Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “training data”

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 1,765 records · Page 98
All sourcesNASA NTRSDOE OSTI
DOE OSTI2026

High-Fidelity Building Emulator

This dataset provides high-fidelity time series data for an emulated commercial office building sited in the Chicago, IL area during a Typical Meteorological Year (TMY). This dataset consists of air-side HVAC measurements and control inputs, and it includes normal operations as well as various implemented faults (with associated ground truth measurements) implemented on selected days. This data could be used to quantify and compare the impacts of different faults, and it could also be used as training or validation data for machine learning algorithms (e.g., reduced-order modelling, fault detection and diagnosis).

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
NASA NTRS2005

Applications of Principled Search Methods in Climate Influences and Mechanisms

Forest and grass fires cause economic losses in the billions of dollars in the U.S. alone. In addition, boreal forests constitute a large carbon store; it has been estimated that, were no burning to occur, an additional 7 gigatons of carbon would be sequestered in boreal soils each century. Effective wildfire suppression requires anticipation of locales and times for which wildfire is most probable, preferably with a two to four week forecast, so that limited resources can be efficiently deployed. The United States Forest Service (USFS), and other experts and agencies have developed several measures of fire risk combining physical principles and expert judgment, and have used them in automated procedures for forecasting fire risk. Forecasting accuracies for some fire risk indices in combination with climate and other variables have been estimated for specific locations, with the value of fire risk index variables assessed by their statistical significance in regressions. In other cases, the MAPSS forecasts [23, 241 for example, forecasting accuracy has been estimated only by simulated data. We describe alternative forecasting methods that predict fire probability by locale and time using statistical or machine learning procedures trained on historical data, and we give comparative assessments of their forecasting accuracy for one fire season year, April- October, 2003, for all U.S. Forest Service lands. Aside from providing an accuracy baseline for other forecasting methods, the results illustrate the interdependence between the statistical significance of prediction variables and the forecasting method used.

Glymour, Clark↗
DOE OSTI2019

The NOvA Test Beam Program

NOvA is a two-detector long-baseline neutrino oscillation experiment which aims to make a determination of the neutrino mass hierarchy, the octant of ${\theta}_{23}$, and measure possible CP violation. The NOvA Test Beam program consists of a scaled-down NOvA detector placed in a beamline capable of delivering 0.3 - 2.0 GeV/c protons, electrons, pions, and kaons. The beamline detectors provide us with particle identification and momentum measurements so we can study our detector technology with known inputs. Studying these particles will provide us a more detailed understanding of our calibration, detector response, and energy scale, which are some of the largest sources of systematic uncertainty in NOvA analyses. We will also collect a selection of single-particle data events for training particle identification algorithms. I will present the current status of the NOvA Test Beam program and discuss plans for data taking and analysis.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2014

TU-C-12A-02: Development of a Multiparametric Statistical Response Map for Quantitative Imaging

Purpose: Quantitative imaging biomarkers (QIB) are becoming increasingly utilized in early phase clinical trials as a means of non-invasively assessing treatment response and associated response heterogeneity. The aim of this study was to develop a flexible multiparametric statistical framework to predict voxel-by-voxel response of several potential MRI QIBs. Methods: Patients with histologically proven glioblastomas (n=11) were treated with chemoradiation (with/without bevacizumab) and underwent one baseline and two mid-treatment (3–4wks) MRIs. Dynamic contrast-enhanced (3D FSPGR, 6.3sec/phase, 0.1 mmol/kg Gd-DTPA), dynamic susceptibility contrast (2D GRE-EPI, 1.5sec/phase, 0.2mmol/kg Gd-DTPA), and diffusion tensor (2D DW-EPI, b=0, 1200 s/mm{sup 2}, 27 directions) imaging acquisitions were obtained during each study. Mid-treatment and pre-treatment images were rigidly aligned, and regions of partial response (PR), stable disease (SD), and progressive disease (PD) were contoured in consensus by two experienced radiation oncologists. Voxels in these categories were used to train ordinal (PR<SD<PD) and logistic (PR/SD<PD) models incorporating the apparent diffusion coefficient, fractional anisotropy, Ktrans, ve, vp, and relative cerebral blood volume imaging biomarkers, as well as treatment type. Leave-one-out cross-validation was performed at the patient level to assess model prediction accuracy. Results: Ordinal regression resulted in model prediction accuracies of 60% (PR), 0% (SD), 81% (PD), and 69% (overall), with coefficients of variation (COV) of 9.4%, 9.6%, and 23.6%, respectively. Logistic regression resulted in accuracies of 82.0% (PR/SD), 46.2% (PD), and 76.2% (overall) with COVs of 22.4%, 45.7%, and 23.8%, respectively. Conclusion: Despite limited patient numbers, this feasibility pilot study demonstrates that ordinal and logistic regression models potentially provide a flexible statistical framework for incorporating longitudinal multiparametric quantitative imaging data and, for these particular data, can potentially distinguish PR from PD. Additional training data exhibiting a full range of treatment responses is expected to allow prediction of all three treatment outcomes. Moreover, these models are extensible to multi-modality, multiparametric analyses.

60 APPLIED LIFE SCIENCES↗
DOE OSTI2025

Knowledge Oriented Graph Unified Transformer (KOGUT) v0.1

KOGUT — Knowledge Oriented Graph Unified Transformer KOGUT implements the Relational Graph Transformer (RelGT) architecture for knowledge graph link prediction in biological domains, with a primary focus on microbial growth media prediction. While the original RelGT (arXiv:2505.10960) targets relational tables, time series, and multi-table databases, KOGUT adapts this architecture for heterogeneous biological knowledge graphs, providing first-in-class AI predictive models for microbial cultivation. Key Adaptations Beyond Original RelGT: - Knowledge Graph Focus: Applied to biological KGs with semantic node types (taxa, chemicals, media, phenotypes, environments) versus generic relational database tables, trained on the KG-Microbe knowledge graph (1.3M entities, 2.9M edges, 24 relation types). - Multimodal Node Encoding: Integrates node labels, categories, descriptions, and synonyms from KG metadata through learned embedding layers—adapting relational column features to graph node attributes with textual semantics. - Extended K-Hop Subgraph Strategy: Optimized neighborhood sampling (3-hop default, configurable up to 200 nodes) tuned for sparse biological networks, building on the original local-global attention framework with biological relation preservation. - Biolink Predicate Preservation: Type-specific transformations for 24 biological edge semantics (occurs_in, consumes, produces, has_phenotype, subclass_of) beyond standard relational foreign keys, enabling multi-relation link prediction. - Inductive Learning Support: Enables zero-shot predictions for novel taxa through feature-based embeddings (temperature, oxygen requirements, gram stain, cell shape), extending the original transductive relational benchmark scope to uncultured microorganisms. CheapSOTA Performance Optimizations (This Distribution): - VQ-EMA Centroid Attention: Vector quantization with exponential moving average for improved global context modeling (+5-10% MRR improvement). - HDF5 Precomputed Data Loading: One-time preprocessing of k-hop subgraphs to eliminate redundant graph traversals (2-5× training speedup). - Distributed Data Parallel Training: Multi-GPU support for scaling to larger knowledge graphs (tested on 4× NVIDIA A100 GPUs at NERSC Perlmutter). - Mixed Precision Training: Automatic mixed precision (AMP) for memory efficiency and faster training. Advantages Over Standard Knowledge Graph Embedding Models: Combines RelGT's proven multi-element tokenization (features, type, hop, structure) with graph-native biological representations, enabling interpretable link prediction across heterogeneous entities that standard embedding models (TransE, RotatE, ComplEx) and table-based transformers cannot directly model. Achieves near-perfect performance on microbial growth media prediction (MRR: 0.9966, Precision@1: 0.9932, Hit@10: 1.0000) while maintaining explainability through attention-based reasoning over biological pathways. Training Data: - KG-Microbe merged knowledge graph: 1,379,337 nodes, 2,960,472 edges - 24 biological relation types including taxonomic hierarchies, metabolic interactions, phenotype associations, and environmental relationships - Primary prediction task: Growth media suitability for microbial taxa (biolink:occurs_in, 50K edges) - Multi-relation capability: Predicts links for any of the 24 relation types, including chemical consumption/production, phenotype associations, and taxonomic classification Citation: Original RelGT Architecture: Dwivedi et al., "Relational Graph Transformer", arXiv:2505.10960, 2025 KOGUT Implementation: Knowledge Oriented Graph Unified Transformer for Microbial Growth Media Prediction Developed at Lawrence Berkeley National Laboratory (LBNL) Trained on NERSC Perlmutter supercomputer

Joachimiak, Marcin [Lawrence Berkeley National Lab↗
DOE OSTI2021

Heat transport in liquid water from first-principles and deep neural network simulations

In this work, we compute the thermal conductivity of water within linear response theory from equilibrium molecular dynamics simulations, by adopting two different approaches. In one, the potential energy surface (PES) is derived on the fly from the electronic ground state of density functional theory (DFT) and the corresponding analytical expression is used for the energy flux. In the other, the PES is represented by a deep neural network (DNN) trained on DFT data, whereby the PES has an explicit local decomposition and the energy flux takes a particularly simple expression. By virtue of a gauge invariance principle, established by Marcolongo, Umari, and Baroni, the two approaches should be equivalent if the PES were reproduced accurately by the DNN model. We test this hypothesis by calculating the thermal conductivity, at the GGA (PBE) level of theory, using the direct formulation and its DNN proxy, finding that both approaches yield the same conductivity, in excess of the experimental value by approximately 60%. Besides being numerically much more efficient than its direct DFT counterpart, the DNN scheme has the advantage of being easily applicable to more sophisticated DFT approximations, such as meta-GGA and hybrid functionals, for which it would be hard to derive analytically the expression of the energy flux. We find in this way that a DNN model, trained on meta-GGA (SCAN) data, reduces the deviation from experiment of the predicted thermal conductivity by about 50%, leaving the question open as to whether the residual error is due to deficiencies of the functional, to a neglect of nuclear quantum effects in the atomic dynamics, or, likely, to a combination of the two.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2026

Deployment of Traditional and Hybrid Machine Learning for Critical Heat Flux Prediction in the CTF Thermal-Hydraulics Code

Critical heat flux (CHF) marks the transition from nucleate to film boiling, where heat transfer to the working fluid can rapidly deteriorate. Accurate CHF prediction is essential for efficiency, safety, and preventing equipment damage, particularly in nuclear reactors. Although widely used, empirical correlations frequently exhibit discrepancies when compared to experimental data, limiting their reliability in diverse operational conditions. Traditional machine learning (ML) approaches have demonstrated potential for CHF prediction but often suffer from limited interpretability, data scarcity, and insufficient knowledge of physical principles. Hybrid model approaches, which combine data-driven ML with base models, mitigate these concerns by incorporating prior knowledge of the domain. This study integrates an externally trained purely data-driven ML model and two hybrid models (using the Biasi and Bowring CHF correlations) within the CTF subchannel code via a custom Fortran framework. Performance was evaluated using two validation cases: a subset of the Nuclear Regulatory Commission (NRC) CHF database and the Bennett dryout experiments. In both cases, the hybrid models demonstrated significantly lower error metrics compared to conventional empirical correlations, with the best models often reducing relative error by about 5 percentage points. The pure ML model achieved comparable accuracy, outperforming the hybrid Biasi model in the NRC test case (3.3% versus 5.5% relative error) but exhibiting slightly higher error against the hybrid Bowring model in the Bennett test case (7.7% versus 6.1%). Trend analysis of error parity indicated that ML-based models reduced the tendency for CHF overprediction, improving overall accuracy. These results demonstrate that ML-based CHF models can be effectively integrated into subchannel codes and could potentially increase performance compared to conventional methods.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2013

DETECTING ACTIVE GALACTIC NUCLEI USING MULTI-FILTER IMAGING DATA. II. INCORPORATING ARTIFICIAL NEURAL NETWORKS

This is the second paper of the series Detecting Active Galactic Nuclei Using Multi-filter Imaging Data. In this paper we review shapelets, an image manipulation algorithm, which we employ to adjust the point-spread function (PSF) of galaxy images. This technique is used to ensure the image in each filter has the same and sharpest PSF, which is the preferred condition for detecting AGNs using multi-filter imaging data as we demonstrated in Paper I of this series. We apply shapelets on Canada-France-Hawaii Telescope Legacy Survey Wide Survey ugriz images. Photometric parameters such as effective radii, integrated fluxes within certain radii, and color gradients are measured on the shapelets-reconstructed images. These parameters are used by artificial neural networks (ANNs) which yield: photometric redshift with an rms of 0.026 and a regression R-value of 0.92; galaxy morphological types with an uncertainty less than 2 T types for z ≤ 0.1; and identification of galaxies as AGNs with 70% confidence, star-forming/starburst (SF/SB) galaxies with 90% confidence, and passive galaxies with 70% confidence for z ≤ 0.1. The incorporation of ANNs provides a more reliable technique for identifying AGN or SF/SB candidates, which could be very useful for large-scale multi-filter optical surveys that also include a modest set of spectroscopic data sufficient to train neural networks.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2022

Dataset for 'Ombadi, M. & Varadharajan, C. (2022). Urbanization and aridity mediate distinct salinity response to floods in rivers and streams across the Contiguous United States, Water Research'

This package contains data sets and code used to obtain the results in Ombadi, M., & Varadharajan, C. (2022). Urbanization and aridity mediate distinct salinity response to floods in rivers and streams across the Contiguous United States. Water Research, 118664. The folder "data" contains 259 .csv files, each of which has daily time series of concurrent streamflow (Q) and specific conductance (SC) for each of the sites used in this study originally downloaded from the USGS National Water Information System (NWIS; USGS, 2016). The number of data points in each of the files is at least 3650 (i.e. 10 years of daily measurements). The folder "RF_single_data" contains 259 .csv files, each of which include data used to train and test the Random Forest models at individual sites for predicting SC during days of floods. The folder "RF_regional_data" contains 3 .csv files, each of which include scaled data compiled from all sites within each climate zone (arid, temperate and wet). "metadata.csv" contains the physical properties of the 259 catchments corresponding to the sites used in this study; this data was extracted from GAGES-II dataset (Falcone et al., 2010). "RF_implementation.ipynb" is a Jupyter notebook with the code needed to implement the analysis using Random Forest models either for individual sites or for the regional models (for each climate zone). The code utilizes the data in the two folders: "RF_single_data" and "RF_regional_data" and the metadata.csv file.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2020

Optimizing Classifiers for Radionuclide Identification

Identifying threat nuclear materials is a critical for the prevention of acts of nuclear terrorism on the homeland. For this purpose, many radionuclide identification devices are deployed in the field. However, these will not necessarily be in the hands of non-experts, therefore these devices need to provide ready-made answers for the personnel in the field. This is where advanced algorithms are employed to both interpret the data and provide the identification of the nuclear material being interrogated. We took a machine learning approach to identification, by using training and validation data sets to create and optimize classifiers which determine which radionuclide is consistent with the data. The classifiers investigated were the Random Forest, Decision Tree, Support Vector Machine, and XG Boost and their performance was judged using the F1 Score for both hyperparameter tuning and comparison. In the end, we found out that the Random Forest Classifier worked the best based off the F1 Score they got which was 0.98.

45 MILITARY TECHNOLOGY, WEAPONRY, AND NATIONAL DEF↗
NASA NTRS1981

NASA/BLM APT, phase 2. Volume 2: Technology demonstration

Techniques described include: (1) steps in the preprocessing of LANDSAT data; (2) the training of a classifier; (3) maximum likelihood classification and precision; (4) geometric correction; (5) class description; (6) digitizing; (7) digital terrain data; (8) an overview of sample design; (9) allocation and selection of primary sample units; (10) interpretation of secondary sample units; (11) data collection ground plots; (12) data reductions; (13) analysis for productivity estimation and map verification; (14) cost analysis; and (150) LANDSAT digital products. The evaluation of the pre-inventory planning for P.J. is included.

Source record↗
DOE OSTI2024

Seismic Signal Detection on International Monitoring System 3-Component Stations using PhaseNet

In this report we discuss training a deep learning seismic signal detection model on 3-component stations from the International Monitoring System (IMS) using the PhaseNet architecture. Using 14 years of associated signals from the International Data Centre’s (IDC) Late Event Bulletin (LEB), we auto-curated training data consisting of signal windows containing associated arrivals, and noise windows that contain no LEB-associated signals. We trained several models using different waveform window durations (30 seconds and 100 seconds), with and without bandpass filtering. We evaluated the effectiveness of our models using associated signals from the Unconstrained Global Event Bulletin (UGEB) and found that several of our models outperformed the signal detections from the IDC’s Selected Event List 3 (SEL3) arrival table. The SEL3 bulletin evaluated on the UGEB dataset with 100-second waveform windows registered a precision and recall of .15 and .48, respectively, versus .19 and .59 for our filtered-data model. For the 30-second waveform window dataset, the SEL3 bulletin achieved a precision and recall of .31 and .47, respectively, versus .32 and .60 for our filtered-data model. Finally, our models detected signals from all source-to-receiver distances, suggesting it is feasible to use a single PhaseNet model for the IMS network.

58 GEOSCIENCES↗
NASA NTRS1993

LVQ and backpropagation neural networks applied to NASA SSME data

Feedfoward neural networks with backpropagation learning have been used as function approximators for modeling the space shuttle main engine (SSME) sensor signals. The modeling of these sensor signals is aimed at the development of a sensor fault detection system that can be used during ground test firings. The generalization capability of a neural network based function approximator depends on the training vectors which in this application may be derived from a number of SSME ground test-firings. This yields a large number of training vectors. Large training sets can cause the time required to train the network to be very large. Also, the network may not be able to generalize for large training sets. To reduce the size of the training sets, the SSME test-firing data is reduced using the learning vector quantization (LVQ) based technique. Different compression ratios were used to obtain compressed data in training the neural network model. The performance of the neural model trained using reduced sets of training patterns is presented and compared with the performance of the model trained using complete data. The LVQ can also be used as a function approximator. The performance of the LVQ as a function approximator using reduced training sets is presented and compared with the performance of the backpropagation network.

Doniere, Timothy F.↗
NASA NTRS1975

M-DAS: System for multispectral data analysis

M-DAS is a ground data processing system designed for analysis of multispectral data. M-DAS operates on multispectral data from LANDSAT, S-192, M2S and other sources in CCT form. Interactive training by operator-investigators using a variable cursor on a color display was used to derive optimum processing coefficients and data on cluster separability. An advanced multivariate normal-maximum likelihood processing algorithm was used to produce output in various formats: color-coded film images, geometrically corrected map overlays, moving displays of scene sections, coverage tabulations and categorized CCTs. The analysis procedure for M-DAS involves three phases: (1) screening and training, (2) analysis of training data to compute performance predictions and processing coefficients, and (3) processing of multichannel input data into categorized results. Typical M-DAS applications involve iteration between each of these phases. A series of photographs of the M-DAS display are used to illustrate M-DAS operation.

Johnson, R. H.↗
DOE OSTI2026

MTL_TX: A Multi-Task Transformer Model for Improved Radiation Time-Series Estimation

Controlling radiation doses at potential radioactive facilities is critical to ensuring the safety of both personnel and the public. At the Thomas Jefferson National Accelerator Facility (JLab), multiple sensors are deployed around the three experimental halls to monitor key parameters, including single-beam current, energy levels, current leakage, and radiation values during accelerator operations. In this study, we developed a Multi-task Transformer model, MTL_TX, to accurately estimate radiation doses at sensor locations based on historical data, with the aim of enhancing safety in accelerator facilities and surrounding public areas. To improve estimation accuracy, we integrated two innovative components into the proposed model: hierarchical feature embedding (HFE) and multi-level decomposition attention (MDA). Additionally, the multi-task learning (MTL) framework effectively leverages correlations among multiple sensors, enabling individual estimations for each sensor. MTL_TX achieved outstanding results on data collected in 2018, with an MSE of 0.1464, an RMSE of 0.2353, and an R 2 score of 0.8584. Furthermore, when trained on 2018 data, MTL_TX exhibited excellent generalization capability to unseen datasets from 2016 to 2019, achieving an MSE of 0.1407, an RMSE of 0.2263, and an R 2 score of 0.8831. These results demonstrate a significant improvement over existing state-of-the-art models.

Transformer↗
DOE OSTI2018

Identifying Sociomarkers of Pediatric Asthma Patients at Risk of Hospital Revisiting

ObjectiveAsthma is one of the most common chronic childhood diseases in the United States [2, 3]. In addition to its pervasiveness, pediatric asthma shows high sensitivity to the environment. Combining medical-social dataset with machine learning methods we demonstrate how socio-markers play an important role in identifying patients at risk of hospital revisits due to pediatric asthma within a year.IntroductionA socio-marker is a measurable indicator of social conditions where a patient is embedded in and exposed to, being analogous with a biomarker indicating the severity or presence of some disease state. Social factors are one of the most clinical health determinants [1], which play a critical role in explaining health outcomes. Socio-markers can help medical practitioners and researchers to reliably identify high-risk individuals in a timely manner.MethodsWe collected data from three different sources: pediatric asthma encounter records from Jan 1st, 2016 to Dec 31st, 2016 at a children’s hospital, the 2010 U.S census data and neighborhood quality survey data by Memphis Property Hub. After merging these datasets we examine the effect of social features in identifying the patients who visited the hospital more than once during the observation period. We only use the first time visit (3,678 cases) to avoid over-counting of the same patients. In addition to demographic features (age, gender, insurance type, and race (African American and White)), we incorporate the social features such as the proportion of individuals living below the federal poverty level, blight prevalence, neighborhood quality, neighborhood quality inequality, trash dumping presence, the broken window pervasiveness within the zip code area of patients’ residence are included.We then implemented a Support Vector Machine (SVM) based classification model using abovementioned 11 social features. The classification outcome is either patient visits the hospital only one-time (class 0) or revisits the hospital within a year (class 1). Among 3,678 unique patients in the dataset, only 823 pediatric patients revisited hospital with asthma. So, to overcome the class imbalance issue, we have used 823 patients’ data (randomly selected in 1,000 iterations) from each class. Further, to avoid overfitting and ensure generalizability, we divided the dataset as training, test, and validation with a proportion of 60%, 20%, and 20%, respectively. The reported test (5-folds cross-validation using training and testing data) and validation accuracy of the SVM method are averaged over 1,000 iterations to avoid sampling error and bias.ResultsThe proposed socio-marker model resulted in an average classification accuracy of 63.70% for the test set and 63.67 % for the validation set. Further, the average specificity (the total true negative cases divided by the sum of true negative and false positive) and sensitivity (the total number of true positive cases divided by the sum of positive predicted cases) is found to be 62.79% and 64.77%, respectively for the test set and 62.79% and 64.83%, respectively for the validation set. Results of this study suggest that socio-marker features that are not directly related to a patient’s medical conditions can still predict whether the patient will come back to the hospital within a year or not with approximately 64% accuracy.ConclusionsBringing the socio-marker features in the surveillance system may ease the burden of detecting the patients at risk of revisiting the hospital. The results should be interpreted with caution because we only used 12-month period of observation and the visit beyond the observation window is not considered. Also the patients may have visited different hospitals which are not captured in the data.References1. Booske BC, Athens JK, Kindig DA, Park H, Remington PL: Different perspectives for assigning weights to determinants of health. University of Wisconsin: Population Health Institute 2010.2. Subbarao P, Mandhane PJ, Sears MR: Asthma: epidemiology, etiology and risk factors. Canadian Medical Association Journal 2009, 181(9):E181-E190.3. Gold DR, Wright R: Population disparities in asthma. Annu Rev Public Health 2005, 26:89-113.

Shin, Eun Kyong↗
NASA NTRS1993

Functional approximation using artificial neural networks in structural mechanics

The artificial neural networks (ANN) methodology is an outgrowth of research in artificial intelligence. In this study, the feed-forward network model that was proposed by Rumelhart, Hinton, and Williams was applied to the mapping of functions that are encountered in structural mechanics problems. Several different network configurations were chosen to train the available data for problems in materials characterization and structural analysis of plates and shells. By using the recall process, the accuracy of these trained networks was assessed.

Alam, Javed↗
DOE OSTI2023

Active Learning Framework

Machine learning (ML) of interatomic potentials show great promise to accelerate scientific simulation, e.g., by emulating expensive computations at a high accuracy but much reduced computational cost. Training datasets are calculated from computationally expensive ab initio quantum mechanics methods, density functional theory (DFT). Trained on this data, an ML model can be very successful in predicting energy and forces for new atomic configurations. A critical factor is the quality and diversity of the training dataset. Thus, a highly automated approach to dataset construction based on active learning framework is designed suitable for material physics. The active learning scheme begins with fully randomized atomic configurations. Then, many Molecular Dynamics (MD) trajectories are simulated using current ML potentials, where each MD trajectory is initialized to a random disordered configuration. The temperature is varied in order to diversify the sampled configuration during these simulations. The variance of predictions for eight neural networks within an ensemble is analyzed to determine whether the model is operating as expected. This helps in determining whether collecting more data would be helpful to the model by checking the ensemble variance is greater than the threshold. In this case, the MD trajectory is terminated and the final atomic configuration is placed on a queue (SQL database) for DFT calculations and added to training dataset. Periodically, ML model is retrained to the updated training model. This Active Learning loop is iterated until the cost of MD simulations becomes prohibitively expensive. The MD simulations will hopefully be sufficiently robust to support nucleation after many active learning iterations. In this sense, active learning scheme must automatically discover the important low energy and nonequilibrium physics.

Nebgen, Benjamin↗