Search NASA⌕ Search

SEARCH · Search NASA

Results for “Neural Network Potentials”

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 649 records · Page 36
All sourcesNASA NTRSDOE OSTI
NASA NTRSDate not supplied

Unlocking the Mysteries of the Moon’s Shadowed Regions

The Moon poles host large quantities of water-ice deposits in the permanently shadowed regions (PSRs), which are vital for enabling sustainable human space exploration, making these regions high-priority targets for upcoming Artemis missions [1]. Unfortunately, today, the best available orbital lunar imagery [2, 3] lacks the meter-scale resolution and signal needed to understand the geomorphology and trafficability of PSRs, complicating the planning and execution of future missions seeking to explore PSRs. We have developed an image enhancement tool called HORUS (Hyper-effective nOise Removal Unet Software) [4, 5], designed to enhance LRO Narrow-Angle Camera (NAC) optical low-light imagery of permanently shadowed regions by effectively removing the CCD-related, photon, and other residual noises that corrupt the images. The tool is composed of two deep learning neural networks trained on environmental metadata and real and synthetic imagery, the latter generated by a physical noise model (LROC). We demonstrated that HORUS effectively produces low-noise, high-resolution images (~1.5m/px), achieving a 5 to 10x improvement over existing long-exposure images of PSRs. HORUS allows scientists and engineers to identify geomorphic features (e.g., craters and boulders) in shadowed regions as small as 3 meters across as well as to peek inside of small shadowed regions, for the first time. The tool was deployed and thoroughly validated for NASA's VIPER mission [6], where it was applied to 20 candidate target regions across the lunar South Pole. Additionally, we conducted different approaches to validate the resulting HORUS-processed images. With HORUS denoised images, VIPER scientists can increase their confidence on what surface features (previously unseen) exist in the shadowed regions, helping them plan rover traverses more safely and efficiently (e.g., Fig. 1) In this manuscript, we will describe how VIPER scientists are utilizing HORUS denoised images to extract new information from the terrain and increase their confidence in what surface features exist in the shadowed regions. In combination with other high-resolution images and digital elevation maps, HORUS images are helping the team analyze potential lading and science sites, as well as planning traverses more safely and efficiently (e.g., Fig. 1). Additionally, we will describe how HORUS tool unlocks a broad range of scientific and exploration applications to other Artemis and CPLS missions to the lunar poles, including (but not limited to) geomorphic analysis, change detection, surface hazard detection, and terrain relative navigation.

artificial intelligence↗
DOE OSTI2025

Deep Learning–Assisted Multiobjective Optimization of Geological CO 2 Storage Performance under Geomechanical Risks

In geological CO 2 storage, designing the optimal well control strategy for CO 2 injection to maximize CO 2 storage while minimizing the associated geomechanical risks is not trivial. This challenge arises due to pressure buildup, CO 2 plume migration, the highly nonlinear nature of geomechanical responses to rock-fluid interaction, and the high computational cost associated with coupled flow and geomechanics simulations. In this paper, we introduce a novel optimization framework to address these challenges. The optimization problem is formulated as follows: maximize total CO 2 storage while minimizing geomechanical risks by adjusting the injection schedules within bounded constraints. The geomechanical risks are primarily driven by injection-induced pressure build-up, which is characterized by ground displacement and the induced microseismicity. We used the Fourier neural operator (FNO)-based deep learning model to construct surrogate models, replacing the time-consuming coupled flow and geomechanics simulations for evaluating the aforementioned objective functions. The developed surrogate models have been incorporated into a multiobjective optimization framework through a genetic algorithm to reduce the computational burden. The proposed optimization framework reduces the computational cost from approximately 2,400 hours, when using objective function evaluations based on physics-based simulations, to around 20 minutes. A set of Pareto-optimal solutions of the proposed workflow yields nontrivial optimal decisions, reducing the microseismicity potential and the vertical displacement. This Pareto front highlights the optimal trade-offs between CO 2 storage amount, safety, and ground displacement, emphasizing the need for careful optimization and management of injection strategies to achieve a balanced outcome. The novelty of this work is twofold. First, we demonstrate the importance of incorporating the minimization of the geomechanical risks as objective functions into the CO 2 storage optimization workflow to mitigate the potential risk of induced microseismicity and ground displacement. Second, we leverage the FNO-based surrogate models to optimize a real-field CO 2 storage operation.

42 ENGINEERING↗
NASA NTRS2005

Towards Rocket Engine Components with Increased Strength and Robust Operating Characteristics

High-energy rotating machines, powering liquid propellant rocket engines, are subject to various sources of high and low cycle fatigue generated by unsteady flow phenomena. Given the tremendous need for reliability in a sustainable space exploration program, a fundamental change in the design methodology for engine components is required for both launch and space based systems. A design optimization system based on neural-networks has been applied and demonstrated in the redesign of the Space Shuttle Main Engine (SSME) Low Pressure Oxidizer Turbo Pump (LPOTP) turbine nozzle. One objective of the redesign effort was to increase airfoil thickness and thus increase its strength while at the same time detuning the vane natural frequency modes from the vortex shedding frequency. The second objective was to reduce the vortex shedding amplitude. The third objective was to maintain this low shedding amplitude even in the presence of large manufacturing tolerances. All of these objectives were achieved without generating any detrimental effects on the downstream flow through the turbine, and without introducing any penalty in performance. The airfoil redesign and preliminary assessment was performed in the Exploration Technology Directorate at NASA ARC. Boeing/Rocketdyne and NASA MSFC independently performed final CFD assessments of the design. Four different CFD codes were used in this process. They include WIL DCA T/CORSAIR (NASA), FLUENT (commercial), TIDAL (Boeing Rocketdyne) and, a new family (AardvarWPhantom) of CFD analysis codes developed at NASA MSFC employing LOX fluid properties and a Generalized Equation Set formulation. Extensive aerodynamic performance analysis and stress analysis carried out at Boeing Rocketdyne and NASA MSFC indicate that the redesign objectives have been fully met. The paper presents the results of the assessment analysis and discusses the future potential of robust optimal design for rocket engine components.

Marcu, Bogdan↗
DOE OSTI2024

Accelerating Thermochemical Equilibrium Calculations for Nuclear Reactor Applications

Thermochemical properties play a key role in modeling and simulation of several key phenomena in nuclear reactors. There has been an increasing interest in incorporating CALPHAD-based formulations in multiphysics simulations including for Molten Salt Reactors where knowledge of phase evolution of the salt and the chemical potentials of various elements are of utmost importance in source term analyses and redox control. However, the size of such simulations is often limited by the high computational cost of full thermodynamic equilibrium calculations. This work discusses the current efforts aimed at accelerating thermochemical equilibrium calculations for multiphysics simulations performed using the open-source finite element / finite volume code Multiphysics Object Oriented Simulation Environment (MOOSE) [1]. While several methods have been proposed for accelerating phase equilibrium calculations [2], most focus on relatively small systems and often rely on a- priori knowledge of the state-space of the system. Nuclear materials, however, are often multi-component systems owing to the evolution of composition under irradiation and an approach based on a-priori mapping of phase diagram is often not enough. This work is aimed at demonstrating an on-the-fly surrogate modeling framework that uses active learning to reduce the number of full equilibrium calculations that must be performed. By combining with efficient coupling approaches, the surrogate framework helps in reducing the computational cost of thermodynamic equilibrium informed multiphysics simulations of nuclear materials. The performance is benchmarked against full coupling with the thermochemistry library Thermochimica [3]. This work uses a machine learning based approach for constructing surrogate models to predict the stable phases in a multicomponent system. The surrogates were constructed using neural networks and Gaussian process classification. In this work, we compare the relative performance of the two methods. We also demonstrate the use of caching previous calculations by interpolating the values from nearest neighbors. References [1] Lindsay, A.D., et al. "2.0 – MOOSE: Enabling massively parallel multiphysics simulation", SoftwareX, 20 (2022): 101202. [2] Roos, W.A. and Zietsman J.H. "Accelerating complex chemical equilibrium calculations – A Review", Calphad, 77 (2022): 102380. [3] Piro, M.H.A., et al. "The thermochemistry library Thermochimica", Computational Materials Science, 67 (2013): 266-272.

36 MATERIALS SCIENCE↗
DOE OSTI2024

Separable physics-informed DeepONet: Breaking the curse of dimensionality in physics-informed machine learning

The deep operator network (DeepONet) has shown remarkable potential in solving partial differential equations (PDEs) by mapping between infinite-dimensional function spaces using labeled datasets. However, in scenarios lacking labeled data, the physics-informed DeepONet (PI-DeepONet) approach, which utilizes the residual loss of the governing PDE to optimize the network parameters, faces significant computational challenges, particularly due to the curse of dimensionality. This limitation has hindered its application to high-dimensional problems, making even standard 3D spatial with 1D temporal problems computationally prohibitive. Additionally, the computational requirement increases exponentially with the discretization density of the domain. Here, to address these challenges and enhance scalability for high-dimensional PDEs, we introduce the Separable physics-informed DeepONet (Sep-PI-DeepONet). This framework employs a factorization technique, utilizing sub-networks for individual one-dimensional coordinates, thereby reducing the number of forward passes and the size of the Jacobian matrix required for gradient computations. By incorporating forward-mode automatic differentiation (AD), we further optimize computational efficiency, achieving linear scaling of computational cost with discretization density and dimensionality, making our approach highly suitable for high-dimensional PDEs. We demonstrate the effectiveness of Sep-PI-DeepONet through three benchmark PDE models: the viscous Burgers’ equation, Biot’s consolidation theory, and a parameterized heat equation. Our framework maintains accuracy comparable to the conventional PI-DeepONet while reducing training time by two orders of magnitude. Notably, for the heat equation solved as a 4D problem, the conventional PI-DeepONet was computationally infeasible (estimated 289.35 h), while the Sep-PI-DeepONet completed training in just 2.5 h. These results underscore the potential of Sep-PI-DeepONet in efficiently solving complex, high-dimensional PDEs, marking a significant advancement in physics-informed machine learning.

Neural operator↗
NASA NTRS2022

Multi-Mission Terrain Classifier for Safe Rover Navigation and Automated Science

We previously presented Soil Property and Object Classification (SPOC), a machine learning-based terrain classifier for Mars rovers, for automatically segmenting rover images by its surface type such as sand and bedrock. This paper presents a number of practical improvements to pave the way for potential future onboard deployment. First, we achieved 97.0% overall pixel accuracy, evaluated against the classification generated by human experts on images from Mars Science Laboratory (MSL) missions. The substantial increase in accuracy was primarily enabled by the sheer volume of data used for training; we created a new large-scale dataset of Martian terrain labels, namely AI4Mars, which contains more than 400k labels contributed by citizen scientists for 50k images taken by the Mars Exploration Rovers (MER) and Mars Science Laboratory (MSL) rover. Second, we demonstrated that SPOC can quickly adapt to a new mission landed on a previously unseen site. Specifically, we pretrained a model with MER and MSL data from the AI4Mars dataset and then adapted to the Mars 2020 Rover (M2020) by feeding a small volume of data between Sol 0 and 157; the adapted model was tested on Sol 200-203 and resulted in 84.2% overall pixel accuracy and 93.4% reliability (recall) for detecting sand, the most concerning class for rover’s traversability. Third, we found that pretraining can substantially mitigate the decline of accuracy over time. We showed that the performance of a SPOC model pretrained with the ImageNet dataset and then trained by MSL images only up to Sol 390 remains comparable to a model trained by images up to Sol 1689 on the test data after Sol 1689. Fourth, we reimplemented SPOC with a light-weight convolutional neural network (CNN), MobileNetV2, which typically runs within tens of milliseconds (ms) on mobile processors such as Qualcomm’s Snapdragon. Finally, we released the AI4Mars dataset to the public to encourage open innovation.

Ono, Masahiro↗
DOE OSTI2025

On the effectiveness of neural operators at zero-shot weather downscaling

Machine-learning (ML) methods have shown great potential for weather downscaling. These data-driven approaches provide a more efficient alternative for producing high-resolution weather datasets and forecasts compared to physics-based numerical simulations. Neural operators, which learn solution operators for a family of partial differential equations, have shown great success in scientific ML applications involving physics-driven datasets. Neural operators are grid-resolution-invariant and are often evaluated on higher grid resolutions than they are trained on, i.e., zero-shot super-resolution. Given their promising zero-shot super-resolution performance on dynamical systems emulation, we present a critical investigation of their zero-shot weather downscaling capabilities, which is when models are tasked with producing high-resolution outputs using higher upsampling factors than are seen during training. To this end, we create two realistic downscaling experiments with challenging upsampling factors (e.g., 8x and 15x) across data from different simulations: the European Centre for Medium-Range Weather Forecasts Reanalysis version 5 (ERA5) and the Wind Integration National Dataset Toolkit. While neural operator-based downscaling models perform better than interpolation and a simple convolutional baseline, we show the surprising performance of an approach that combines a powerful transformer-based model with parameter-free interpolation at zero-shot weather downscaling. We find that this Swin-Transformer-based approach mostly outperforms models with neural operator layers in terms of average error metrics, whereas an Enhanced Super-Resolution Generative Adversarial Network-based approach is better than most models in terms of capturing the physics of the ground truth data. We suggest their use in future work as strong baselines.

17 WIND ENERGY↗
NASA NTRS2015

A Module for Assimilating Hyperspectral Infrared Retrieved Profiles into the Gridpoint Statistical Interpolation System for Unique Forecasting Applications

Hyperspectral infrared sounder radiance data are assimilated into operational modeling systems however the process is computationally expensive and only approximately 1% of available data are assimilated due to data thinning as well as the fact that radiances are restricted to cloud-free fields of view. In contrast, the number of hyperspectral infrared profiles assimilated is much higher since the retrieved profiles can be assimilated in some partly cloudy scenes due to profile coupling other data, such as microwave or neural networks, as first guesses to the retrieval process. As the operational data assimilation community attempts to assimilate cloud-affected radiances, it is possible that the use of retrieved profiles might offer an alternative methodology that is less complex and more computationally efficient to solve this problem. The NASA Short-term Prediction Research and Transition (SPoRT) Center has assimilated hyperspectral infrared retrieved profiles into Weather Research and Forecasting Model (WRF) simulations using the Gridpoint Statistical Interpolation (GSI) System. Early research at SPoRT demonstrated improved initial conditions when assimilating Atmospheric Infrared Sounder (AIRS) thermodynamic profiles into WRF (using WRF-Var and assigning more appropriate error weighting to the profiles) to improve regional analysis and heavy precipitation forecasts. Successful early work has led to more recent research utilizing WRF and GSI for applications including the assimilation of AIRS profiles to improve WRF forecasts of atmospheric rivers and assimilation of AIRS, Cross-track Infrared and Microwave Sounding Suite (CrIMSS), and Infrared Atmospheric Sounding Interferometer (IASI) profiles to improve model representation of tropopause folds and associated non-convective wind events. Although more hyperspectral infrared retrieved profiles can be assimilated into model forecasts, one disadvantage is the retrieved profiles have traditionally been assigned the same error values as the rawinsonde observations when assimilated with GSI. Typically, satellitederived profile errors are larger and more difficult to quantify than traditional rawinsonde observations (especially in the boundary layer), so it is important to appropriately assign observation errors within GSI to eliminate potential spurious innovations and analysis increments that can sometimes arise when using retrieved profiles. The goal of this study is to describe modifications to the GSI source code to more appropriately assimilate hyperspectral infrared retrieved profiles and outline preliminary results that show the differences between a model simulation that assimilated the profiles as rawinsonde observations and one that assimilated the profiles in a module with the appropriate error values.

Berndt, Emily↗
DOE OSTI2024

Machine learning aided line intensity ratio method for helium–hydrogen mixed recombining plasmas

The helium line intensity ratio (LIR) with the help of a collisional radiative (CR) model has long been used to measure the electron density, n e , and temperature, T e , and its potential and limitations for fusion applications have been discussed. However, it has been reported that the CR model approach leads to deviations in helium–hydrogen mixed plasmas and/or recombining plasmas. In this study, a machine learning (ML) aided LIR method is used to measure n e and T e from spectroscopic data of helium–hydrogen mixed recombining plasmas in the divertor simulator Magnum-PSI. To analyze mixed plasmas, which have more complex spectral shapes, the spectroscopy data were used directly for training instead of separating the intensities of each line. Finally, it is shown that the ML approach can provide a robust and simpler analysis method to deduce n e and T e from the visible emissions in helium–hydrogen mixed plasmas.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
NASA NTRS2022

A New Machine Learning Based Analysis for Improving Satellite Retrieved Atmospheric Composition Data: OMI SO2 as an Example

Despite recent progress, satellite retrievals of anthropogenic SO2 still suffer from relatively low signal-tonoise ratios. In this study, we demonstrate a new machine learning data analysis method to improve the quality of satellite SO2 products. In the absence of large ground-truth datasets for SO2, we start from SO2 slant column densities (SCDs) retrieved from the Ozone Monitoring Instrument (OMI) using a data-driven, physically based algorithm and calculate the ratio between the SCD and the root mean square (rms) of the fitting residuals for each pixel. To build the training data, we select presumably clean pixels with small SCD / rms ratios (SRRs) and set their target SCDs to zero. For polluted pixels with relatively large SRRs, we set the target to the original retrieved SCDs. We then train neural networks (NNs) to reproduce the target SCDs using predictors including SRRs for individual pixels, solar zenith, viewing zenith and phase angles, scene reflectivity, and O3 column amounts, as well as the monthly mean SRRs. For data analysis, we employ two NNs: (1) one trained daily to produce analyzed SO2 SCDs for polluted pixels each day and (2) the other trained once every month to produce analyzed SCDs for less polluted pixels for the entire month. Test results for 2005 show that our method can significantly reduce noise and artifacts over background regions. Over polluted areas, the monthly mean NN-analyzed and original SCDs generally agree to within ±15 %, indicating that our method can retain SO2 signals in the original retrievals except for large volcanic eruptions. This is further confirmed by running both the NN-analyzed and original SCDs through a topdown emission algorithm to estimate the annual SO2 emissions for ∼ 500 anthropogenic sources, with the two datasets yielding similar results. We also explore two alternative approaches to the NN-based analysis method. In one, we employ a simple linear interpolation model to analyze the original SCD retrievals. In the other, we develop a PCA–NN algorithm that uses OMI measured radiances, transformed and dimension-reduced with a principal component analysis (PCA) technique, as inputs to NNs for SO2 SCD retrievals. While the linear model and the PCA–NN algorithm can reduce retrieval noise, they both underestimate SO2 over polluted areas. Overall, the results presented here demonstrate that our new data analysis method can significantly improve the quality of existing OMI SO2 retrievals. The method can potentially be adapted for other sensors and/or species and enhance the value of satellite data in air quality research and applications.

Can Li↗
DOE OSTI2025

Fundamental limit of jet tagging

Identifying the origin of high-energy hadronic jets (jet tagging) has been a critical benchmark problem for machine learning in particle physics. Jets are ubiquitous at colliders and are complex objects that serve as prototypical examples of collections of particles to be categorized. Over the last decade, machine learning-based classifiers have replaced classical observables as the state of the art in jet tagging. Increasingly complex machine learning models are leading to increasingly more effective tagger performance. Our goal is to address the question of convergence—are we getting close to the fundamental limit on jet tagging or is there still potential for computational, statistical, and physical insights for further improvements? We address this question using state-of-the-art generative models to create a realistic, synthetic dataset with a known jet tagging optimum. Various state-of-the-art taggers are deployed on this dataset, showing that there is a significant gap between their performance and the optimum. Our dataset and software are made public to provide a benchmark task for future developments in jet tagging and other areas of particle physics.

Artificial intelligence↗
NASA NTRS1996

An Acoustic Emission and Acousto-Ultrasonic Analysis of Impact Damaged Composite Pressure Vessels

The research presented herein summarizes the development of acoustic emission (AE) and acousto-ultrasonic (AU) techniques for the nondestructive evaluation of filament wound composite pressure vessels. Vessels fabricated from both graphite and kevlar fibers with an epoxy matrix were examined prior to hydroburst using AU and during hydroburst using AE. A dead weight drop apparatus featuring both blunt and sharp impactor tips was utilized to produce a single known energy 'damage' level in each of the vessels so that the degree to which the effects of impact damage could be measured. The damage levels ranged from barely visible to obvious fiber breakage and delamination. Independent neural network burst pressure prediction models were developed from a sample of each fiber/resin material system. Here, the cumulative AE amplitude distribution data collected from low level proof test (25% of the expected burst for undamaged vessels) were used to measure the effects of the impact on the residual burst pressure of the vessels. The results of the AE/neural network model for the inert propellant filled graphite/epoxy vessels 'IM7/3501-6, IM7/977-2 and IM7/8553-45' demonstrated that burst pressures can be predicted from low level AE proof test data, yielding an average error of 5.0%. The trained network for the IM7/977-2 class vessels was also able to predict the expected burst pressure of taller vessels (three times longer hoop region length) constructed of the same material and using the same manufacturing technique, with an average error of 4.9%. To a lesser extent, the burst pressure prediction models could also measure the effects of impact damage to the kevlar/epoxy 'Kevlar 49/ DPL862' vessels. Here though, due to the higher attenuation of the material, an insufficient amount of AE amplitude information was collected to generate robust network models. Although, the worst case trial errors were less than 6%, when additional blind predictions were attempted, errors as high as 50% were produced. An acousto-ultrasonic robotic evaluation system (AURES) was developed for mapping the effects of damage on filament wound pressure vessels prior to hydroproof testing. The AURES injects a single broadband ultrasonic pulse into each vessel at preprogrammed positions and records the effects of the interaction of that pulse on the material volume with a broadband receiver. A stress wave factor in the form of the energy associated with the 750 to 1000 kHz and 1000 to 1250 kHz frequency bands were used to map the potential failure sites for each vessel. The energy map associated with the graphite/epoxy vessels was found to decrease in the region of the impact damage. The kevlar vessels showed the opposite trend, with the energy values increasing around the damage/failure sites.

Walker, James L.↗
DOE OSTI2025

ELECTRIFICATION OF A HEAVY-DUTY OFF-ROAD MATERIAL HANDLER: ENERGY SAVINGS AND EMISSION REDUCTIONS

Federal regulations are driving the adoption of electrification technologies to reduce carbon dioxide equivalent (CO2e) emissions, a metric that quantifies the global warming potential of various greenhouse gases in terms of carbon dioxide (CO2). Although no specific CO2 regulations exist for heavy-duty off-road machines, future reductions are likely, given stricter emissions standards for on-road vehicles. The heavy-duty off-road sector offers significant fuel-saving potential, as its focus has traditionally been on reliability and performance rather than fuel efficiency. This dissertation examines fuel and CO2e savings opportunities on a heavy-duty off-road material handler, the Pettibone Cary-Lift 204i, from stock configuration to simple modifications to a complete teardown and reconfiguration of the machine with a plug-in series hybrid architecture using electrified hydraulics. The study begins by modeling the baseline machine’s fuel and energy consumption, calibrating with experimental data from custom operating cycles. An energy analysis identifies key areas for fuel savings. Two simple powertrain modifications result in a combined 16.2% fuel savings. Next, a Pugh-style analysis narrows a list of electrified architectures, leading to high-fidelity models that evaluate total lifetime CO2e and costs. Higher electrification levels reduce CO2e emissions but increase costs, and electricity grid emissions significantly impact CO2e for plug-in architectures. A plug-in series hybrid is chosen for the project. In its base control form, 49% fuel and 29% CO2e savings are expected from the plug-in series hybrid compared to the baseline machine. Further savings are pursued through regenerative braking (6.3%) and load-following hydraulic control (17.8%), totaling 24.1% fuel savings, and leading to a total of 61% fuel and 41% CO2e savings compared to the baseline. Battery chemistries and charging strategies are also analyzed for cost and CO2e impacts, finding LFP batteries as superior due to longevity, and overnight level 2 charging usually at a lower cost but resulting in higher emissions than opportunity DC fast-charging (DCFC). DCFC emissions are highly dependent on grid emissions, and DCFC cost is highly dependent on grid demand charges. Finally, artificial intelligence is applied to operating cycle recognition. Neural network accuracy ranges from 81% to 99%, with applications to worksite efficiency and safety improvements.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2024

Structural and mechanical properties of monolayer amorphous carbon and boron nitride

Amorphous materials exhibit various characteristics that are not featured by crystals and can sometimes be tuned by their degree of disorder (DOD). Here, we report results on the mechanical properties of monolayer amorphous carbon (MAC) and monolayer amorphous boron nitride (maBN) with different DOD. The pertinent structures are obtained by kinetic-Monte-Carlo (kMC) simulations using machine-learning potentials (MLP) with density-functional-theory (DFT)-level accuracy. An intuitive order parameter, namely the areal fraction F x occupied by crystallites within the continuous random network, is proposed to describe the DOD. We find that F x captures the essence of the DOD: Samples with the same F x but different sizes and arrangements of crystallites, obtained using two distinct kMC procedures, have virtually identical radial distributions functions as well as bond-length and bond-angle distributions. Furthermore, by simulating the fracture process with molecular dynamics, we found that the mechanical responses of MAC and maBN before fracture are mainly determined by F x and are insensitive to the sizes and specific arrangements and to some extent the numbers and area distributions of the crystallites. The behavior of cracks in the two materials is analyzed and found to mainly propagate in meandering paths in the CRN region and to be influenced by crystallites in distinct ways that toughen the material. Furthermore, the present results reveal the relation between structure and mechanical properties in amorphous monolayers and may provide a universal toughening strategy for 2D materials.

2-dimensional systems↗
DOE OSTI2025

FPGA-accelerated SpeckleNN with SNL for real-time X-ray single-particle imaging

We present the implementation of a specialized version of our previously published unified embedding model, SpeckleNN, for real-time speckle pattern classification in X-ray Single-Particle Imaging (SPI), using the SLAC Neural Network Library (SNL) on an FPGA platform. This hardware realization transitions SpeckleNN from a prototypic model into a practical edge solution, optimized for running inference near the detector in high-throughput X-ray free-electron laser (XFEL) facilities, such as those found at the Linac Coherent Light Source (LCLS). To address the resource constraints inherent in FPGAs, we developed a more specialized version of SpeckleNN. The original model, which was designed for broader classification across multiple biological samples, comprised ~5.6 million parameters. The new implementation, while reducing the parameter count to 64.6K (a 98.8% reduction), focuses on maintaining the model's essential functionality for real-time operation, achieving an accuracy of 90%. Furthermore, we compressed the latent space from 128 to 50 dimensions. This implementation was demonstrated on the KCU1500 FPGA board, utilizing 71% of available DSPs, 75% of LUTs, and 48% of FFs, with an average power consumption of 9.4W according to the Vivado post-implementation report. The FPGA performed inference on a single image with a latency of 45.015 microseconds at a 200 MHz clock rate. In comparison, running the same inference on an NVIDIA A100 GPU resulted in an average power consumption of ~73W and an image processing latency of around 400 microseconds. Our FPGA-accelerated version of SpeckleNN demonstrated significant improvements, achieving an 8.9 × speedup and a 7.8 × reduction in power consumption compared to the GPU implementation. Key advancements include model specialization and dynamic weight loading through SNL, which eliminates the need for time-consuming FPGA design re-synthesis, allowing fast and continuous deployment of models (re)trained online. These innovations enable real-time adaptive classification and efficient vetoing of speckle patterns, making SpeckleNN more suited for deployment in XFEL facilities. This implementation has the potential to significantly accelerate SPI experiments and enhance adaptability to evolving experimental conditions.

47 OTHER INSTRUMENTATION↗
DOE OSTI2024

Quadrature Based Neural Network Learning of Stochastic Hamiltonian Systems

Hamiltonian Neural Networks (HNNs) provide structure-preserving learning of Hamiltonian systems. In this paper, we extend HNNs to structure-preserving inversion of stochastic Hamiltonian systems (SHSs) from observational data. We propose the quadrature-based models according to the integral form of the SHSs’ solutions, where we denoise the loss-by-moment calculations of the solutions. The integral pattern of the models transforms the source of the essential learning error from the discrepancy between the modified Hamiltonian and the true Hamiltonian in the classical HNN models into that between the integrals and their quadrature approximations. This transforms the challenging task of deriving the relation between the modified and the true Hamiltonians from the (stochastic) Hamilton–Jacobi PDEs, into the one that only requires invoking results from the numerical quadrature theory. Meanwhile, denoising via moments calculations gives a simpler data fitting method than, e.g., via probability density fitting, which may imply better generalization ability in certain circumstances. Numerical experiments validate the proposed learning strategy on several concrete Hamiltonian systems. The experimental results show that both the learned Hamiltonian function and the predicted solution of our quadrature-based model are more accurate than that of the corrected symplectic HNN method on a harmonic oscillator, and the three-point Gaussian quadrature-based model produces higher accuracy in long-time prediction than the Kramers–Moyal method and the numerics-informed likelihood method on the stochastic Kubo oscillator as well as other two stochastic systems with non-polynomial Hamiltonian functions. Moreover, the Hamiltonian learning error εH arising from the Gaussian quadrature-based model is lower than that from Simpson’s quadrature-based model. These demonstrate the superiority of our approach in learning accuracy and long-time prediction ability compared to certain existing methods and exhibit its potential to improve learning accuracy via applying precise quadrature formulae.

Mathematics↗
DOE OSTI2025

Leveraging structure-informed machine learning for fast steric zipper propensity prediction across whole proteomes

Predicting the amyloid fold and the propensity of peptide segments to adopt amyloid-like structures remain a challenge. However, recent progress has facilitated structure-based prediction of steric zipper propensity and the use of machine learning to accelerate the calculation of predictive models across many scientific areas. Leveraging these advances, we have developed a new approach for rapid proteome-wide assessment of zipper profiles that is informed by four million steric zipper predictions collected over ten years. This collection is used to build a machine learning model capable of rapidly predicting steric zipper propensity, and allowing for the assessment of zippers at both the protein and proteome level. Our predictions show enrichment for zipper forming segments in proteins involved in cell wall reorganization in yeast, highlighting a potential category of interest for experimental characterization. Overall, our predictive model allows for the exploration of amyloid formation across the tree of life and provides a tool for assessment of both novel and designed sequences for zipper density.

Biochemistry & Molecular Biology↗
DOE OSTI2024

Application of Modified Meshgraphnets for Subsurface Prediction during CO2 Sequestration

In the face of the increasingly dire consequences of anthropogenic climate change, capturing and storing carbon dioxide is paramount. However, several impediments exist to the safe and effective subsurface storage of CO2, such as cost of transport, identification of suitable sites for subsurface storage, and assessment of long-term risk from storage in subsurface aquifers. Accurate subsurface modeling is necessary to ensure that CO2 storage is both safe and effective. Still, such modeling has traditionally required either substantial time and computational power (numerical simulation) or a substantial amount of pre-existing data for training (machine learning models). Additionally, these models lack flexibility in dealing with both changes in discretization of the input data and generalizability beyond the data on which they are trained. In order to address these issues, this research applies graph neural networks (GNNs) to predict subsurface saturation and pressure during CO₂ injection in a model of the Illinois Basin-Decatur Project (IBDP). GNNs provide a flexible, intuitive method for representing and manipulating complex unstructured data, which is often found in many practical domain problems such as fluid flow and subsurface characterization. These unstructured grids are easily represented in GNNs by representing spatially-localized features such as permeability, porosity, saturation, and pressure as nodes in a graph and relationships between these properties as edges connecting these nodes. This research applies a specific GNN model called MeshGraphNets (MGN) to model the change in CO2 saturation and pressure over a 50-month time period (36 months of injection, 14 months post-injection). The MGN model leverages a message passing process that allows the network to learn both the spatial and temporal dynamics of this system simultaneously. Additionally, training on a limited dataset (64 realizations, 20 time points each) resulted in a high degree of accuracy in saturation prediction both within the same timeframe as the training (20 months, 0.039 average RMSE) and when projecting out to the end of injection (36 months, 0.053 average RMSE). Temporal predictions such as those generated by MGNs and other similar models are prone to accumulated error over time; in order to address this, a multi-step rollout (MSR) training process was applied to calculate training loss. This method mimics the forward prediction during inference by “rolling out” multiple time points in a single training step using the previous prediction as input to the MGN model. By calculating the loss several time steps forward from the current prediction, the model is forced to find a more stable state over time. Application of MSR to the MGN model resulted in an average 15% reduction in inference error over time during forward prediction. This study showcases the immense potential of GNNs as a game-changing methodology for predicting pressure and saturation evolution in CCS projects, ultimately paving the way for more sustainable and effective carbon storage solutions. Presentation prepared for the 2024 AiChE Annual Meeting, October 27 to November 1 2024, San Diego, CA.

Holcomb, Paul↗