performed at RTI’s Bench-scale Gas Absorption System (BsGAS). During the BP1, the aerosol generation system and monitoring equipment were installed at BsGAS to produce and determine the aerosol characteristics during the NAS CO 2 capture process. The aerosol produced by this setup produced aerosols with the peak diameter and concentration of 50 micron and 1.2E10 7 cm -3 , respectively. These particle sizes and concentrations are matched to those observed in the actual coal-fired power plant flue gases and expected to be found at the absorber inlet of the CO 2 capture system. Over 1,300 hours of parametric testing have been conducted to evaluate the impact of the aerosols and operating conditions during the CO 2 capture with NAS on the overall amine emissions in the treated flue gas. At the worse condition tested, the presence of the aerosols in the flue gas could increase the overall emissions by 10X compared to the baseline emissions from NAS’s vapor pressure. CO 2 capture rate was found to be a main factor impacting the overall emissions as well as aerosol size and concentrations in the absorber off-gas. The higher CO 2 capture rate, the higher amine emissions in the treated gas. The temperature difference between the temperature bulge seen in the absorber and the water wash temperature also impacts the particle growth where the larger the temperature difference, the more amine emissions from aerosols in the treated gas. The majority of the aerosols did not grow substantially in the system, and the particle concentrations remained nearly constant between the absorber inlet and wash outlet. Only a small portion of the particles were found to grow significantly. The high efficiency demister with mesh size of 5-10 micron can be installed to remove a portion of the aerosols from the gas stream leaving the water wash. Overall, these results from parametric testing have established the emission baseline and validate our assumption on the need of emission control technologies (ECT) in order to minimize the emissions from the baseline NAS CO 2 capture process. Over 2,000 of BsGAS operating hours was used to investigate a handful of process improvements which led to a selection of the vital few changes that effectively control the amine emissions. These process improvements are lime-coated-filters for absorber gas inlet, advanced demister at the top of the absorber, a second water wash with amine recovery unit were designed, installed, and tested at BsGAS at the end of BP1. The result showed that the NAS CO 2 capture process with these additional emission control devices could lower the amine emission in the treated gas to about 1 ppm using a simulated coal-fire flue gas stream. The main contributor in lowering the amine emission came from the second water wash with amine recovery unit where the amine concentration in the scrubbing water was kept below 2 wt% through a continuous amine removal via an adsorbent bed, resulting in a low amine vapor pressure. The adsorbent bed was regenerated via a direct steam regeneration and the recover amine was returned to the absorber to minimize wastewater and makeup amine. A flue gas generation system was designed and installed during the first half of BP2 to support the emission testing using a real coal-derived flue gas. The system is capable of generating both coal- and natural gas- derived flue gases with the composition of the gaseous species highly resemble to that of the power plant flue gases. The particulates detected in the coal-derived flue gas showed the mean diameter of 1 micron. The CO 2 capture operating was then proceed using the real coal-derived flue gas where the amine emission was controlled to be about 0-3 ppm for the total run time of about 200 hours. Similar testing was conducted with natural gas-derived flue gas and the result showed a highly amine emission of 30 ppm under the total run time of 200 hours. The Principal Component Analysis (PCA) and the Partial Least Squares Projection to Latent Structures (PLS) techniques were applied to the parametric testing data to derive a multivariate statistical model. The model was validated and trained with half of the data collected, and the predictive ability of the model was evaluated using the remaining half of the data. The resulting empirical model was capable of predicting the overall emissions from the NAS process without the ECTs with ±15% accuracy (average absolute deviation, AAD) in BP1. As more emission data were obtained under the real coal-flue gas in the BP2, the model incorporated these new set of data to reflect the final process configuration, operating parameters, and amine emission. This results in the updated empirical model predicting the amine emission from the NAS CO 2 capture process with 84% goodness-of-fit (R 2 ), 85% predictability (Q 2 ), and 15% AAD. The study evaluates the use of RTI’s Non-Aqueous Solvent technology for 90% CO 2 capture from a net 650 MWe pulverized coal power plant, downstream of the flue-gas desulfurization unit. The captured CO 2 has a purity of > 95% CO 2 , and is dried, compressed to 15.3 MPa (2,215 psia), ready for sequestration. The analysis uses Case B12B from the DOE Baseline study on Bituminous Coal, Revision 4 where the Cansolv CO 2 capture plant is replaced by the RTI CO 2 Capture plant. The CO 2 capture plant has been sized to capture >90% CO 2 from flue gas derived from a net 650 MWe supercritical pulverized coal power plant. The CO 2 capture plant is equipped with emission control technologies that limits the amine emissions to < 1 ppm. Two different cases were evaluated for the technoeconomic study. The key difference between the two cases is the regenerator pressure. In Case 1, the regenerator operates at 0.195 MPa (28.3 psia), whereas in Case 2, the regenerator pressure is 0.44 MPa (64 psia) thus removing the need for the first stage of compression of the eight-stage compression train. Results from the TEA are compared against the DOE reference cases for SCPC plant with and without CO 2 Capture (Case B12A and Case B12B of the DOE Baseline study, respectively). Case 2 with CO 2 regeneration at higher pressure results in the lower cost of CO 2 capture. The total capital cost of the capture process has been estimated using 2018 dollars in Aspen Process Economic Analyzer and was estimated to be $579 MM. The capture plant operation leads to a total parasitic power loss rate of 96 MWe, resulting in a decrease in pulverized coal power plant efficiency of 7.8% points. The resulting cost of electric power increases from 64.4 mills/kWh, for no capture, to 97.5 mills/kWh, with 90% capture, an increase of 51% in the COE. The cost of capturing 90% CO 2 was estimated to be $38.2/tonne-CO 2 , and meets the DOE target of $40/t-CO 2 . Emission control technologies (ECT) investigated in this project includes a second water wash with use of activated carbon beds for removal of amine from the wash water prior to recirculation in the water wash. These ECT allow operation of the CO 2 capture plant with < 1 ppm amine emissions with the treated flue gas and contributes to $2.4/t-CO 2 captured. Amine emissions derived from thermal and oxidative degradations were investigated under this project along with the emissions derived from aerosols for the NAS system. The thermally degraded of the lean NAS showed less than 4% decreased of the original total amine content in the NAS at 150 °C while the result obtained at 120 °C showed no drop in total amine content, suggesting that thermal degradation of the NAS is minimal. These results also suggested that the thermally degraded species are not likely formed and contributed to the emissions due to the low regeneration temperature of the NAS at 90-105 °C. The oxidative degradation, on the other hand, could become problematic as some of these oxidative degraded species were observed during the NAS-5 testing at National Carbon Capture Center (NCCC) and SINTEF in our previous project. The rapid screening of selected inhibitors suggested that oxidative degradation of NAS can be suppressed using thiol containing compounds in amounts of at least 1 mol%. The detailed mechanistic degradation pathway was conceived for a specific amine used in NAS formulation during BP2. he reduction of the nitrosamines caused by the NO x present in the flue gas was also examined. The study suggested that the thermo-chemical treatment of the NAS solvent would be a more effective and economically viable compared to removing NO x at the DCC.

01 COAL, LIGNITE, AND PEAT↗
DOE OSTI2022

Physics-Informed AI for Climate and Weather Risk Prediction (Final Report)

As part of the work Terrafuse developed 1) a model for wildfire risk in California, and 2) a model for downscaled wind fields from Numerical Weather Prediction (NWP) numerical models. The wildfire model is based on 20 years of historical data and captures the dependence of wildfire incidence and spread on climatic, weather and land-use variables by training a machine learning model. Nonlinear relationships between input features are learned and expressed by the model and model transparency allows features to be ranked and interpreted. The fire model is of use for accurately predicting real-time daily and long-term wildfire risk for use cases in energy and insurance. The downscaled wind model is a spatio-temporal deep learning model that emulates the influence of high-resolution variables on wind speed, allowing coarse-resolution operational NWP models to be accurately expressed on fine grids at high resolution, with application to wind energy and weather forecasting.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2025

Applications of Federated Learning in Semiconductor Manufacturing [Poster]

As semiconductor manufacturers explore advanced data analytics and modeling techniques and data hungry machine learning models increase in popularity due to their accuracy in solving generalized problems and ability to learn complex relationships, federated learning emerges as a privacy preserving machine learning technique for preserving data privacy and ensuring intellectual property protection. Federated Learning is a machine learning technique focused on training models using distributed data that never needs to be centrally stored, allowing the use of advanced machine learning techniques without compromising data privacy, and in the semiconductor manufacturing industry advanced machine learning techniques can reduce cost and time, but maintaining data privacy is essential to maintaining a competitive advantage. This paper systematically reviews existing literature on applications of federated learning in the semiconductor manufacturing industry with a focus on identifying common themes, algorithms, and gaps within the literature to drive future research directions. The findings reveal five key themes, including improvements in quality assurance, virtual models, privacy preservation, reliable data practices, and emerging trends and developments. By identifying key themes in literature on federated learning and semiconductor manufacturing and analyzing gaps and discussed methodologies, this study highlights several potential future research directions to expand the application of federated learning techniques in the semiconductor manufacturing domain.

42 ENGINEERING↗
DOE OSTI2025

Data Quality Monitoring for the Hadron Calorimeters Using Transfer Learning for Anomaly Detection

The proliferation of sensors brings an immense volume of spatio-temporal (ST) data in many domains, including monitoring, diagnostics, and prognostics applications. Data curation is a time-consuming process for a large volume of data, making it challenging and expensive to deploy data analytics platforms in new environments. Transfer learning (TL) mechanisms promise to mitigate data sparsity and model complexity by utilizing pre-trained models for a new task. Despite the triumph of TL in fields like computer vision and natural language processing, efforts on complex ST models for anomaly detection (AD) applications are limited. In this study, we present the potential of TL within the context of high-dimensional ST AD with a hybrid autoencoder architecture, incorporating convolutional, graph, and recurrent neural networks. Motivated by the need for improved model accuracy and robustness, particularly in scenarios with limited training data on systems with thousands of sensors, this research investigates the transferability of models trained on different sections of the Hadron Calorimeter of the Compact Muon Solenoid experiment at CERN. The key contributions of the study include exploring TL’s potential and limitations within the context of encoder and decoder networks, revealing insights into model initialization and training configurations that enhance performance while substantially reducing trainable parameters and mitigating data contamination effects.

47 OTHER INSTRUMENTATION↗
DOE OSTI2022

Neural Networks for Nuclear Reactions in MAESTROeX

We demonstrate the use of neural networks to accelerate the reaction steps in the MAESTROeX stellar hydrodynamics code. A traditional MAESTROeX simulation uses a stiff ODE integrator for the reactions; here, we employ a ResNet architecture and describe details relating to the architecture, training, and validation of our networks. Our customized approach includes options for the form of the loss functions, a demonstration that the use of parallel neural networks leads to increased accuracy, and a description of a perturbational approach in the training step that robustifies the model. We test our approach on millimeter-scale flames using a single-step, 3-isotope network describing the first stages of carbon fusion occurring in Type Ia supernovae. We train the neural networks using simulation data from a standard MAESTROeX simulation, and show that the resulting model can be effectively applied to different flame configurations. This work lays the groundwork for more complex networks, and iterative time-integration strategies that can leverage the efficiency of the neural networks.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2025

ML-driven Strong Lens Discoveries: Down to θ E ~ $0^{_{''}}_{^.}03$ and M halo < 10 11 M ⊙

We present results on extending the strong lens discovery space down to much smaller Einstein radii ( θ E ≲ $0^{_{''}}_{^.}03$) and much lower halo mass (M halo < 10 11 M ⊙ ) through the combination of JWST observations and machine learning (ML) techniques. First, we forecast detectable strong lenses with JWST using CosmoDC2 as the lens catalog, and a source catalog down to 29th magnitude. By further incorporating the VELA hydrodynamical simulations of high-redshift galaxies, we simulate strong lenses. We train a ResNet on these images, achieving near-100% completeness and purity for “conventional” strong lenses ( θ E ≳ $0^{_{''}}_{^.}05$), applicable to JWST, the Hubble Space Telescope (HST), the Roman Space Telescope, and Euclid VIS. For the first time, we also search for very low halo mass strong lenses (M halo < 10 11 M ⊙ ) in simulations, with θ E << $0^{_{''}}_{^.}05$, down to the best resolution ($0^{_{''}}_{^.}03$) and depth (10,000 s) limits of JWST using ResNet. A U-Net model is employed to pinpoint these small lenses in images, which are otherwise virtually impossible for human detection. Our results indicate that JWST can find ∼17/deg 2 such low-halo-mass lenses, with the locations of ∼1.1/deg 2 of these detectable by the U-Net at ∼100% precision (and ∼7.0/deg 2 at a 99.0% precision). To validate our model for finding “conventional” strong lenses, we apply it to HST images, discovering two new strong lens candidates previously missed by human classifiers in a crowdsourcing project (E. O. Garvin et al. 2022). This study demonstrates the (potentially “superhuman”) advantages of ML combined with current and future space telescopes for detecting conventional, and especially, low-halo-mass strong lenses, which are critical for testing cold dark matter models.

Silver, Ethan [Harvard University, Cambridge, MA (↗
DOE OSTI2026

Selective Amnesia using Contrastive Subnet Erasure for Class Level Unlearning in Vision Models

We study concept-level forgetting in pretrained vision models: removing an entire semantic category so the system no longer recognizes that object in unseen images and contexts, rather than merely forgetting specific training examples. Prior work either applies blunt global projections or fine-tunes parameters, which can introduce collateral damage to unrelated features, add compute, and become unstable as forgetting strength increases. We introduce Contrastive Subnet Erasure (CSE), a training-free, encoder-centric edit that targets a compact set of channels most responsible for the class and attenuates them in a calibrated manner. The modification is algebraically folded into the subsequent layer, yielding no inference-time overhead and leaving task heads unchanged. To evaluate whether forgetting generalizes beyond the data used to specify the class, we introduce a cross dataset protocol in which the class is defined on a source dataset and performance is measured on a disjoint target dataset drawn from a different distribution with no shared images. This setup tests whether the model still fails to recognize the object when it looks different or appears in new scenes, and it helps avoid overfitting to patterns in the source dataset. Across CIFAR 10, CIFAR 100, and ImageNet under this protocol, CSE achieves stronger forgetting of the target class while better preserving non target utility than existing baselines in both single class and multi class settings. Overall, CSE provides a simple, stable, and deployment-ready mechanism for class-level unlearning in vision.

Kotevska, Olivera [ORNL] (ORCID:0000000316772243)↗
DOE OSTI2021

Convolutional neural network based non-iterative reconstruction for accelerating neutron tomography *

Abstract Neutron computed tomography (NCT), a 3D non-destructive characterization technique, is carried out at nuclear reactor or spallation neutron source-based user facilities. Because neutrons are not severely attenuated by heavy elements and are sensitive to light elements like hydrogen, neutron radiography and computed tomography offer a complementary contrast to x-ray CT conducted at a synchrotron user facility. However, compared to synchrotron x-ray CT, the acquisition time for an NCT scan can be orders of magnitude higher due to lower source flux, low detector efficiency and the need to collect a large number of projection images for a high-quality reconstruction when using conventional algorithms. As a result of the long scan times for NCT, the number and type of experiments that can be conducted at a user facility is severely restricted. Recently, several deep convolutional neural network (DCNN) based algorithms have been introduced in the context of accelerating CT scans that can enable high quality reconstructions from sparse-view data. In this paper, we introduce DCNN algorithms to obtain high-quality reconstructions from sparse-view and low signal-to-noise ratio NCT data-sets thereby enabling accelerated scans. Our method is based on the supervised learning strategy of training a DCNN to map a low-quality reconstruction from sparse-view data to a higher quality reconstruction. Specifically, we evaluate the performance of two popular DCNN architectures—one based on using patches for training and the other on using the full images for training. We observe that both the DCNN architectures offer improvements in performance over classical multi-layer perceptron as well as conventional CT reconstruction algorithms. Our results illustrate that the DCNN can be a powerful tool to obtain high-quality NCT reconstructions from sparse-view data thereby enabling accelerated NCT scans for increasing user-facility throughput or enabling high-resolution time-resolved NCT scans.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗