Search NASA⌕ Search

SEARCH · Search NASA

Results for “machine learning uncertainty”

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 163 records · Page 9
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Parameter uncertainties for imperfect surrogate models in the low-noise regime

Abstract Bayesian regression determines model parameters by minimizing the expected loss, an upper bound to the true generalization error. However, this loss ignores model form error, or misspecification, meaning parameter uncertainties are significantly underestimated and vanish in the large data limit. As misspecification is the main source of uncertainty for surrogate models of low-noise calculations, such as those arising in atomistic simulation, predictive uncertainties are systematically underestimated. We analyze the true generalization error of misspecified, near-deterministic surrogate models, a regime of broad relevance in science and engineering. We show that posterior parameter distributions must cover every training point to avoid a divergence in the generalization error and design a compatible ansatz which incurs minimal overhead for linear models. The approach is demonstrated on model problems before application to thousand-dimensional datasets in atomistic machine learning. Our efficient misspecification-aware scheme gives accurate prediction and bounding of test errors in terms of parameter uncertainties, allowing this important source of uncertainty to be incorporated in multi-scale computational workflows.

Swinburne, Thomas D. (ORCID:0000000232554257)↗
DOE OSTI2024

Bridging the time scale in exascale computing of chemical systems (Final Technical Report)

This report summarizes the work carried out with support of the United States Department of Energy under Award DE-SC0019441. The theme of this project was to develop and apply methods that allowed for the acceleration of atomistic calculations, particularly in challenging areas such as multiphase systems, electrified interfaces, uncertainty estimation, and applications requiring chemical accuracy, which tend to be applications where simulation time is severely bottlenecked by the computational time requirements. Much of the focus was on the application of emerging machine-learning methodologies, although a wide range of methodologies were employed. This report has two major sections. The first focuses on the methodological advances themselves. Within this part, we report a number of major advances, a few examples of which are described here. We report the first machine-learning scheme for the acceleration of electronically grand-canonical calculations (that is, those applicable to electrochemistry). We report new methods of performing transfer learning, in which physics-based priors can be used to provide predictions, often with uncertainty estimates, of images well outside of training sets; we also offer ways to fine-tune these transfer-learning models. We provide a new systematic means to generate and apply minimal training data sets to very large (10,000’s of atoms) systems, with only small training sets appropriate for electronic structure. We developed new methodologies to integrate surface vibrations into surface adsorption calculations. We made advances to the applicability of diffusion Monte Carlo methods to allow (learned) force prediction, finite-size error correction, and force-free means of searching for transition states. We integrated machine-learned atomistic predictions into mechanism generation codes. Additionally, we released new software including AmpTorch, a modernized version of our original atomistic machine-learning code Amp. The second part of this report focuses on the scientific applications that accompanied, and were often enabled by, the methodological advances described earlier. A few examples follow, but full details are in the individual chapters of the report. For example, we developed a general theory of phonon-induced friction on molecular adsorbates. We showed fundamentally how solvent influences the adsorption and desorption process and how it differs from the processes typically involved at the solid–gas interface, making aqueous-phase and electrocatalysis different from traditional thermocatalysis. We examined how metal–insulator and magnetic transitions can be probed, and accelerated exciton dynamics via Frenkel Hamiltonian parameters. We showed that the nearsighted force-training approach, developed within this project, can predict both the stability and reactivity of large nanoparticles, and can also lead to insights on catalyst coverage on binding energies and entropies. These applied studies, which generally integrated with our method development, allowed us to push forward the theoretical understanding of several reaction classes.

08 HYDROGEN↗
DOE OSTI2024

Electronic structure prediction of multi-million atom systems through uncertainty quantification enabled transfer learning

The ground state electron density — obtainable using Kohn-Sham Density Functional Theory (KS-DFT) simulations — contains a wealth of material information, making its prediction via machine learning (ML) models attractive. However, the computational expense of KS-DFT scales cubically with system size which tends to stymie training data generation, making it difficult to develop quantifiably accurate ML models that are applicable across many scales and system configurations. Here, we address this fundamental challenge by employing transfer learning to leverage the multi-scale nature of the training data, while comprehensively sampling system configurations using thermalization. Our ML models are less reliant on heuristics, and being based on Bayesian neural networks, enable uncertainty quantification. We show that our models incur significantly lower data generation costs while allowing confident — and when verifiable, accurate — predictions for a wide variety of bulk systems well beyond training, including systems with defects, different alloy compositions, and at multi-million-atom scales. Moreover, such predictions can be carried out using only modest computational resources.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2021

AI-Batt (Autonomous Identification of Battery Life Models) [SWR 21-36]

Autonomous Identification of Battery Life Models (AI-Batt) AI-Batt is a MATLAB code base for developing lifetime models for batteries from accelerated aging data. The code base provides many functions for processing, visualizing, and modeling battery aging data, making the data processing, exploration, and modeling workflow substantially faster. These tools are tailored for working with battery aging data sets, which usually consist of many separate time-series for each cell, with many test conditions and possible replicates at each condition, which makes it difficult to simply process or visualize the data set. Complex modeling tasks, such as cross-validation, sensitivity analysis, and uncertainty quantification have been implemented to enable thorough statistical investigation of model predictions. Additionally, several machine-learning algorithms are implemented to autonomously identify suitable models via symbolic regression. Data processing functions automatically cast data from the struct data type, which is commonly used to store experimental data, but is not an acceptable input for most algorithms, to the table data type, which can be easily used as input to any optimization algorithm. Also, the data can be separated into time-invariant and time-variant data tables, which is helpful for exploring the data set as well as developing separate models for time-variant and time-invariant aging mechanisms. For example, in aging tests with constant temperature, temperature is a time-invariant experimental condition. Visualization tools enable plotting of data, model fits, and model simulations possible with single-line function calls, empowering data exploration of complex data sets with both time-varying and time-invariant trends. Plots can be automatically generated for the whole data set, or separated by data group (groups of test replicates) or individual data series. Data points or data series can be automatically colored by the value of a variable with a variety of color maps, and model predictions can also be colored by the value of a fit statistic. Comparisons between data sets and the predictions/simulations of different models on the same data set can be easily plotted as well. Distributions of parameter values from bootstrap resampling can be plotted to visualize the reliability of parameter estimation, or determine any correlations between parameters. Modeling tools handle the complex task of creating and parsing symbolic equations for modeling battery lifetime. Equations are parsed to grab relevant data variables, parameter values, or specified sub-models for input into optimization, evaluation, or simulation functions. Models can be optimized locally (one set of parameters for each data series), bi-level (some parameters shared across the data set), or globally (single set of parameters for all data). Functions implementing symbolic regression algorithms help users to discover effective model equations, even in poorly sampled, high-dimensional data.

Smith, Kandler [National Renewable Energy Lab. (NR↗
DOE OSTI2025

Public Reference Data for Megawatt-Scale Hydrogen Electrolysis - Simulated Wave

The U.S. Department of Energy and the National Laboratory of the Rockies (NLR) demonstrate hydrogen electrolysis, hydrogen compression and storage, and variable hydrogen fuel cell power production using megawatt-scale equipment at NLR’s Flatirons Campus as part of the Advanced Research on Integrated Energy Systems (ARIES) initiative. This dataset represents part of that effort and is intended for academic, national laboratory, industrial, and other stakeholders to plan, design, and validate models of megawatt-scale hydrogen technologies and diverse energy infrastructure nationwide. These data provide a baseline for how existing hydrogen electrolysis technologies perform when coupled with various energy technologies. Future datasets will demonstrate how existing hydrogen fuel cell technologies can provide controllable, dispatchable, and variable power output for artificial intelligence (AI) data centers and other variable loads. This dataset entry describes hydrogen production using a single, simulated wave energy conversion device. The electrolyzer is a 1.25-MW proton exchange membrane type MC250 system manufactured by Nel Hydrogen. While the unit supports up to 2.5 MW of electrolysis, NLR only has a single 1.25-MW electrolysis stack. For the wave energy, NLR used a wave energy converter model from PacWave. These devices can be equipped with accumulators and pressure relief values to smooth the power output by storing and releasing hydraulic energy. Using a peak power output of 10 MW, the model created two 25-minute profiles: one with and one without the accumulators and pressure relief valves. To down select the profile data from the native resolution of 20 Hz to 1 Hz, NLR took the mean of every 20 data points. NLR experimented with two simulated wave energy power plants: one that peaks at 10 MW, and one that peaks at 5 MW. These profiles were scaled for the physical 1.25 MW electrolyzer by multiplying the original profiles by one eighth and one quarter, respectively. The first profile matches the capacity rating of eight of the 1.25 MW electrolyzers, while the second matches four electrolyzers. Finally, NLR experimented with two settings for the electrolyzer power supply minimum and maximum current ramp rates (gain and slew): 200 and 400 amperes per second. The simulated profiles were translated from power (kilowatts) to current (amperes) using a curve fit with calibration data and sent to the electrolyzer power supply at 1-Hz frequency. These datasets report relevant hydrogen balance-of-plant and system data, all captured at 1 Hz, including hydrogen mass production measured with an Emerson Coriolis flow meter. Each .zip file represents a single wave electrolysis experiment and is formatted as follows: {technology}-{accumulator?}_{number of 1.25 MW electrolyzers connected}-{electrolyzer ramp rate in amperes/second} For instance, “wavePacWave-Noacc_4-400.zip” represents the 25 minute-long experiment using the PacWave’s wave energy converter model, equipped with no accumulator, connected to four 1.25-MW electrolyzers with their power supplies set to a maximum current ramp rate (gain and slew) of 400 A/s. Each .zip folder contains the following files: A .csv file containing raw data. An .xlsx file explaining all the fields in the raw data. A .png plot showing the time series of hydrogen production in kilograms per hour, electrolysis power consumption, and input wave power. An experiment, labeled “characterization_200.zip”, demonstrates the MC250 electrolyzer steady-state response with 30 minute load steps for a total duration of 5 hours. Finally, a .csv file is provided with all wave profiles combined into one dataset labeled "combined_wave_experiments.csv". NLR also built an AI/machine-learning predictive model based on these datasets. The model ingests the electrolyzer current command in amperes, as well as various pressures and temperatures across the system, and predicts hydrogen output in kilograms per hour. The complete model can be found at https://huggingface.co/NatLabRockies/ptmelt-hydrogen-electrolysis.

08 HYDROGEN↗
DOE OSTI2026

Public Reference Data for Megawatt-Scale Hydrogen Electrolysis - NLR Historical Wind

The U.S. Department of Energy and the National Laboratory of the Rockies (NLR) demonstrate hydrogen electrolysis from variable sources, hydrogen compression and storage, and hydrogen fuel cell power production using megawatt-scale equipment at NLR’s Flatirons Campus as part of the Advanced Research on Integrated Energy Systems (ARIES) initiative. This dataset represents part of that effort and is intended for academic, national laboratory, industrial, and other stakeholders to plan, design, and validate models of megawatt-scale hydrogen technologies and diverse energy infrastructure nationwide. These data provide a baseline for how existing hydrogen electrolysis technologies perform when coupled with various energy technologies. Future datasets will demonstrate how existing hydrogen fuel cell technologies can provide controllable, dispatchable, and variable power output for artificial intelligence (AI) data centers and other variable loads. This dataset entry describes hydrogen production by conducting a statistical analysis of historical wind data over a five-year period (2020-2025) from a single 1.5MW turbine manufactured by General Electric (GE) located at NLR’s Flatirons Campus, to generate an experimental test profile that was deployed on a 1.25-MW proton exchange membrane type MC250 electrolyzer system manufactured by Nel Hydrogen . [1] While the electrolyzer balance-of-plant supports up to 2.5 MW of electrolysis, NLR only has a single 1.25-MW electrolysis stack. The historical wind data provided several metrics, however, the analysis particularly focused on the measured power output by the wind turbine. The power output time series of data for each day was categorized by total energy generation and standard deviation, and the day that represented the highest combination of these two metrics was chosen – December 25th, 2022. This process was then repeated for a moving four-hour window within this day to identify the most statistically variable period. Finally, this four-hour period was scaled by 65% to match the 1.25 MW electrolyzer. The electrolysis system controls hydrogen production by varying DC current applied to the stack, from a maximum of 3000 A to a minimum safe operation of 300 A, or 10%. Because the current – voltage characteristic changes as the stack ages and efficiency degrades, the actual minimum safe operating power changes over time. The historical wind profiles were translated from power (kilowatts) to current (amperes) using a curve fit with calibration data and sent to the electrolyzer power supply at 1 Hz frequency. For more details on the statistical analysis process, see the presentation labeled “ Public Reference Data for Megawatt-Scale Hydrogen Electrolysis” provided with each data entry. These datasets report relevant hydrogen balance-of-plant and system data, all captured at 1 Hz, including hydrogen mass production measured with an Emerson Coriolis flow meter. Each .zip file represents a single wind turbine electrolysis experiment and is formatted as follows: {technology}_{scaling factor}-{electrolyzer ramp rate in amperes/second} For instance, “wind-GE1.5MW_0.65-400.zip” represents the hour-long experiment using historical data from the wind-GE1.5MW turbine, scaled to 65%, with the electrolyzer power supply set to a maximum ramp rate (gain and slew) of 400 A/s. Each .zip folder contains the following files: A .csv file containing raw data An .xlsx file explaining all the fields in the raw data. A .png plot showing the time series of hydrogen production, electrolysis power consumption, and wind power input. A PDF file detailing the historical wind data statistical analysis used to generate the wind profile. An experiment labeled “characterization_200.zip” demonstrates the MC250 electrolyzer steady-state response with 30-minute load steps for a total duration of 5 hours. Finally, a .csv file is provided with all simulated wind experiments combined into one dataset labeled "combined_historical_wind_experiments.csv". NLR also built an AI/machine-learning predictive model based on these datasets. The model ingests the electrolyzer current command in amperes, as well as various pressures and temperatures across the system, and predicts hydrogen output in kilograms per hour. The complete model can be found at https://huggingface.co/NatLabRockies/ptmelt-hydrogen-electrolysis [1] nelhydrogen.com/product/mc-series-electrolyser .

08 HYDROGEN↗
DOE OSTI2025

Rare Lepton Decays and Differentiable Hadronization Models - From Signatures of New Physics to Data-driven Event Generation

This dissertation is partitioned into two parts: phenomenological studies focused on rare lepton decays as probes of heavy and light new physics, and the development of differentiable, data-driven hadronization models. Part I develops the phenomenology of new physics signatures stemming from rare charged lepton flavor violating decays probed by experiments at the intensity frontier. These include interactions mediated by both high-scale effective operators and light new physics, manifesting in multi-lepton final states ($\mu \to 5e$), elastic nuclear transitions ($\mu \to e$ conversion), baryon-number-violating muon capture, and time-dependent signals from ultralight dark matter ($\mu \to e \phi, \tau \to \ell \phi$). Part II develops two distinct strategies for advancing differentiable and data-driven hadronization models. One involves comprehensive reweighting frameworks for hadronization that enable efficient uncertainty estimation, facilitate parameter tuning, and interface naturally with differentiable programming paradigms. The other introduces machine-learning-based methods for extracting microscopic fragmentation dynamics directly from macroscopic observables through the deformation of existing models -- effectively providing solutions to the inverse problem of hadronization. Altogether, these studies advance the interpretability, flexibility, and precision of theoretical predictions for both high-intensity and high-energy experiments.

Menzo, Tony [Cincinnati U.] (ORCID:000000022013457↗
DOE OSTI2025

Toward equitable environmental exposure modeling through convergence of data, open, and citizen sciences: an example of air pollution exposure modeling amidst increasing wildfire smoke

Exposure modeling is critical in environmental epidemiology and human health but may face challenges (e.g., skewed data, unequal error, context-insensitive validation, and computational demands). Modeling decisions reflect the intended use of the models and the values that modelers prioritize. We aimed to provide a conceptual framework and machine learning (ML) modeling protocols that address these issues. With 500m-gridded hourly PM 2.5 and O 3 levels in Illinois before, during, and after the 2023 Canadian wildfire season as a motivating example, we conducted modeling experiments to evaluate modeling methods, guided by three domains we propose based on theories of science: 1) Data Diversity, leveraging open and citizen science data to enhance inclusivity, parsimony, and representativeness; 2) Equitable Accuracy, ensuring fairly distributed uncertainties across subpopulations; and 3) Sustainable Modeling, balancing accuracy with reducing computational demands to promote accessibility for under-resourced researchers. Here, we found that ML with publicly available data can achieve high accuracy. Depending on methods, performance may vary substantially, even with identical input data. Large but skewed data may reduce performance. Misuse of cross-validation protocols can underestimate prediction error; although we observed R 2 s of ∼98 %, the modeled estimates varied significantly, indicating the need for careful model validation. By using new modeling protocols including representativeness-considered training and validation data and a new loss function, we achieved high agreement between estimates and ground-based measurements (e.g., R 2 = ∼90 % for PM 2.5 ; ∼80 % for O 3 ), equally distributed errors across sociodemographic strata and urban–rural divides, and reduction in computation time—from several weeks or months to a few days.

Exposure assessment↗
DOE OSTI2024

AEOLUS: Advances in Experimental Design, Optimal Control, and Learning for Uncertain Complex Systems

Sustained advances in the mathematics of modeling and simulation have resulted in the capability today for routine simulation of a number of large scale complex DOE-relevant systems. As remarkable as this capability for solving the so-called forward problem is, it is typically only the first step-an inner loop within an outer loop that explores the simulation model's parameter space and decision space to characterize uncertainty in the model's predictions, learn unknown model parameters from data, design the most informative experiments, determine optimal control strategies, and create optimal designs. Broadly, what unifies all of these outer loop problems is that they are, in one form or another, optimization problems over parameter/control/design space that are constrained by complex uncertain models. To fully realize the power of scientific simulation as a basis for scientific discovery, technological innovation, and rational decision-making, it is imperative to move beyond simulation to tackle the outer loop of optimization for learning from data, experimental design, and control with complex uncertain models. When the models under consideration are large-scale and complex, and when the optimization variable and uncertain parameter spaces are high (or infinite) dimensional, this constitutes a grand challenge of the highest order, and is intractable with conventional methods. To overcome these challenges, the AEOLUS Center was established to develop a unified mathematical, computational, and statistical framework for (1) Learning predictive models from complex data via Bayesian inference and optimization, and (2) Optimizing experiments, processes, and designs using the resulting uncertain models. These problems are intractable with conventional methods, for several reasons: (1) The simulation problems that govern the inner loops of the optimization problems are expensive to execute (due to severe nonlinearity, heterogeneity, multiphysics/multiscale coupling); (2) The optimization variable and uncertain parameter spaces are high dimensional, often stemming from discretizations of infinite dimensional fields such as initial conditions, sources, or material properties. We argue that the key to overcoming these challenges is to develop new mathematical, computational, and statistical methods that exploit the structure of the Bayesian inference and optimization problems mediated by their underlying complex uncertain models. This structure includes the regularity, sparsity, geometry, low intrinsic dimensionality, and multifidelity nature of the maps from uncertain parameter/optimization variable spaces to the specific objectives targeted: Bayesian inference, optimal experimental design, and optimal control design. Black box methods developed as generic tools are incapable of exploiting this structure. To be successful, we must create, integrate, and cross-fertilize ideas across multiple areas of applied math--including approximation theory, Bayesian inference, data science, experimental design, information theory, machine learning, model reduction, optimal control theory, parallel algorithms, PDE-constrained optimization, randomized algorithms, stochastic optimization, and uncertainty quantification--all while exploiting the structure of the problems at hand. With this goal in mind, we have marshaled a team of leading authorities in these areas. While the methods we develop will be broadly applicable across a wide spectrum of DOE problems in which experiments inform models and the systems those models describe must be optimized under uncertainty, we have chosen a specific area, advanced manufacturing and materials, to drive our work. AMM is characterized by complex models across multiple scales, and is a rich source of challenging problems in inference, experimental design, and optimal control, requiring multifaceted and integrated advances in applied mathematics. As such, AMM serves as an excellent vehicle to motivate and demonstrate the advances in applied mathematics developed by our center.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Polarized and unpolarized gluon PDFs: Generative machine learning applications for lattice QCD matrix elements at short distance and large momentum

Lattice quantum chromodynamics (QCD) calculations share a defining challenge by requiring a small finite range of spatial separation z between quark/gluon bilinears for controllable power corrections in the perturbative QCD factorization, and a large hadron boost p z for a successful determination of collinear parton distribution functions (PDFs). However, these two requirements make the determination of PDFs from lattice data very challenging. We present the application of generative machine learning algorithms to estimate the polarized and unpolarized gluon correlation functions utilizing short-distance data and extending the correlation up to z p z ≲ 14 , surpassing the current capabilities of lattice QCD calculations. We train physics-informed machine learning algorithms to learn from the short-distance correlation at z ≲ 0.36 fm and take the limit, p z → ∞ , thereby minimizing possible contamination from the higher-twist effects for a successful reconstruction of the polarized gluon PDF. We also expose the bias and problems with underestimating uncertainties associated with the use of model-dependent and overly constrained functional forms, such as x α ( 1 − x ) β and its variants to extract PDFs from the lattice data. We propose the use of generative machine learning algorithms to mitigate these issues and present our determination of the polarized and unpolarized gluon PDFs in the nucleon. Published by the American Physical Society 2025

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2024

Denoising diffusion probabilistic models for generative alloy design

Inverse material design is an extremely challenging optimization task made difficult by, in part, the highly nonlinear relationship linking performance with composition. Quantitative approaches have improved significantly owing to advances in high throughput experimentation and computational thermodynamics. However, existing physics-based tools are mostly forward models; input a chemistry and obtain a prediction. More recently the materials community has leveraged advances in the machine learning community to establish novel inverse design frameworks. Very recently denoising diffusion probabilistic models have been shown to be extremely powerful generators producing synthetic data of various modalities e.g. images, text, audio, tables, etc.. In this work a novel framework for alloy design and optimization is proposed leveraging these class of models. Five key generative tasks are demonstrated (1) unconditional generation (2) composition conditioned generation (3) property conditioned generation (4) multi-feedstock conditioned generation and (5) generative optimization. These methods were tested on three case studies: high entropy alloy design, superalloy binder jet additive manufacturing, and in-situ dual-feedstock wire-arc additive manufacturing. Results indicate that the established models are extremely flexible, expressive, and robust. The architecture’s flexibility and training procedure empower the model to learn complex intra-compositional and composition-property relationships. Furthermore, the probabilistic nature of these models makes them well suited for addressing solution non-uniqueness and tackling uncertainty quantification tasks. While the fidelity and quantity of the underlying training data is paramount, we envision that future alloy design frameworks will make extensive use of these kinds of machine learning models as “search” tools bolstering the utility of experimental and computational approaches.

36 MATERIALS SCIENCE↗
DOE OSTI2024

Accelerating the Discovery of New, Single Phase High Entropy Ceramics via Active Learning

High-entropy ceramics have garnered interest due to their remarkable hardness, compressive strength, thermal stability, and fracture toughness; yet the discovery of new high-entropy ceramics (out of a tremendous number of possible elemental permutations) still largely requires costly, inefficient, trial-and-error experimental and computational approaches. The entropy forming ability (EFA) factor was recently proposed as a computational descriptor that positively correlates with the likelihood that a 5-metal high-entropy carbide (HECs) will form the desired single phase, homogeneous solid solution; however, discovery of new compositions is computationally expensive. If you consider 8 candidate metals, the HEC EFA approach uses 49 optimizations for each of the 56 unique 5-metal carbides, requiring a total of 2744 costly density functional theory calculations. Here, we describe an orders-of-magnitude more efficient active learning (AL) approach for identifying novel HECs. To begin, we compared numerous methods for generating composition-based feature vectors (e.g., magpie and mat2vec), deployed an ensemble of machine learning (ML) models to generate an average and distribution of predictions, and then utilized the distribution as an uncertainty. Here we then deployed an AL approach to extract new training data points where the ensemble of ML models predicted a high EFA value or was uncertain of the prediction. Our approach has the combined benefit of decreasing the amount of training data required to reach acceptable prediction qualities and biases the predictions toward identifying HECs with the desired high EFA values, which are tentatively correlated with the formation of single phase HECs. Using this approach, we increased the number of 5-metal carbides screened from 56 to 15,504, revealing 4 compositions with record-high EFA values that were previously unreported in the literature. Our AL framework is also generalizable and could be modified to rationally predict optimized candidate materials/combinations with a wide range of desired properties (e.g., mechanical stability, thermal conductivity).

36 MATERIALS SCIENCE↗
DOE OSTI2025

LTAU-FF: Loss Trajectory Analysis for Uncertainty in atomistic Force Fields

Model ensembles are effective tools for estimating prediction uncertainty in deep learning atomistic force fields. However, their widespread adoption is hindered by high computational costs and overconfident error estimates. In this work, we address these challenges by leveraging distributions of per-sample errors obtained during training and employing a distance-based similarity search in the model latent space. Our method, which we call LTAU (Loss Trajectory Analysis for Uncertainty), efficiently estimates the full probability distribution function of errors for any test point using the logged training errors, achieving speeds that are 2–3 orders of magnitudes faster than typical ensemble methods and allowing it to be used for tasks where training or evaluating multiple models would be infeasible. We apply LTAU towards estimating parametric uncertainty in atomistic force fields (LTAU-FF), demonstrating that it produces well-calibrated confidence intervals and predicts errors that correlate strongly with the true errors for data near the training domain. Furthermore, we show that the errors predicted by LTAU-FF can be used in practical applications for detecting out-of-domain data, tuning model performance, and predicting failure during simulations. We believe that LTAU will be a valuable tool for uncertainty quantification in atomistic force fields and is a promising method that should be further explored in other domains of machine learning.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Computational materials reliability assessment of hydrogen fueled gas turbine power generation engines

The use of blended fuel sources in land based gas turbine engines drives variations in the resulting operational profile (temperatures and pressures) which can impact engine reliability. Furthermore, variability in the manufacture of components affects the resulting microstructure which directly impacts material performance and reliability. Currently, data-driven models are typically used for maintaining and inspecting fleets of engines. Without explicitly capturing material and operational sources of variability conservatism must be used in developing component-level reliability models. Therefore, there exists an opportunity to use information from materials-scale physics models to better inform reliability modeling and reduce conservatism; the impact is more cost-efficient operation and maintenance of current and future fleets. Specifically, this work establishes a computational framework for evaluating the probabilistic high temperature creep performance of hot-section Ni-based superalloys where uncertainty comes from both microstructural and operational variability. A novel high-fidelity physics model which phenomenologically captures grain-boundary sensitive phenomena has been established. A probabilistic calibration procedure was used to calibrate the model and capture uncertainty in the parameterized model coefficients. A design of experiments methodology was established for identifying informative microstructural digital representations for suitable for forward model evaluation. Results show that training a machine-learning surrogate using this design criteria outperforms random selection of microstructural representations. Finally, two surrogate models were developed: (1) a deterministic surrogate model which predicts the local field response given microstructure, constitutive model parameters, and operating conditions (stress, temperature) and (2) a probabilistic model, where uncertainty comes from constitutive law uncertainty, built using denoising diffusion probabilistic models which samples responses given (1) microstructure and (2) operating conditions. These surrogate models enable partner Siemens Energy to rapidly perform UQ analysis specific to creep deformation across a range of microstructures and operating conditions. The impact is that these ML and physics codes can be used to establish more advanced reliability models for the inspection, servicing, and maintenance of land based gas turbine engines.

36 MATERIALS SCIENCE↗
DOE OSTI2026

Quantifying mean, variability, and uncertainty in indoor radon exposure in Pennsylvania using random forest and quantile regression forest models

Radon is a naturally occurring radioactive gas that poses a serious health risk as the primary cause of lung cancer in non-smokers. Despite the well-known adverse association with health outcomes, current radon exposure assessments are limited to county-level or average-level estimates, which fail to capture regional variability. This study uses Machine Learning models, including Random Forest (RF) and Quantile Regression Forest (QRF), to estimate the indoor radon concentrations at the ZCTA (Zip code tabulation area)-level and characterize uncertainties in model estimates. Incorporating geological, meteorological, and building-specific data, the models aim to improve radon risk assessment by capturing mean exposure, variability, and extreme concentration levels. Processed radon test data (n = 718,111) were analyzed using average, variability, and quantile prediction methods. Models that estimate the average radon exposure at the ZCTA-level can yield promising model-fit results, but they do not capture the underlying variability of indoor radon exposure within a ZCTA. We utilize volatility analyses to identify characteristics indicative of high variability of indoor radon exposure. We also show that a QRF model can be used to estimate upper quantiles of residential radon exposure, thereby uncovering localized areas of elevated exposure that were not apparent in mean estimates. The results highlighted the need for a deep characterization of exposure risk and show that regions with moderate average exposure levels could still harbor extreme outliers with implications for evaluating health risks. Utilizing multiple radon exposure models allows for a deeper characterization of radon risk within a geographic area and can better identify high-risk areas. The results from this study provide a foundation for developing mitigation strategies and examining associations between radon exposure and health outcomes at fine scales. Future research should extend the geographic scope and incorporate additional environmental risk factors to establish a comprehensive framework for risk assessment.

Lee, Heechan [ORNL]↗
DOE OSTI2025

Multi-head physics-informed neural networks for learning functional priors and uncertainty quantification

In numerous applications, the integration of prior knowledge and historical information is essential, particularly for tasks requiring the solution of ordinary or partial differential equations (ODEs/PDEs) in data-sparse or noisy environments. For instance, achieving accurate solutions to time-dependent PDEs with limited initial condition measurements necessitates an effective strategy for embedding prior knowledge. Hard-parameter sharing architectures in neural networks (NNs) have demonstrated success in both traditional and scientific machine learning domains, facilitating the learning of informative representations. Here, in this study, we introduce a novel, yet efficient, method to enhance physics-informed neural networks (PINNs) by incorporating a multi-head structure that enables the learning of functional priors from both empirical data and governing physical laws. This prior information can then be used to address data sparsity and high-level noise in solving ODE/PDE problems with uncertainty quantification (UQ). The approach, termed Multi-Head PINN (MH-PINN), consists of a shared body NN and multiple head NNs, each corresponding to an individual PINN instance. Our framework for functional prior learning is carried out in two stages: (1) training the MH-PINNs to develop a shared body NN alongside multiple head NNs, and (2) employing these trained head NNs to estimate a prior distribution through a normalizing flow-based density estimator. The learned functional prior can then be applied as a regularization mechanism in deterministic contexts or as an informative prior within a Bayesian inference framework, aiding in the resolution of subsequent ODE/PDE tasks. We evaluate the efficacy of MH-PINNs across five benchmark problems, including a high-dimensional parametric PDE, all characterized by data sparsity or substantial noise levels. Our findings reveal that MH-PINNs deliver accurate solutions and robust UQ, demonstrating adaptability across a range of complex and challenging scenarios.

Bayesian inference↗
DOE OSTI2026

Low Activity Waste Glass Optimization with Property Models from Machine Learning, Part 2: Experimental Validation and Active Learning

The United States Department of Energy is responsible for managing legacy nuclear waste stored in underground tanks at the Hanford Site. To treat the waste, it is planned as the current baseline to separately vitrify low-activity waste (LAW) and high-level waste fractions. Previously, machine learning (ML) based glass property models (e.g., chemical durability, viscosity, electrical conductivity and SO3 solubility) were developed with prediction uncertainties. A waste glass optimization approach was then established to enable the capability of using these ML models in LAW glass formulation. In this study, the previous ML models were first experimentally validated, and the results were incorporated back into the database to update the ML models. The updated models and formulations showed increased waste loading while reducing the failure rate, demonstrating improved predictive accuracy, reduced uncertainties, and the effectiveness of active learning in guiding high-dimensional, nonlinear LAW glass design. This represents the first experimental validation of ML based LAW glass formulation, with practical benefits such as higher waste loading, shorter mission duration, and lower operational risk.

Lu, Xiaonan (ORCID:0000000179708148)↗
DOE OSTI2025

Spatiotemporal Downscaling Model for Solar Irradiance Forecast Using Nearest-Neighbor Random Forest and Gaussian Process

Accurate solar photovoltaic (PV) capacity estimation requires high-resolution, site-specific solar irradiance data to account for localized variability. However, global datasets, such as the National Solar Radiation Database (NSRDB), provide regional averages that fail to capture the fine-scale fluctuations critical for large-scale grid integration. This limitation is particularly relevant in the context of increasing distributed energy resources (DERs) penetration, such as rooftop PV. Additionally, it is critical to the implementation of the U.S. Federal Energy Regulatory Commission (FERC) Order 2222, which facilitates DER participation in U.S. bulk power markets. To address this challenge, this study evaluates Nearest-Neighbor Random Forest (NNRF) and Nearest-Neighbor Gaussian Process (NNGP) models for spatiotemporal downscaling of global solar irradiance data. By leveraging historical irradiance and meteorological data, these models incorporate spatial, temporal, and feature-based correlations to enhance local irradiance predictions. The NNRF model, a machine-learning approach, prioritizes computational efficiency and predictive accuracy, while the NNGP model offers a level of interpretability and prediction uncertainty by numerically quantifying correlations and dependencies in the data. Model validation was conducted using day-ahead predictions. The results showed that the average Goodness of Fit (GoF) of the NNRF model of 90.61% across all eight sites outperformed the GoF of the NNGP of 85.88%. Additionally, the computational speed of NNRF was 2.5 times faster than the NNGP. Finally, the NNGP displayed polynomial scaling while the NNRF scaled linearly with increasing number of nearest neighbors. Additional validation of the model on five sites in Puerto Rico further confirmed the superiority of the NNRF model over the NNGP model. These findings highlight the robustness and computational efficiency of NNRF for large-scale solar irradiance downscaling, making it a strong candidate for improving PV capacity estimation and real-time electricity market integration for DERs.

Asiedu, Shadrack (ORCID:0009000646004826)↗