Search NASA⌕ Search

SEARCH · Search NASA

Results for “generalization via transfer learning”

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 37 records · Page 2

Quantum biology revisited

Photosynthesis is a highly optimized process from which valuable lessons can be learned about the operating principles in nature. Its primary steps involve energy transport operating near theoretical quantum limits in efficiency. Recently, extensive research was motivated by the hypothesis that nature used quantum coherences to direct energy transfer. This body of work, a cornerstone for the field of quantum biology, rests on the interpretation of small-amplitude oscillations in two-dimensional electronic spectra of photosynthetic complexes. This Review discusses recent work reexamining these claims and demonstrates that interexciton coherences are too short lived to have any functional significance in photosynthetic energy transfer. Instead, the observed long-lived coherences originate from impulsively excited vibrations, generally observed in femtosecond spectroscopy. These efforts, collectively, lead to a more detailed understanding of the quantum aspects of dissipation. Nature, rather than trying to avoid dissipation, exploits it via engineering of exciton-bath interaction to create efficient energy flow.

59 BASIC BIOLOGICAL SCIENCES↗

NSFnets (Navier-Stokes flow nets): Physics-informed neural networks for the incompressible Navier-Stokes equations

In the last 50 years there has been a tremendous progress in solving numerically the Navier-Stokes equations using finite differences, finite elements, spectral, and even meshless methods. Yet, in many real cases, we still cannot incorporate seamlessly (multi-fidelity) data into existing algorithms, and for industrial-complexity applications the mesh generation is time consuming and still an art. Moreover, solving ill-posed problems (e.g., lacking boundary conditions) or inverse problems is often prohibitively expensive and requires different formulations and new computer codes. Here, we employ physics-informed neural networks (PINNs), encoding the governing equations directly into the deep neural network via automatic differentiation, to overcome some of the aforementioned limitations for simulating incompressible laminar and turbulent flows. We develop the Navier-Stokes flow nets (NSFnets) by considering two different mathematical formulations of the Navier-Stokes equations: the velocity-pressure (VP) formulation and the vorticity-velocity (VV) formulation. Since this is a new approach, we first select some standard benchmark problems to assess the accuracy, convergence rate, computational cost and flexibility of NSFnets; analytical solutions and direct numerical simulation (DNS) databases provide proper initial and boundary conditions for the NSFnet simulations. The spatial and temporal coordinates are the inputs of the NSFnets, while the instantaneous velocity and pressure fields are the outputs for the VP-NSFnet, and the instantaneous velocity and vorticity fields are the outputs for the VV-NSFnet. This is unsupervised learning and, hence, no labeled data are required beyond boundary and initial conditions and the fluid properties. The residuals of the VP or VV governing equations, together with the initial and boundary conditions, are embedded into the loss function of the NSFnets. No data is provided for the pressure to the VP-NSFnet, which is a hidden state and is obtained via the incompressibility constraint without extra computational cost. Unlike the traditional numerical methods, NSFnets inherit the properties of neural networks (NNs), hence the total error is composed of the approximation, the optimization, and the generalization errors. Here, we empirically attempt to quantify these errors by varying the sampling (“residual”) points, the iterative solvers, and the size of the NN architecture. For the laminar flow solutions, we show that both the VP and the VV formulations are comparable in accuracy but their best performance corresponds to different NN architectures. The initial convergence rate is fast but the error eventually saturates to a plateau due to the dominance of the optimization error. For the turbulent channel flow, we show that NSFnets can sustain turbulence at , but due to expensive training we only consider part of the channel domain and enforce velocity boundary conditions on the subdomain boundaries provided by the DNS data base. We also perform a systematic study on the weights used in the loss function for balancing the data and physics components, and investigate a new way of computing the weights dynamically to accelerate training and enhance accuracy. In the last part, we demonstrate how NSFnets should be used in practice, namely for ill-posed problems with incomplete or noisy boundary conditions as well as for inverse problems. We obtain reasonably accurate solutions for such cases as well without the need to change the NSFnets and at the same computational cost as in the forward well-posed problems. As a result, we also present a simple example of transfer learning that will aid in accelerating the training of NSFnets for different parameter settings.

97 MATHEMATICS AND COMPUTING↗

Onboard Hyperspectral Image Classification via Transfer Learning for Communication-Limited Spacecraft

Employing deep-learning and artificial-intelligence (AI) techniques onboard spacecraft can dramatically improve priority data selection to ensure more effective use of the available downlink. However, deployment of effective deep-learning models requires significant training on the ground, which may not be feasible, due to limited data available in an unexplored environment. Therefore, this research explores building robust classification models for onboard data processing where training data is highly limited using transfer-learning techniques. In this paper, we focus on the use case of hyperspectral imaging for remote sensing, a domain where the high dimensionality of the data from the sensor can rapidly saturate the downlink bandwidth. With this bottleneck, there is an impending need to autonomously and robustly classify data onboard to optimize downlink of high-impact measurements, thus maximizing the scientific utility per bit transmitted to the ground. This paper examines the use of deep neural networks onboard for hyperspectral image classification in a communication-limited scenario to analyze how the models perform with limited training data. The use of transfer learning can ameliorate the issue of poor generalization by transferring features learned from training on a large source dataset for one classification task to the target classification task with limited training data. For two deep-learning models from literature, we compare the accuracy of the models trained using transfer learning to models trained from scratch using a random weight initialization with varying amounts of training data. We demonstrate the feasibility and performance of running inference of the deep-learning models on representative flight-like hardware.

Advanced Avionics, Machine Learning, Data Processi↗

Data Curation for Machine Learning Applied to Geothermal Power Plant Operational Data for GOOML: Geothermal Operational Optimization with Machine Learning: Preprint

Geothermal Operational Optimization with Machine Learning (GOOML) is a transferable and extensible component-based geothermal asset modeling framework that considers complex steamfield relationships and identifies optimization prospects using a data-driven approach to physics-guided, data-centric machine learning. This framework has been used to develop digital twins that provide steamfield operators with operational environments to analyze and understand historical and forecasted power production, explore new steamfield configuration possibilities, and seek optimal asset management in real world applications. To create, test, and apply the GOOML framework, diverse time-series datasets spanning multiple years were sourced from various geothermal power plant components within several complex real-world geothermal operations. These operations are based in the United States and New Zealand and include a variety of technologies, end-uses and configurations, collectively covering nearly all relevant operating conditions for modern geothermal fields. Datasets were acquired from multiple sources to ensure that machine learning experiments generalized properly to various operating conditions. It was found that the data varied in quality, format, and completeness. To ensure consistency between the various datasets, a standardized data curation process was developed to reliably streamline data preparation. This paper will discuss best practices as learned from the GOOML data curation process which takes the following steps: 1) acquisition of large quantities of data from power plant operators, 2) digestion of data to gain an initial understanding of what is included, 3) data transformation, which includes converting the data into a standardized machine-readable format so that they can be visualized, quality checked, and cleaned, 4) quality assurance and quality control, involving identification of significant data gaps and apparent anomalies through mapping of data features to real world componentry via the GOOML historical model, followed by discussion with modelers and power plant operators to identify additional data needs and to resolve issues, 5) use in machine learning algorithms, and 6) repetition of steps one through five until all data needs are met and data are deemed suitable for producing trustworthy modeling results which may be disseminated, ideally along with the curated dataset. This iterative process is focused on improving the quality of the data rather than tuning machine learning model parameters and supports a shift towards data-centric AI as a means to improving real-world applicability of geothermal machine learning projects.

access↗

Predicting the propensity for thermally activated β events in metallic glasses via interpretable machine learning

Abstract The elementary excitations in metallic glasses (MGs), i.e., β processes that involve hopping between nearby sub-basins, underlie many unusual properties of the amorphous alloys. A high-efficacy prediction of the propensity for those activated processes from solely the atomic positions, however, has remained a daunting challenge. Recently, employing well-designed site environment descriptors and machine learning (ML), notable progress has been made in predicting the propensity for stress-activated β processes (i.e., shear transformations) from the static structure. However, the complex tensorial stress field and direction-dependent activation could induce non-trivial noises in the data, limiting the accuracy of the structure-property mapping learned. Here, we focus on the thermally activated elementary excitations and generate high-quality data in several Cu-Zr MGs, allowing quantitative mapping of the potential energy landscape. After fingerprinting the atomic environment with short- and medium-range interstice distribution, ML can identify the atoms with strong resistance or high compliance to thermal activation, at a high accuracy over ML models for stress-driven activation events. Interestingly, a quantitative “between-task” transferring test reveals that our learnt model can also generalize to predict the propensity of shear transformation. Our dataset is potentially useful for benchmarking future ML models on structure-property relationships in MGs.

36 MATERIALS SCIENCE↗

Evolution from a respiratory ancestor to fill syntrophic and fermentative niches: comparative fenomics of six Geobacteraceae species

Background: The anaerobic degradation of organic matter in natural environments, and the biotechnical use of anaerobes in energy production and remediation of subsurface environments, both require the cooperative activity of a diversity of microorganisms in different metabolic niches. The Geobacteraceae family contains members with three important anaerobic metabolisms: fermentation, syntrophic degradation of fermentation intermediates, and anaerobic respiration. Results: In order to learn more about the evolution of anaerobic microbial communities, the genome sequences of six Geobacteraceae species were analyzed. The results indicate that the last common Geobacteraceae ancestor contained sufficient genes for anaerobic respiration, completely oxidizing organic compounds with the reduction of external electron acceptors, features that are still retained in modern Geobacter and Desulfuromonas species. Evolution of specialization for fermentative growth arose twice, via distinct lateral gene transfer events, in Pelobacter carbinolicus and Pelobacter propionicus. Furthermore, P. carbinolicus gained hydrogenase genes and genes for ferredoxin reduction that appear to permit syntrophic growth via hydrogen production. The gain of new physiological capabilities in the Pelobacter species were accompanied by the loss of several key genes necessary for the complete oxidation of organic compounds and the genes for the c-type cytochromes required for extracellular electron transfer. Conclusion: The results suggest that Pelobacter species evolved parallel strategies to enhance their ability to compete in environments in which electron acceptors for anaerobic respiration were limiting. More generally, these results demonstrate how relatively few gene changes can dramatically transform metabolic capabilities and expand the range of environments in which microorganisms can compete.

59 BASIC BIOLOGICAL SCIENCES↗

Influence of Solutocapillary Convection on Macrovoid Defect Formation in Polymeric Membranes

Macrovoids (MVs) are large (10-50 micrometers) pores often found in polymeric membranes prepared via phase-inversion techniques. They are generally considered undesirable, as they adversely affect the permeability properties and performance of polymeric membranes for microfiltration, ultrafiltration, and reverse osmosis. However, MVs can be useful in certain thin-film applications in which vapor transmission is necessary, or for use as reservoirs for enzymes or liquid membrane material. If more could be learned about the nature and causes of MV formation, it might be possible to devise techniques to control and/or prevent MV formation that are more effective than those currently employed. Two hypotheses for the MV growth mechanism have been advanced. Reuvers proposed that once initiated, MV growth can be attributed to diffusion of (primarily) solvent to the MV nuclei. Because this mechanism does not involve gross movement of the MV, the presence or absence of body forces such as buoyancy should not significantly affect MV growth. On the other hand, Shojaie et al. proposed that solutocapillary convection induced by a steep surface-tension gradient along the MV/bulk solution interface enhances mass transfer to the growing MV. This interfacial convection exerts a force that pulls the growing MV downward into the casting solution. Both buoyancy and viscous drag hinder MV growth by inhibiting this motion. Thus, removing the buoyancy force by casting in microgravity should augment MV growth according to this hypothesis. Whereas neither surface tension nor gravity has a significant effect on MV growth according to the first hypothesis, buoyancy forces should be important if the second hypothesis is correct. The overall goal of this research is to test these two hypotheses in order to improve our understanding of the MV growth processing solvent-cast polymeric membranes. Studying MV growth in low-gravity conditions is pivotal to our ability to discriminate between these two hypotheses.

Pekny, M. R.↗

Developing strong concurrent multiphysics multiscale coupling to understand the impact of microstructural mechanisms on the structural scale

The heterogeneity in mechanical fields introduced by microstructure plays a critical role in the localization of deformation. To resolve this incipient stage of failure, it is therefore necessary to incorporate microstructure with sufficient resolution. On the other hand, computational limitations make it infeasible to represent the microstructure in the entire domain at the component scale. In this study, the authors demonstrate the use of concurrent multi- scale modeling to incorporate explicit, finely resolved microstructure in a critical region while resolving the smoother mechanical fields outside this region with a coarser discretization to limit computational cost. The microstructural physics is modeled with a high-fidelity model that incorporates anisotropic crystal elasticity and rate-dependent crystal plasticity to simulate the behavior of a stainless steel alloy. The component-scale material behavior is treated with a lower fidelity model incorporating isotropic linear elasticity and rate-independent J 2 plas- ticity. The microstructural and component scale subdomains are modeled concurrently, with coupling via the Schwarz alternating method, which solves boundary-value problems in each subdomain separately and transfers solution information between subdomains via Dirichlet boundary conditions. Beyond cases studies in concurrent multiscale, we explore progress in crystal plastic- ity through modular designs, solution methodologies, model verification, and extensions to Sierra/SM and manycore applications. Advances in conformal microstructures having both hexahedral and tetrahedral workflows in Sculpt and Cubit are highlighted. A structure-property case study in two-phase metallic composites applies the Materials Knowledge System to local metrics for void evolution. Discussion includes lessons learned, future work, and a summary of funded efforts and proposed work. Finally, an appendix illustrates the need for two-way coupling through a single degree of freedom.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

A general mechanistic framework for cross-scale understanding of hot spots and hot moments in carbon and water fluxes

Semi-arid ecosystems, like those in the American Southwest, exert a massive impact on the interannual variability of carbon and water cycling. Unfortunately, these carbon and water fluxes are notoriously difficult to predict due to their high spatial and temporal variability, which is poorly captured by the current generation of vegetation models. Indeed, this region is exemplified by the ‘hot spots and hot moments’ concept, which states that small areas in space (‘hot spots’) and transient moments in time (‘hot moments’) exert an outsized influence on biogeochemical cycling. However, the factors that regulate these pulses in biogeochemical activity are unknown, as is their variability across space and time. These uncertainties severely limit efforts to better represent hot spots and hot moments in models. Here, we seek to develop a generalized method for detecting and quantifying the importance of hot spots and hot moments from individual plant to regional scales. Underpinning this method is our recently developed statistical approach for identifying hot spots and hot moments. By applying this method to semi-continuous measurements of plant water status, a depth profile of soil water potential, and ecosystem fluxes via eddy covariance, we will track the fate of water through the soil-plant-atmosphere continuum and identify the mechanistic drivers of these transient pulses in biogeochemical activity. Then, we will expand this approach across a broad network of Ameriflux towers, and apply a machine learning approach that will allow us to upscale measurements of hot spots and hot moments across the American Southwest and quantify their impact on carbon and water cycles. These products will allow us to identify hot spots and hot moments across spatio-temporal scales and will serve as crucial data sources for validating a new generation of models that can better capture highly dynamic carbon and water fluxes. The proposed method will be easily transferable across biomes and will serve as a framework for future research on hot spots and hot moments across the plant ecophysiology, biometeorology, and vegetation modeling communities.

54 ENVIRONMENTAL SCIENCES↗

Advances in Hyperspectral Image Classification Methods for Vegetation and Agricultural Cropland Studies

Hyperspectral data are becoming more widely available via sensors on airborne and unmanned aerial vehicle (UAV) platforms, as well as proximal platforms. While space-based hyperspectral data continue to be limited in availability, multiple spaceborne Earth-observing missions on traditional platforms are scheduled for launch, and companies are experimenting with small satellites for constellations to observe the Earth, as well as for planetary missions. Land cover mapping via classification is one of the most important applications of hyperspectral remote sensing and will increase in significance as time series of imagery are more readily available. However, while the narrow bands of hyperspectral data provide new opportunities for chemistry-based modeling and mapping, challenges remain. Hyperspectral data are high dimensional, and many bands are highly correlated or irrelevant for a given classification problem. For supervised classification methods, the quantity of training data is typically limited relative to the dimension of the input space. The resulting Hughes phenomenon, often referred to as the curse of dimensionality, increases potential for unstable parameter estimates, overfitting, and poor generalization of classifiers. This is particularly problematic for parametric approaches such as Gaussian maximum likelihood–based classifiers that have been the backbone of pixel-based multispectral classification methods. This issue has motivated investigation of alternatives, including regularization of the class covariance matrices, ensembles of weak classifiers, development of feature selection and extraction methods, adoption of nonparametric classifiers, and exploration of methods to exploit unlabeled samples via semi-supervised and active learning. Data sets are also quite large, motivating computationally efficient algorithms and implementations. This chapter provides an overview of the recent advances in classification methods for mapping vegetation using hyperspectral data. Three data sets that are used in the hyperspectral classification literature (e.g., Botswana Hyperion satellite data and AVIRIS airborne data over both Kennedy Space Center and Indian Pines) are described in Section 3.2 and used to illustrate methods described in the chapter. An additional high-resolution hyperspectral data set acquired by a SpecTIR sensor on an airborne platform over the Indian Pines area is included to exemplify the use of new deep learning approaches, and a multiplatform example of airborne hyperspectral data is provided to demonstrate transfer learning in hyperspectral image classification. Classical approaches for supervised and unsupervised feature selection and extraction are reviewed in Section 3.3. In particular, nonlinearities exhibited in hyperspectral imagery have motivated development of nonlinear feature extraction methods in manifold learning, which are outlined in Section 3.3.1.4. Spatial context is also important in classification of both natural vegetation with complex textural patterns and large agricultural fields with significant local variability within fields. Approaches to exploit spatial features at both the pixel level (e.g., co-occurrence–based texture and extended morphological attribute profiles [EMAPs]) and integration of segmentation approaches (e.g., HSeg) are discussed in this context in Section 3.3.2. Recently, classification methods that leverage nonparametric methods originating in the machine learning community have grown in popularity. An overview of both widely used and newly emerging approaches, including support vector machines (SVMs), Gaussian mixture models, and deep learning based on convolutional neural networks is provided in Section 3.4. Strategies to exploit unlabeled samples, including active learning and metric learning, which combine feature extraction and augmentation of the pool of training samples in an active learning framework, are outlined in Section 3.5. Integration of image segmentation with classification to accommodate spatial coherence typically observed in vegetation is also explored, including as an integrated active learning system. Exploitation of multisensor strategies for augmenting the pool of training samples is investigated via a transfer learning framework in Section 3.5.1.2. Finally, we look to the future, considering opportunities soon to be provided by new paradigms, as hyperspectral sensing is becoming common at multiple scales from ground-based and airborne autonomous vehicles to manned aircraft and space-based platforms.

Pasolli, Edoardo↗

Increasing Cognitive Ability/Reserve Using Software – Pilot (ICARUS-Pilot)

BACKGROUND This research study was competitively awarded under the 2022 JSC Innovation Charge Account (ICA) program administered by NASA Johnson Space Center’s Joint Technology Working Group. Study period of performance was May through September 2022, with a maximum allowed procurement budget of $10K. The study sought to quantify and assess the potential benefit of using commercial-off-the-shelf (COTS) cognitive training software to improve cognitive performance in an astronaut-like terrestrial population. METHODS Five volunteer research participants were recruited from the JSC employee population to mimic certain demographic characteristics of the NASA astronaut population (age, education/discipline). Participant cognitive performance was assessed before and after executing eighteen sessions of remote cognitive training executed nominally three times per week using six exercises within an adaptive app-based COTS software package (BrainHQ, Posit Science) on study-provided tablets. Pre- and post-training cognitive performance was measured using internal assessments in BrainHQ as well as Cognition Test Battery (CTB) version ISS B01 v3 (3.0.9-201710021500), an independent software test developed specifically for NASA and used currently in research studies on astronauts. BrainHQ exercises were posited to map well or partially to several CTB sub-tests. Participants provided feedback on their study experience formally via semi-structured interview at the conclusion of testing and informally throughout the study if they encountered issues. RESULTS The enrolled ICARUS-Pilot study participants generally matched Artemis crew demographic characteristics. Four of five participants have completed study training and assessment activities as of the writing of this abstract. These test participants complied well with desired training session frequency and duration yielding an average cumulative active training duration of 15 hours over an average of 45 days; participants showed 78% average improvement in metric performance for the six trained exercises, with an associated overall 33%ile ranking increase against performance of the entire BrainHQ subscribing population for internal pre/post assessment, agreeing with post-study survey self-reported performance increases. CTB overall feedback scoring, not corrected for learning effects, showed an average of 19% performance improvement across its 10 performance measures over the training period for the completed participants. Detailed analyses will be conducted once participant data collection for the study is complete and the resulting dataset is fully populated. DISCUSSION These preliminary results provide a positive trend for the effectiveness of the training approach, but further analysis will be needed to establish significance, investigate far transfer, and suggest the needed participant pool size for subsequent efforts to achieve statistically significant outcomes given similar results. The pilot study has already been helpful by allowing the study team to learn a great deal about the capabilities and limitations of the COTS software package that will be reflected in future proposals along with revised timelines for study execution and test participant management. From participant feedback, one common thread regarding the COTS training was that it felt overly repetitive – future proposals should reassess overall training duration, available levels for each trained exercise, and the behavior of the BrainHQ internal scheduler in determining which exercises should be trained and for how long. If the final analysis of this feasibility study ultimately supports it, the study team will recommend further investigation to fully evaluate this potential countermeasure and optimize its implementation. Future proposals would cite this feasibility study’s outcome and would seek to refine the training protocol and obtain statistically significant results for cognitive performance increases as well as retention data.

cognitive training↗

Increasing Cognitive Ability/Reserve Using Software – Pilot (ICARUS-Pilot)

Background: This research study was competitively awarded under the 2022 JSC Innovation Charge Account (ICA) program administered by NASA Johnson Space Center’s Joint Technology Working Group. Study period of performance was May through September 2022, with a maximum allowed procurement budget of $10K. The study sought to quantify and assess the potential benefit of using commercial-off-the-shelf (COTS) cognitive training software to improve cognitive performance in an astronaut-like terrestrial population. Methods: Five volunteer research participants were recruited from the JSC employee population to mimic certain demographic characteristics of the NASA astronaut population (age, education/discipline). Participant cognitive performance was assessed before and after executing eighteen sessions of remote cognitive training executed nominally three times per week using six exercises within an adaptive app-based COTS software package (BrainHQ, Posit Science) on study-provided tablets. Pre- and post-training cognitive performance was measured using internal assessments in BrainHQ as well as Cognition Test Battery (CTB) version ISS B01 v3 (3.0.9-201710021500), an independent software test developed specifically for NASA and used currently in research studies on astronauts. BrainHQ exercises were posited to map well or partially to several CTB sub-tests. Participants provided feedback on their study experience formally via semi-structured interview at the conclusion of testing and informally throughout the study if they encountered issues. Results: The enrolled ICARUS-Pilot study participants generally matched Artemis crew demographic characteristics. Four of five participants have completed study training and assessment activities as of the writing of this abstract. These test participants complied well with desired training session frequency and duration yielding an average cumulative active training duration of 15 hours over an average of 45 days; participants showed 78% average improvement in metric performance for the six trained exercises, with an associated overall 33%ile ranking increase against performance of the entire BrainHQ subscribing population for internal pre/post assessment, agreeing with post-study survey self-reported performance increases. CTB overall feedback scoring, not corrected for learning effects, showed an average of 19% performance improvement across its 10 performance measures over the training period for the completed participants. Detailed analyses will be conducted once participant data collection for the study is complete and the resulting dataset is fully populated. Discussion: These preliminary results provide a positive trend for the effectiveness of the training approach, but further analysis will be needed to establish significance, investigate far transfer, and suggest the needed participant pool size for subsequent efforts to achieve statistically significant outcomes given similar results. The pilot study has already been helpful by allowing the study team to learn a great deal about the capabilities and limitations of the COTS software package that will be reflected in future proposals along with revised timelines for study execution and test participant management. From participant feedback, one common thread regarding the COTS training was that it felt overly repetitive – future proposals should reassess overall training duration, available levels for each trained exercise, and the behavior of the BrainHQ internal scheduler in determining which exercises should be trained and for how long. If the final analysis of this feasibility study ultimately supports it, the study team will recommend further investigation to fully evaluate this potential countermeasure and optimize its implementation. Future proposals would cite this feasibility study’s outcome and would seek to refine the training protocol and obtain statistically significant results for cognitive performance increases as well as retention data.

cognitive training↗

Exploratory analysis and performance prediction of big data transfer in High-performance Networks

Big data transfer in large-scale scientific and business applications is increasingly carried out over connections with guaranteed bandwidth provisioned in High-performance Networks (HPNs) via advance bandwidth reservation. Provisioning agents need to carefully schedule data transfer requests, compute network paths, and allocate appropriate bandwidths. Such reserved bandwidths, if not fully utilized, could be simply wasted due to the exclusive access during the approved time window, and cause extra overhead and complexity for resource management. This calls for accurate performance prediction to reserve bandwidths that match actual needs and avoid over-provisioning. We employ machine learning algorithms to predict big data transfer performance based on extensive performance measurements collected in the past several years from data transfer tests using different protocols and toolkits between various end sites on several real-life physical or emulated testbeds. We first analyze the performance patterns in response to a comprehensive list of parameters in end-host systems, network connections, and data transfer applications, which motivate the use of machine learning and also help us identify the effects of latent factors. We then propose threshold- and clustering-based methods to eliminate negative effects of latent factors in data preprocessing and build a robust performance predictor based on customized domain-oriented loss functions. The performance of the proposed methods is verified by extensive experiments using SVR and RFR as well as theoretical analysis of the general performance bound.

97 MATHEMATICS AND COMPUTING↗

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↗

An algorithm for physics informed scan path optimization in additive manufacturing

Site specific microstructure control is a critical research area within the field of additive manufacturing due to its potential to revolutionize part performance. One way to achieve site specific microstructure control is through control of the solidification conditions via the construction of intricate scan paths; however, the search space for such a problem is large. Previous attempts only considered the solidification conditions at the top surface while also requiring either lots of manual-fine tuning or large amounts of computational resources. This paper introduces a general method for scan path optimization which considers the solidification conditions in the bulk of the material without an increase in computational expense. This method consists of three core components:1. A heat transfer model for simulating the temperature field at a given time.2. A surrogate model which takes scan pattern information and temperature data and predicts the solidification conditions of the bulk as well as the meltpool depths for a spot melt.3. A decision algorithm to decide which spot melt should be printed next based on the outputs of the surrogate model.Each of these components can be changed without changing the overall method. Within this work, this method is applied in the creation of an algorithm containing a semi-analytic heat transfer model to simulate the temperature field, a fully convolutional neural network (FCNN) as the surrogate model, and a greedy decision algorithm. The resulting algorithm produced complex scan patterns which gave strong results for simulated microstructure control.

36 MATERIALS SCIENCE↗

Less is more: Sampling chemical space with active learning

We present the development of accurate and transferable machine learning (ML) potentials for predicting molecular energetics is a challenging task. The process of data generation to train such ML potentials is a task neither well understood nor researched in detail. In this work, we present a fully automated approach for the generation of datasets with the intent of training universal ML potentials. It is based on the concept of active learning (AL) via Query by Committee (QBC), which uses the disagreement between an ensemble of ML potentials to infer the reliability of the ensemble’s prediction. QBC allows the presented AL algorithm to automatically sample regions of chemical space where the ML potential fails to accurately predict the potential energy. AL improves the overall fitness of ANAKIN-ME (ANI) deep learning potentials in rigorous test cases by mitigating human biases in deciding what new training data to use. AL also reduces the training set size to a fraction of the data required when using naive random sampling techniques. To provide validation of our AL approach, we develop the COmprehensive Machine-learning Potential (COMP6) benchmark (publicly available on GitHub) which contains a diverse set of organic molecules. Active learning-based ANI potentials outperform the original random sampled ANI-1 potential with only 10% of the data, while the final active learning-based model vastly outperforms ANI-1 on the COMP6 benchmark after training to only 25% of the data. Finally, we show that our proposed AL technique develops a universal ANI potential (ANI-1x) that provides accurate energy and force predictions on the entire COMP6 benchmark. Finally, this universal ML potential achieves a level of accuracy on par with the best ML potentials for single molecules or materials, while remaining applicable to the general class of organic molecules composed of the elements CHNO.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

Unifying Combinatorial and Graphical Methods in Artificial Intelligence

Recently, a new graph Laplacian, called the inner product Laplacian, was introduced which generalizes many existing Laplacians, including the normalized and combinatorial Laplacian and their weighted variants. The key observation behind the inner product Laplacian is that by defining appropriate inner product spaces on the vertices and edges, the standard Laplacians can be recovered as Hodge Laplacians over the simplicial complex formed by the edges and vertices. These inner product spaces form a natural way to incorporate non-combinatorial information into the definition of a domain-specific Laplacian. In particular, in contrast to current domain-specific weighting schemes which rely solely on edge weights, information regarding the similarity of non-adjacent vertices and arbitrary pairs of edges can be effectively incorporated into the Laplacian. In order to illustrate this approach we consider the problem of calculating the potential energy of an atomistic configuration using Graph Neural Networks. In comparison with start-of-the-art approaches, such as SchNet, our approach replaces a learned (via auto-encoder) representation of the atom types with an inner product space on atoms based on scientific knowledge (e.g., electronegativity). We will illustrate how this approach captures key chemical properties of the molecules and compare the energy calculations with state-of-the-art neural network approaches. However, to compute the resulting Laplacian involves a mixture of sparse and dense matrix computation and yields a dense matrix as the basis for the graph convolution. This dense convolutional kernel necessitates moving away from the standard message passing framework for graph neural networks and increases the computational cost of applying the kernel. In order to mitigate these costs we investigate means of leveraging the mixed sparse and dense computations to reduce the overall computational cost and how these approaches can be automatically transferred to energy efficient hardware (e.g., field programmable gate arrays (FPGAs)).

97 MATHEMATICS AND COMPUTING↗

Molière scattering in quark-gluon plasma: finding point-like scatterers in a liquid

By finding rare (but not exponentially rare) large-angle deflections of partons within a jet produced in a heavy ion collision, or of such a jet itself, experimentalists can find the weakly coupled short-distance quark and gluon particles (scatterers) within the strongly coupled liquid quark-gluon plasma (QGP) produced in heavy ion collisions. This is the closest one can come to probing QGP via a scattering experiment and hence is the best available path toward learning how a strongly coupled liquid emerges from an asymptotically free gauge theory. The short-distance, particulate, structure of liquid QGP can be revealed in events in which a jet parton resolves, and scatters off, a parton from the droplet of QGP. The probability for picking up significant transverse momentum via a single scattering was calculated previously, but only in the limit of infinite parton energy which means zero angle scattering. Here, we provide a leading order perturbative QCD calculation of the Molière scattering probability for incident partons with finite energy, scattering at a large angle. We set up a thought experiment in which an incident parton with a finite energy scatters off a parton constituent within a “brick” of QGP, which we treat as if it were weakly coupled, as appropriate for scattering with large momentum transfer, and compute the probability for a parton to show up at a nonzero angle with some energy. We include all relevant channels, including those in which the parton that shows up at a large angle was kicked out of the medium as well as the Rutherford-like channel in which what is seen is the scattered incident parton. The results that we obtain will serve as inputs to future jet Monte Carlo calculations and can provide qualitative guidance for how to use future precise, high statistics, suitably differential measurements of jet modification in heavy ion collisions to find the scatterers within the QGP liquid.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗