Search NASA⌕ Search

SEARCH · Search NASA

Results for “computationally expensive function”

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 487 records · Page 27
All sourcesNASA NTRSDOE OSTI
DOE OSTI2015

PanFP: Pangenome-based functional profiles for microbial communities

For decades there has been increasing interest in understanding the relationships between microbial communities and ecosystem functions. Current DNA sequencing technologies allows for the exploration of microbial communities in two principle ways: targeted rRNA gene surveys and shotgun metagenomics. For large study designs, it is often still prohibitively expensive to sequence metagenomes at both the breadth and depth necessary to statistically capture the true functional diversity of a community. Although rRNA gene surveys provide no direct evidence of function, they do provide a reasonable estimation of microbial diversity, while being a very cost effective way to screen samples of interest for later shotgun metagenomic analyses. However, there is a great deal of 16S rRNA gene survey data currently available from diverse environments, and thus a need for tools to infer functional composition of environmental samples based on 16S rRNA gene survey data. As a result, we present a computational method called pangenome based functional profiles (PanFP), which infers functional profiles of microbial communities from 16S rRNA gene survey data for Bacteria and Archaea. PanFP is based on pangenome reconstruction of a 16S rRNA gene operational taxonomic unit (OTU) from known genes and genomes pooled from the OTU s taxonomic lineage. From this lineage, we derive an OTU functional profile by weighting a pangenome s functional profile with the OTUs abundance observed in a given sample. We validated our method by comparing PanFP to the functional profiles obtained from the direct shotgun metagenomic measurement of 65 diverse communities via Spearman correlation coefficients. These correlations improved with increasing sequencing depth, within the range of 0.8 0.9 for the most deeply sequenced Human Microbiome Project mock community samples. PanFP is very similar in performance to another recently released tool, PICRUSt, for almost all of survey data analysed here. But, our method is unique in that any OTU building method can be used, as opposed to being limited to closed reference OTU picking strategies against specific reference sequence databases. In conclusion, we developed an automated computational method, which derives an inferred functional profile based on the 16S rRNA gene surveys of microbial communities. The inferred functional profile provides a cost effective way to study complex ecosystems through predicted comparative functional metagenomes and metadata analysis. All PanFP source code and additional documentation are freely available online at GitHub.

16S rRNA survey↗
NASA NTRS2002

Application of the generalized vertical coordinate ocean model for better representing satellite data

It is found that two adaptive parametric functions can be introduced into the basic ocean equations for utilizing the optimal or hybrid features of commonly used z-level, terrain- following, isopycnal, and pressure coordinates in numerical ocean models. The two parametric functions are formulated by combining three techniques: the arbitrary vertical coordinate system of Kasahara (1 974), the Jacobian pressure gradient formulation of Song (1 998), and a newly developed metric factor that permits both compressible (non-Boussinesq) and incompressible (Boussinesq) approximations. Based on the new formulation, an adaptive modeling strategy is proposed and a staggered finite volume method is designed to ensure conservation of important physical properties and numerical accuracy. Implementation of the combined techniques to SCRUM (Song and Haidvogel1994) shows that the adaptive modeling strategy can be applied to any existing ocean model without incurring computational expense or altering the original numerical schemes. Such a generalized coordinate model is expected to benefit diverse ocean modelers for easily choosing optimal vertical structures and sharing modeling resources based on a common model platform. Several representing oceanographic problems with different scales and characteristics, such as coastal canyons, basin-scale circulation, and global ocean circulation, are used to demonstrate the model's capability for multiple applications. New results show that the model is capable of simultaneously resolving both Boussinesq and non-Boussinesq, and both small- and large-scale processes well. This talk will focus on its applications of multiple satellite sensing data in eddy-resolving simulations of Asian Marginal Sea and Kurosio. Attention will be given to how Topex/Poseidon SSH, TRMM SST; and GRACE ocean bottom pressure can be correctly represented in a non- Boussinesq model.

numerical ocean model generalized coordinate syste↗
DOE OSTI2018

Comprehensive Phase Diagrams of MoS 2 Edge Sites Using Dispersion-Corrected DFT Free Energy Calculations

A comprehensive set of surface phase diagrams addressing the catalytically relevant edges of the (100) surface of MoS 2 catalysts is developed using dispersion-corrected density functional theory and ab initio thermodynamic modeling. The results of the temperature-dependent, free energy-based thermodynamic model are presented over the full range of catalytically relevant temperatures and pressures, in addition to S- and H-coverages ranging from 0 to 100%. The results of this work allow for a full thermodynamic analysis to be performed at the conditions relevant to any promising reaction involving MoS 2 , ranging from hydrodesulfurization to dehydrogenation to electrocatalysis. Several methodological recommendations are discussed and implemented with the goal of improving the accuracy of the surface phase diagrams at minimal computational expense. A library of the most stable S- and H-adsorption modes is also developed so that linear scaling relationships can be used to correlate thermodynamic stability with kinetic activity. Applying the results to C–H bond activation of methane with a S 2 oxidant, here we predict S-coverages near 100% on the Mo- and S-edges to be thermodynamically favored and S monomers on edge sites with high S-coverages to be kinetically favorable. For H-abstraction on surface S atoms, the Mo-edge is also predicted to be more active than the S-edge.

36 MATERIALS SCIENCE↗
DOE OSTI2023

Novel Wavefunction Approaches for Studying Actinides and Other Heavy Elements (Final Report (2012-2020))

The theoretical characterization of actinide molecules is a crucial complement to their experimental study; actinides are vital to issues of national security and energy, but they are expensive and hazardous to study experimentally. However, because actinides contain a large number of electrons and are often strongly correlated, their theoretical description is difficult. We have had success studying actinide chemistry and physics using density functional theoretical approaches over the past several years, but we have found that such methods are incapable of treating strong correlations adequately. We thus shifted our research focus to the development of tractable wavefunction methods for strong correlations of actinides. Our early wavefunction-based approaches focused on extending symmetry-adapted, single-reference coupled cluster for treating strong correlations with feasible computational cost. While we have had some success along these lines, we have recently developed spin-projected unrestricted coupled cluster, which is essentially a black-box multi-reference coupled cluster theory that is superior to unrestricted coupled cluster for small- to medium-sized systems across practically all physical correlation strengths, yet preserves good quantum numbers. For actinides, however, where spin-orbit coupling is often nonnegligible, spin is no longer a fundamental symmetry. We thus propose to develop time reversal and point group projected coupled cluster, which will preserve the fundamental symmetries for actinides. Time reversal projection is achieved via the product of half-spin projection, or spin flip, and complex conjugation projection. Along with point group, these are discrete symmetries, i.e. non-continuous quantum numbers, which, in addition to being the correct symmetries for actinides, can be implemented in lower computational cost than full spin projection. We propose the development of these theories within a synergistic collaboration to use the new methods to elucidate difficult actinide chemistry.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2022

Active learning with multifidelity modeling for efficient rare event simulation

Here, while multifidelity modeling provides a cost-effective way to conduct uncertainty quantification with computationally expensive models, much greater efficiency can be achieved by adaptively deciding the number of required high-fidelity (HF) simulations, depending on the type and complexity of the problem and the desired accuracy in the results. We propose a framework for active learning with multifidelity modeling emphasizing the efficient estimation of rare events. Our framework works by fusing a low-fidelity (LF) prediction with an HF-inferred correction, filtering the corrected LF prediction to decide whether to call the high-fidelity model, and for enhanced subsequent accuracy, adapting the correction for the LF prediction after every HF model call. The framework does not make any assumptions as to the LF model type or its correlations with the HF model. In addition, for improved robustness when estimating smaller failure probabilities, we propose using dynamic active learning functions that decide when to call the HF model. We demonstrate our framework using several academic case studies (including some high-dimensional problems) and two finite element model case studies: estimating Navier-Stokes velocities using the Stokes approximation and estimating stresses in a transversely isotropic model subjected to displacements via a coarsely meshed isotropic model. Across these case studies, not only did the proposed framework estimate the failure probabilities accurately, but compared with either Monte Carlo or a standard variance reduction method, it also required only a small fraction of the calls to the HF model.

42 ENGINEERING↗
DOE OSTI2016

First principles modeling of the metal-electrolyte interface: A novel approach to the study of the electrochemical interface

The research objective of this proposal is the computational modeling of the metal-electrolyte interface purely from first principles. The accurate calculation of the electrostatic potential at electrically biased metal-electrolyte interfaces is a current challenge for periodic “ab-initio” simulations. It is also an essential requisite for predicting the correspondence between the macroscopic voltage and the microscopic interfacial charge distribution in electrochemical fuel cells. This interfacial charge distribution is the result of the chemical bonding between solute and metal atoms, and therefore cannot be accurately calculated with the use of semi-empirical classical force fields. The project aims to study in detail the structure and dynamics of aqueous electrolytes at metallic interfaces taking into account the effect of the electrode potential. Another side of the project is to produce an accurate method to simulate the water/metal interface. While both experimental and theoretical surface scientists have made a lot of progress on the understanding and characterization of both atomistic structures and reactions at the solid/vacuum interface, the theoretical description of electrochemical interfaces is still lacking behind. A reason for this is that a complete and accurate first principles description of both the liquid and the metal interfaces is still computationally too expensive and complex, since their characteristics are governed by the explicit atomic and electronic structure built at the interface as a response to environmental conditions. This project will characterize in detail how different theoretical levels of modeling describer the metal/water interface. In particular the role of van der Waals interactions will be carefully analyzed and prescriptions to perform accurate simulations will be produced.

36 MATERIALS SCIENCE↗
DOE OSTI2021

A REDUCED ORDER MODELING APPROACH TO PROBABILISTIC CREEP-DAMAGE PREDICTIONS IN FINITE ELEMENT ANALYSIS

This paper introduces a computationally efficient Reduced Order Modeling (ROM) approach for the probabilistic prediction of creep-damage failure. Component-level probabilistic simulations are needed to assess the reliability and safety of high-temperature components. Full-scale probabilistic creep-damage modeling in finite element (FE) approach is computationally expensive requiring many hundreds of simulations to replicate the uncertainty of component failure. To that end, ROM is proposed to minimize the elevated computational cost while controlling the loss of accuracy. It is proposed that full-scale probabilistic simulations can be completed in 1D at a reduced cost, the extremum conditions extracted, and those conditions applied for lower-cost 2D/3D probabilistic simulations of components that capture the mean and uncertainty of failure. The probabilistic Sine-hyperbolic (Sinh) model is selected which in previous work was calibrated to alloy 304 stainless steel. The Sinh model includes probability density functions (pdfs) for test condition (stress and temperature), initial damage (i.e. microstructure), and material properties uncertainty. The Sinh model is programmed into ANSYS finite element software using the USERCREEP.F material subroutine. First, the Sinh model and FE code are subject to verification and validation to ensure the accuracy of the simulations. Numerous Monte Carlo simulations are executed in a 1D model to generate probabilistic creep deformation, damage, and rupture data. This data is analyzed and the probabilistic parameters corresponding to extreme creep response are extracted. The ROM concept is applied where only the extreme conditions are applied in the 2D probabilistic prediction of a component. The probabilistic predictions between the 1D and 2D geometry is compared to assess ROM for creep. The accuracy of the probabilistic prediction employing the ROM approach will potentially reduce the time and cost of simulating complex engineering systems. Future studies will introduce multi-stage Sinh, stochasticity, and spatial uncertainty for improved prediction.

36 MATERIALS SCIENCE↗
DOE OSTI2025

Analyzing and Exploring Training Recipes for Large-Scale Transformer-Based Weather Prediction

Abstract The rapid rise of deep learning (DL) in numerical weather prediction (NWP) has led to a proliferation of models which forecast atmospheric variables with comparable or superior skill than traditional physics-based NWP. However, among these leading DL models, there is a wide variance in both the training settings and architecture used. Further, the lack of thorough ablation studies makes it hard to discern which components are most critical to success. In this work, we show that it is possible to attain high forecast skill even with relatively off-the-shelf architectures, simple training procedures, and moderate compute budgets. Specifically, we train a minimally modified Swin Transformer V2 (SwinV2) on ERA5 data and find that it attains superior skill in terms of mean-square errors of deterministic forecasts when compared against the European Centre for Medium-Range Weather Forecasts’ Integrated Forecasting System (IFS). Almost all DL–NWP systems share a core set of hyperparameters and design decisions. To aid and expedite future DL–NWP research, we present an in-depth, systematic exploration of different loss functions, model sizes and depths, patch sizes, and multistep training objectives. We also examine the model performance with metrics beyond the typical accuracy (ACC) and RMSE and investigate how the performance scales with model size. Through our open-source code, scoring pipelines, and models, we share our findings on key aspects of the training pipeline. These ablations reduce the necessity for expensive hyperparameter tuning and lower the barrier to entry for future DL–NWP research. Significance Statement This study investigates the potential of using large-scale transformer-based models for weather prediction, showing that it is possible to achieve high forecast accuracy with simpler, off-the-shelf architectures. By training a minimally modified SwinV2 transformer on ERA5 data, we show that the model achieves competitive forecast skill in terms of mean-square error for key variables, outperforming the European Centre for Medium-Range Weather Forecasts’ Integrated Forecasting System (IFS) at all lead times. Our findings suggest that effective training strategies, such as multistep fine-tuning and channel-weighted losses, significantly enhance the model’s performance. However, we also highlight that these improvements come with trade-offs in other areas, such as ensemble spread and high-frequency spatial detail. This work highlights the promise of deep learning in improving weather forecasts, which could lead to better preparedness and response to weather events, ultimately benefiting society by providing more reliable weather predictions.

Willard, Jared D. [Lawrence Berkeley National Labo↗
DOE OSTI2023

Application of unsupervised deep learning to image segmentation and in-situ contact angle measurements in a CO 2 -water-rock system

Rock surface wettability is a critical property that regulates multiphase flows in porous media, which can be quantified using the surface contact angle (CA). X-ray micro-computed tomography (μCT) provides an effective approach to in-situ measurements of surface CAs. However, the CA measurement accuracy depends significantly on the quality of CT image segmentation, which is the clustering of CT pixels into separate phases. Inspired by this, we developed a deep learning (DL)-based CA measurement workflow. Motivated by the recent tremendous progress in unsupervised learning techniques and aiming to avoid expensive manual data annotations, an unsupervised DL pipeline for CT image segmentation was proposed and implemented, which includes unsupervised model training and post-processing. The unsupervised model training was driven by a novel loss function constrained with feature similarity and spatial continuity and implemented by iterative forward and backward paths; the former clustered the pixel-wise feature vectors extracted by convolution neural networks, whereas the latter updated the parameters using gradient descent. An over-segmentation strategy was adopted for model training. The post-processing steps based on agglomerative hierarchical clustering (AHC) were implemented to further merge the over-segmented model output to the desired cluster number, which is intended to improve the efficiency of image segmentation. The developed unsupervised DL pipeline was compared with other commonly-used image segmentation methods using pixel-wise and physics-based evaluation metrics on a synthetic raw-image dataset, which had a known ground truth. The unsupervised DL pipeline showed the best performance. Next, the segmented images were input to an automatic CA measurement tool, and the results were validated by comparisons with manual measurements. The CA values from the manual and automatic measurements showed similar distributions and statistical properties. The automatic measurement demonstrated a wider spectrum because of the much larger number of measurement data points. The primary novelty of the unsupervised DL pipeline developed in this study lies in the novel loss function and the over-segmentation strategy associated with AHC post-processing. Finally, the workflow has been proven an efficient tool for pore-scale wettability characterization, which has a wide range of applications in fundamental studies of multiphase flows in natural porous media, which have critical implications to geological carbon sequestration, hydrocarbon energy recovery, and contaminant transport in groundwater.

42 ENGINEERING↗
DOE OSTI2020

On the Relationship Between Shallow Cumulus Cloud Field Properties and Surface Solar Irradiance

Abstract Shallow cumulus clouds exhibit highly three‐dimensional (3‐D) spatial structure leading to complex variability in the surface solar irradiance (SSI) beneath. This variability is captured by the typically bimodal shape of the SSI probability density function (PDF). Using large eddy simulation to generate well‐resolved cloud fields and Monte Carlo 3‐D radiative transfer to reproduce realistic associated SSI PDFs, we seek direct relationships between the cloud field properties and the SSI PDF shape. Applying both random forest and artificial neural network algorithms, we find variations in the two modes of the SSI PDF are well predicted by just a handful of cloud field properties. The two algorithms utilize cloud properties similarly, with indistinguishable performance despite their different architectures. These results offer a marked improvement in realism relative to one‐dimensional radiative transfer while bypassing computationally expensive 3‐D radiative transfer, with immediate application to renewable energy assessments, and potential for several other geophysical applications.

58 GEOSCIENCES↗
NASA NTRS2019

A Communication Channel Density Estimating Generative Adversarial Network

Autoencoder-based communication systems use neural network channel models to backwardly propagate message reconstruction error gradients across an approximation of the physical communication channel. In this work, we develop and test a new generative adversarial network (GAN) architecture for the purpose of training a stochastic channel approximating neural network. In previous research, investigators have focused on additive white Gaussian noise (AWGN) channels and/or simplified Rayleigh fading channels, both of which are linear and have well defined analytic solutions. Given that training a neural network is computationally expensive, channel approximation networks— and more generally the autoencoder systems—should be evaluated in communication environments that are traditionally difficult. To that end, our investigation focuses on channels that contain a combination of non-linear amplifier distortion, pulse shape filtering, intersymbol interference, frequency-dependent group delay, multipath, and non-Gaussian statistics. Each of our models are trained without any prior knowledge of the channel. We show that the trained models have learned to generalize over an arbitrary amplifier drive level and constellation alphabet. We demonstrate the versatility of our GAN architecture by comparing the marginal probability density function of several channel simulations with that of their corresponding neural network approximations

Smith, Aaron↗
DOE OSTI2022

Kinetic simulations of collision-less plasmas in open magnetic geometries

Laboratory plasmas in open magnetic geometries can be found in many different applications such as (a) scrape-of-layer (SOL) and divertor regions in toroidal confinement fusion devices, (b) linear divertor simulators, (c) plasma-based thrusters and (d) magnetic mirrors etc. A common feature of these plasma systems is the need to resolve, in addition to velocity space, at least one physical dimension (e.g. along flux lines) to capture the relevant physics. In general, this requires a kinetic treatment. Fully kinetic particle-in-cell (PIC) simulations can be applied but at the expense of large computational effort. A common way to resolve this is to use a hybrid approach: kinetic ions and fluid electrons. In the present work, the development of a hybrid PIC computational tool suitable for open magnetic geometries is described which includes (a) the effect of non-uniform magnetic fields, (b) finite fully-absorbing boundaries for the particles and (c) volumetric particle sources. Analytical expressions for the momentum transport in the paraxial limit are presented with their underlying assumptions and are used to validate the results from the PIC simulations. A general method is described to construct discrete particle distribution functions in a state of mirror-equilibrium. This method is used to obtain the initial state for the PIC simulation. Collisionless simulations in a mirror geometry are performed. The results show that the effect of magnetic compression is correctly described and momentum is conserved. The self-consistent electric field is calculated and is shown to modify the ion velocity distribution function in a manner consistent with analytic theory. Based on this analysis, the ion distribution function is understood in terms of a loss-cone distribution and an isotropic Maxwell-Boltzmann distribution driven by a volumetric plasma source. Finally, the inclusion of a Monte Carlo based Fokker-Planck collision operator is discussed in the context of future work.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2016

The clustering of galaxies in the completed SDSS-III Baryon Oscillation Spectroscopic Survey: Single-probe measurements from DR12 galaxy clustering – towards an accurate model

We analyze the broad-range shape of the monopole and quadrupole correlation functions of the BOSS Data Release 12 (DR12) CMASS and LOWZ galaxy sample to obtain constraints on the Hubble expansion rate H(z), the angular-diameter distance DA(z), the normalised growth rate f(z)σ 8 (z), and the physical matter density Ω m h 2 . In addition, we adopt wide and flat priors on all model parameters in order to ensure the results are those of a `single-probe' galaxy clustering analysis. We also marginalize over three nuisance terms that account for potential observational systematics affecting the measured monopole. However, such Monte Carlo Markov Chain analysis is computationally expensive for advanced theoretical models, thus we develop a new methodology to speed up our analysis.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2025

Raptor

Raptor is an efficient Python-based tool for predicting the formation and morphology of stochastic lack of fusion defects in metal AM processes. A major obstacle for the qualification and certification of additively manufactured parts in critical applications continues to be performance variability caused in part by porosity-related defects. High-fidelity process models that could predict these defect features are currently too computationally expensive for component-level analysis. To address this, Raptor employs a high-performance geometric method to model the dynamic melt pool rather than relying on computationally intensive thermal fluid dynamics. This allows Raptor to rapidly identify regions of unmelted material that correspond to lack of fusion pores. The efficiency of this approach significantly reduces the time and resources needed for generating 3D defect predictions, which enables users to conduct large-scale parameter studies and evaluate how process variations affect part quality. The framework offers operational flexibility; users can execute simulations through a simple command line interface or integrate core functions as a library within larger computational workflows. Simulation outputs include 3D porosity maps for visualization and tools for quantitative morphological analysis. These results are suitable for direct comparison with experimental characterization data from methods such as X-ray computed tomography and can be used for statistical process optimization.

Subraveti, Vamsi [Vanderbilt Univ., Nashville, TN ↗
NASA NTRS2000

A Solution Methodology and Computer Program to Efficiently Model Thermodynamic and Transport Coefficients of Mixtures

A solution methodology has been developed to efficiently model multi-specie, chemically frozen, thermally perfect gas mixtures. The method relies on the ability to generate a single (composite) set of thermodynamic and transport coefficients prior to beginning a CFD solution. While not fundamentally a new concept, many applied CFD users are not aware of this capability nor have a mechanism to easily and confidently generate new coefficients. A database of individual specie property coefficients has been created for 48 species. The seven coefficient form of the thermodynamic functions is currently used rather then the ten coefficient form due to the similarity of the calculated properties, low temperature behavior and reduced CPU requirements. Sutherland laminar viscosity and thermal conductivity coefficients were computed in a consistent manner from available reference curves. A computer program has been written to provide CFD users with a convenient method to generate composite specie coefficients for any mixture. Mach 7 forebody/inlet calculations demonstrated nearly equivalent results and significant CPU time savings compared to a multi-specie solution approach. Results from high-speed combustor analysis also illustrate the ability to model inert test gas contaminants without additional computational expense.

Ferlemann, Paul G.↗
NASA NTRS2010

Statistical and Probabilistic Extensions to Ground Operations' Discrete Event Simulation Modeling

NASA's human exploration initiatives will invest in technologies, public/private partnerships, and infrastructure, paving the way for the expansion of human civilization into the solar system and beyond. As it is has been for the past half century, the Kennedy Space Center will be the embarkation point for humankind's journey into the cosmos. Functioning as a next generation space launch complex, Kennedy's launch pads, integration facilities, processing areas, launch and recovery ranges will bustle with the activities of the world's space transportation providers. In developing this complex, KSC teams work through the potential operational scenarios: conducting trade studies, planning and budgeting for expensive and limited resources, and simulating alternative operational schemes. Numerous tools, among them discrete event simulation (DES), were matured during the Constellation Program to conduct such analyses with the purpose of optimizing the launch complex for maximum efficiency, safety, and flexibility while minimizing life cycle costs. Discrete event simulation is a computer-based modeling technique for complex and dynamic systems where the state of the system changes at discrete points in time and whose inputs may include random variables. DES is used to assess timelines and throughput, and to support operability studies and contingency analyses. It is applicable to any space launch campaign and informs decision-makers of the effects of varying numbers of expensive resources and the impact of off nominal scenarios on measures of performance. In order to develop representative DES models, methods were adopted, exploited, or created to extend traditional uses of DES. The Delphi method was adopted and utilized for task duration estimation. DES software was exploited for probabilistic event variation. A roll-up process was used, which was developed to reuse models and model elements in other less - detailed models. The DES team continues to innovate and expand DES capabilities to address KSC's planning needs.

Trocine, Linda↗
DOE OSTI2021

The Rise of Neural Networks for Materials and Chemical Dynamics

Machine learning (ML) is quickly becoming a premier tool for modeling chemical processes and materials. ML-based force fields, trained on large data sets of high-quality electron structure calculations, are particularly attractive due their unique combination of computational efficiency and physical accuracy. This Perspective summarizes some recent advances in the development of neural network-based interatomic potentials. Designing high-quality training data sets is crucial to overall model accuracy. One strategy is active learning, in which new data are automatically collected for atomic configurations that produce large ML uncertainties. Another strategy is to use the highest levels of quantum theory possible. Transfer learning allows training to a data set of mixed fidelity. A model initially trained to a large data set of density functional theory calculations can be significantly improved by retraining to a relatively small data set of expensive coupled cluster theory calculations. These advances are exemplified by applications to molecules and materials.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2023

Accelerating Scientific Simulations with Bi-Fidelity Weighted Transfer Learning

High-fidelity modeling is an essential design tool for many engineering applications. However, for complex systems, computational cost can be a limiting factor. Analyzing parameter sensitivity, uncertainty quantification, and design optimization require many model evaluations. Surrogate models are often used to develop the relationship between model parameters and quantities of interest. However, in the case of complex systems, surrogate models require several degrees of freedom and, thus, a large number of data points to determine the correct dependencies. For many applications, this may be prohibitively expensive. The reduction of computational requirements can be achieved by leveraging low-fidelity models. Low-fidelity models represent the system at a coarser resolution with the advantage of computational efficiency. Therefore, a bi-fidelity modeling paradigm, which augments the accuracy of a low-fidelity model in a computationally efficient manner by invoking limited runs of a high-fidelity model, can be leveraged to sufficiently balance the accuracy and computational requirements. In this work, a bi-fidelity weighted transfer learning method using neural networks was applied to a computational fluid dynamics heat transfer modeling problem. The transfer learning advantage was investigated as a function of hyperparameters. Our main finding is that the use of a bi-fidelity modeling paradigm achieves accuracy close to that of a high-fidelity Gaussian process model while significantly reducing computational cost. The bi-fidelity model achieves comparable performance with 90 high-fidelity samples-that is, 60% less than the samples needed to achieve similar accuracy without the use of bi-fidelity modeling,

Borowiec, Katarzyna↗