SEARCH · Engineering Papers
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.
The purpose of lessons learned is to identify insight gained during a project – successes or failures – that can be applied on future projects. Lessons learned can contribute to the overall success of a project by building on approaches that have worked well and avoiding previous mistakes. Below are examples of lessons learned during ERDF’s ARRA-funded expansion project.
The cost of nuclear power is a straightforward yet complicated topic. It is straightforward in that the cost of nuclear power is a function of the cost to build the nuclear power plant, the cost to operate and maintain it, and the cost to provide fuel for it. It is complicated in that some of those costs are not necessarily known, introducing uncertainty into the analysis. For large light water reactor (LWR)-based nuclear power plants, the uncertainty is mainly contained within the cost of construction. The typical costs of operations and maintenance (O&M), as well as fuel, are well known based on the current fleet of LWRs. However, the last currently operating reactor to come online was Watts Bar 1 in May 1996; thus, the expected construction costs for gigawatt (GW)-class reactors in the United States are based on information nearly two decades old. Extrapolating construction, O&M, and fuel costs from GW-class LWRs to LWR-based small modular reactors (SMRs) introduces even more complication. The per-installed-kilowatt construction costs for SMRs are likely to be higher than those for the GW-class reactors based on the property of the economy of scale. Generally speaking, the economy of scale is the tendency for overall costs to increase slower than the overall production capacity. For power plants, this means that doubling the power production capacity would be expected to cost less than twice as much. Applying this property in the opposite direction, halving the power production capacity would be expected to cost more than half as much. This can potentially make the SMRs less competitive in the electricity market against the GW-class reactors, as well as against other power sources such as natural gas and subsidized renewables. One factor that can potentially aid the SMRs in achieving economic competitiveness is an economy of numbers, as opposed to the economy of scale, associated with learning curves. The basic concept of the learning curve is that the more a new process is repeated, the more efficient the process can be made. Assuming that efficiency directly relates to cost means that the more a new process is repeated successfully and efficiently, the less costly the process can be made. This factor ties directly into the factory fabrication and modularization aspect of the SMR paradigm—manufacturing serial, standardized, identical components for use in nuclear power plants can allow the SMR industry to use the learning curves to predict and optimize deployment costs.
Limited placement of real-time monitoring devices in the distribution grid, recent trends notwithstanding, has prevented the easy implementation of demand-response and other smart grid applications. Part I of this paper discusses the problem of learning the operational structure of the grid from nodal voltage measurements. In this work (Part II), the learning of the operational radial structure is coupled with the problem of estimating nodal consumption statistics and inferring the line parameters in the grid. Based on a Linear-Coupled(LC) approximation of AC power flows equations, polynomial time algorithms are designed to identify the structure and estimate nodal load characteristics and/or line parameters in the grid using the available nodal voltage measurements. Then the structure learning algorithm is extended to cases with missing data, where available observations are limited to a fraction of the grid nodes. The efficacy of the presented algorithms are demonstrated through simulations on several distribution test cases.
Intrusion/anomaly detection systems are among the first lines of cyber defense. Commonly, they either use signatures or machine learning (ML) to identify threats, but fail to account for sophisticated attackers trying to circumvent them. We propose to embed machine learning within a game theoretic framework that performs adversarial modeling, develops methods for optimizing operational response based on ML, and integrates the resulting optimization codebase into the existing ML infrastructure developed by the Hybrid LDRD. Our approach addresses three key shortcomings of ML in adversarial settings: 1) resulting classifiers are typically deterministic and, therefore, easy to reverse engineer; 2) ML approaches only address the prediction problem, but do not prescribe how one should operationalize predictions, nor account for operational costs and constraints; and 3) ML approaches do not model attackers’ response and can be circumvented by sophisticated adversaries. The principal novelty of our approach is to construct an optimization framework that blends ML, operational considerations, and a model predicting attackers reaction, with the goal of computing optimal moving target defense. One important challenge is to construct a realistic model of an adversary that is tractable, yet realistic. We aim to advance the science of attacker modeling by considering game-theoretic methods, and by engaging experimental subjects with red teaming experience in trying to actively circumvent an intrusion detection system, and learning a predictive model of such circumvention activities. In addition, we will generate metrics to test that a particular model of an adversary is consistent with available data.
Efficient parallel implementations of scientific applications on multi-core CPUs with accelerators such as GPUs and Xeon Phis is challenging. This requires - exploiting the data parallel architecture of the accelerator along with the vector pipelines of modern x86 CPU architectures, load balancing, and efficient memory transfer between different devices. It is relatively easy to meet these requirements for highly structured scientific applications. In contrast, a number of scientific and engineering applications are unstructured. Getting performance on accelerators for these applications is extremely challenging because many of these applications employ irregular algorithms which exhibit data-dependent control-ow and irregular memory accesses. Furthermore, these applications are often iterative with dependency between steps, and thus making it hard to parallelize across steps. As a result, parallelism in these applications is often limited to a single step. Numerical simulation of charged particles beam dynamics is one such application where the distribution of work and memory access pattern at each time step is irregular. Applications with these properties tend to present significant branch and memory divergence, load imbalance between different processor cores, and poor compute and memory utilization. Prior research on parallelizing such irregular applications have been focused around optimizing the irregular, data-dependent memory accesses and control-ow during a single step of the application independent of the other steps, with the assumption that these patterns are completely unpredictable. We observed that the structure of computation leading to control-ow divergence and irregular memory accesses in one step is similar to that in the next step. It is possible to predict this structure in the current step by observing the computation structure of previous steps. In this dissertation, we present novel machine learning based optimization techniques to address the parallel implementation challenges of such irregular applications on different HPC architectures. In particular, we use supervised learning to predict the computation structure and use it to address the control-ow and memory access irregularities in the parallel implementation of such applications on GPUs, Xeon Phis, and heterogeneous architectures composed of multi-core CPUs with GPUs or Xeon Phis. We use numerical simulation of charged particles beam dynamics simulation as a motivating example throughout the dissertation to present our new approach, though they should be equally applicable to a wide range of irregular applications. The machine learning approach presented here use predictive analytics and forecasting techniques to adaptively model and track the irregular memory access pattern at each time step of the simulation to anticipate the future memory access pattern. Access pattern forecasts can then be used to formulate optimization decisions during application execution which improves the performance of the application at a future time step based on the observations from earlier time steps. In heterogeneous architectures, forecasts can also be used to improve the memory performance and resource utilization of all the processing units to deliver a good aggregate performance. We used these optimization techniques and anticipation strategy to design a cache-aware, memory efficient parallel algorithm to address the irregularities in the parallel implementation of charged particles beam dynamics simulation on different HPC architectures. Experimental result using a diverse mix of HPC architectures shows that our approach in using anticipation strategy is effective in maximizing data reuse, ensuring workload balance, minimizing branch and memory divergence, and in improving resource utilization.
Although increased availability of computational resources has enabled high-fidelity simulations (e.g., large eddy simulations) of turbulent flows, the Reynolds-Averaged Navier–Stokes (RANS) models are still the dominant tools in industrial applications. However, the predictive capabilities of RANS models are limited by large model-form discrepancies due to the Reynolds stress closure. Recently, a Physics-Informed Machine Learning (PIML) approach has been proposed to learn the functional form of Reynolds stress discrepancy in RANS simulations based on available data. It has been demonstrated that the learned discrepancy function can be used to improve Reynolds stresses in new flows where data are not available. However, due to a number of challenges, the improvements are only demonstrated in the Reynolds stress prediction but not in corresponding propagated quantities of interest (e.g., velocity field). In this work, we investigate the prediction performance on the velocity field by propagating the corrected Reynolds stresses in PIML approach. To enrich the input features, an integrity basis of invariants is implemented. The fully developed turbulent flow in a square duct is used as the test case. The discrepancy model is trained on flow fields from several Reynolds numbers and evaluated on a duct flow at a higher Reynolds number than any of the training cases. The predicted Reynolds stresses are propagated to velocity field via RANS equations. Numerical results show excellent predictive performances in both Reynolds stresses and their propagated velocities, demonstrating the merits of the PIML approach in predictive turbulence modeling.
Within the DOE complex, 3013 canisters are used to store Pu-bearing waste. Wall penetration due to corrosion, specifically stress corrosion cracking, is considered to be the most likely cause of failure over the 50-year lifetime of the canisters. This project has 2 objectives: The first is the development of machine learning algorithms to identify the presence of corrosion from a very large set of images generated by laser confocal microscope scanning of 3013 canister used to store Pu oxides. This portion of the LDRD constitutes image analysis of a metal surface for the presence of corrosion. The image processing algorithms developed for this project will provide a suitable basis for analysis of other types of corrosion data, which can be produced in vast quantities using modern devices. The second component of the LDRD consists of the development of machine learning algorithms that obtain molecular mechanics force-fields from ab-initio Density Functional Theory (DFT) calculations for corrosive attack by chlorides on 304L or 316L stainless steel. The goal of this latter component is to determine means for mitigating corrosion on a fundamental level, including coatings, welding methods, metal composition, etc. Force-field modeling is necessary for this endeavor because the incipience and progression of corrosion is the governed by molecular structures, grain boundaries, dislocations and surface structures represented by large numbers of atoms. Although DFT calculations are extremely adept at describing molecular scale processes, they are computationally expensive making them ill-suited for calculations of more than several hundred atoms, especially for the repeated applications inherent in material design. Fortunately, force-field methods provide an avenue for viable molecular-scale calculations involving the numbers of atoms involved in corrosion processes. The accuracy of the force-field calculations depends strongly on the accuracy of the force-field model, which is extremely difficult and time-consuming to derive from either data or ab-initio calculations. The use of machine learning algorithms has the potential to make the calculation of force-fields much more efficient but has not been explored significantly for corrosion processes. If shown to be viable, the technique would have wide -ranging impact on design of corrosion resistant materials and on the mitigation of corrosion in existing process systems.
Large-scale particle physics experiments face challenging demands for high-throughput computing resources both now and in the future. New heterogeneous computing paradigms on dedicated hardware with increased parallelization, such as Field Programmable Gate Arrays (FPGAs), offer exciting solutions with large potential gains. The growing applications of machine learning algorithms in particle physics for simulation, reconstruction, and analysis are naturally deployed on such platforms. We demonstrate that the acceleration of machine learning inference as a web service represents a heterogeneous computing solution for particle physics experiments that requires minimal modification to the current computing model. As examples, we retrain the ResNet-50 convolutional neural network to demonstrate state-of-the-art performance for top quark jet tagging at the LHC, and apply a ResNet-50 model with transfer learning for neutrino event classification. Using Project Brainwave b y Microsoft to accelerate the ResNet-50 image classification model, we achieve average inference times of 60 (10) milliseconds with our experimental physics software framework using Brainwave as a cloud (edge or on-premises) service, representing an improvement by a factor of approximately 30 (175) in model inference latency over traditional CPU inference in current experimental hardware. A single FPGA service accessed by many CPUs achieves a throughput of 600--700 inferences per second using an image batch of one, comparable to large batch-size GPU throughput and significantly better than small batch-size GPU throughput. Deployed as an edge or cloud service for the particle physics computing model, coprocessor accelerators can have a higher duty cycle and are potentially much more cost-effective.
The search for simple and elegant equations to explain and predict observable natural phenomena has been a driving force in scientific theory for several centuries. The great successes of Newton's laws, Maxwell's equations, the Schrodinger equation, and the theory of General Relativity have encouraged this perspective, and perhaps a beautiful equation comprising a “theory of everything” remains to be discovered. But now scientists are grappling with systems of ever-increasing complexity, fraught with nonlinearity, and with the quantities of interest often only indirectly observed. Nature may resist a simple description, and the most important discoveries of the next century may be complex theories with countless variables and parameters. The era of big data opens up a promising new approach to scientific discovery in this setting. The predictions of modern theories, even if complex or nonlinear, can be examined through detailed and computationally-intensive simulations that take days or weeks to run and fill petabytes of disk space. The upcoming generation of scientific instruments will provide petabytes of observations at unprecedented resolution and depth. The challenge is to compare data to simulations so as to test theories and identify those that best match the observations. The data are numerous and rich but are subject to noise and various systematic biases. The simulations are large and detailed but are so costly that the effective sample size is small. Moreover, the mapping from data to simulation to the parameters of interest is typically complicated and ill-conditioned. Although these challenges, and the solutions we propose, apply across the sciences, we will ground our efforts by focusing on cosmology. Here, we are truly entering the era of big data in several senses. Ever larger experiments survey larger and larger cosmological volumes, resulting in enormous amounts of pixel data to process and extract the information of cosmological interest. These include large imaging surveys, which collect all the light in broad wavelength ranges over significant fractions of the sky, and spectroscopic surveys, which measure the light as a function of wavelength to map out 3D distances. These experiments are designed to shed light on the big unanswered questions in cosmology. Distilling the petabytes of data into observed quantities for comparison with theory is a daunting task. But these large experiments do not only result in large amounts of data to analyze: if we wish to compare to theories of physics, then we need increasingly extensive predictions for the observed quantities as well, which requires ever-larger cosmological simulations that take substantial -- and at times prohibitive -- computational resources to produce. Our goal is to develop statistical and machine learning methods for using observed and simulated data to advance machine learning with applications cosmology. In particular, we focus our research efforts on the following tasks: The challenge for automated science is that it is computationally impossible and statistically dangerous to consider every possible model in order to find the best one. We will develop Bayesian Optimization based active-learning methods that accelerate both the execution of the simulations and the search for best-fitting parameters. The key idea is to make the simulations adaptive -- across resolutions, time, and parameters -- using the data to search as the simulation runs; Most ML methods operate on simple finite dimensional feature vectors. However, many cosmology and other science applications require ML methods that can operate on more complex objects as inputs or outputs such as functions, distributions, or sets and point clouds. Our goal is to develop efficient ML methods for this problem and demonstrate their applicability in Cosmology, Astrophysics, and other science problems.
ExxonMobil Research and Engineering Company, via a technical services agreement, engages Idaho National Laboratory (INL) to provide general industry knowledge, best practices, and lessons learned in cellulosic ethanol plant design, project execution, and operations. INL will also assess publicly proposed cellulosic plant designs. The scope of work will be accomplished in three tasks: Task 1, Communicate industry knowledge; Task 2, Present high-level lessons learned; Task 3, Provide an engineering assessment of publicly proposed ethanol plant designs. This Task 1 report provides a high-level summary of general industry knowledge, best practice, and lessons learned in biomass feedstock logistics and integration with bioconversion processes.
Software flaw detection using multimodal deep learning models has been demonstrated as a very competitive approach on benchmark problems. In this work, we demonstrate that even better performance can be achieved using neural architecture search (NAS) combined with multimodal learning models. We adapt a NAS framework aimed at investigating image classification to the problem of software flaw detection and demonstrate improved results on the Juliet Test Suite, a popular benchmarking data set for measuring performance of machine learning models in this problem domain.
The development of algorithms for machine learning and data analysis for the 3013 MIS corrosion surveillance program is a collaborative effort by SRNL, USC and GT. For corrosion detection, LCM image data is extracted from large binary files, with software written to convert the data to physical attributes (i.e. height, color and grayscale values; all as functions of a location in a plane projection). The user interface for the software permits selective downloading of binary data and interrogation of attributes. User input thresholds are used to flag attributes of interest. Machine learning algorithms, developed for this application, are used to determine whether the features are the result of corrosion. To address the fundamental mechanisms of corrosion, machine learning algorithms are being developed to derive interatomic potential force-fields from ab-initio DFT calculations. The goal is to apply molecular modeling on a large enough scale to guide the design of resistant materials.
We present a novel machine learning CNN architecture that can learn from limited data combined appropriately with physics and statistical priors (e.g., forward models and noise models). To address the limited availability of training data we adopt a 3D patch-based approach for our models. Patch-based learning is central to several image reconstruction methods and demands fewer training data than DL approaches, as a single data volume can be broken into several millions of overlapping 3D sub-volumes or patches. This creates a very large number of training sub-volumes from a limited number of overall image volumes. A 3D Generative Adversarial Networks (GAN) is then trained to remove artifacts at the sub-volume level. The combination of a sub-volume-based approach with DL allows us to exploit the richness of the latter in extracting and representing image features, while avoiding risks associated with overfitting due to limited training data.
Focal Areas: We plan to use machine learning (ML) to disentangle the impact of climate change on plant water use efficiency (WUE) across the leaf level, structured canopy, and community ecosystem scales. Due to the complex behavior of modeling plant WUE’s connection to carbon storage (and hidden co-varying interactions) we suggest that physics-based deep learning must be utilized. The second topic we propose is, based on a better understanding of WUE response to climate change, there is a need to predict how WUE will limit or enhance the recycling of water from transpiration by vegetation back to the land surface worldwide, and thus impact water storage capacity (i.e., natural and reservoirs), using machine learning optimization techniques such as genetic algorithms.
This whitepaper addresses the following focal area: (1) Data acquisition and assimilation enabled by machine learning, AI, and advanced methods including experimental/network design/optimization, unsupervised learning (including deep learning), and hardware-related efforts involving AI (e.g., edge computing)
Learning 3D representations that generalize well to arbitrarily oriented inputs is a challenge of practical importance in applications varying from computer vision to physics and chemistry. We propose a novel multi-resolution convolutional architecture for learning over concentric spherical feature maps, of which the single sphere representation is a special case. Our hierarchical architecture is based on alternatively learning to incorporate both intra-sphere and inter-sphere information. We show the applicability of our method for two different types of 3D inputs, mesh objects, which can be regularly sampled, and point clouds, which are irregularly distributed. We also propose an efficient mapping of point clouds to concentric spherical images, thereby bridging spherical convolutions on grids with general point clouds. We demonstrate the effectiveness of our approach in improving state-of-the-art performance on 3D classification tasks with rotated data.
This project aimed to develop and demonstrate a Data Integration, Assimilation, and Learning framework for geologic carbon sequestration projects (DIAL-GCS). DIAL-GCS is an intelligence monitoring system (IMS) for automating GCS closed-loop management by leveraging recent developments in machine learning technologies, complex event processing (CEP), and reduced-order modeling. The safe and efficient operation of GCS repositories requires integrated monitoring to track the injected CO¬2 as it moves within a storage reservoir. GCS projects are data intensive, as a result of proliferation of digital instrumentation and smart-sensing technologies. GCS projects are also resource intensive, often requiring multidisciplinary teams performing different monitoring, verification, accounting (MVA) tasks throughout the lifecycle of a project to ensure secure containment of injected CO2. The success of GCS thus depends in a large part on our ability to access, assimilate, and analyze heterogeneous data and information sources in a timely manner. This project included a number of meaningful and necessary tasks to transform the human domain knowledge into machine-interpretable rules for automating knowledge extraction and discovery in GCS. The specific technical objectives of the proposed DIAL-GCS project were to develop an ontology-driven GCS data management module for storing, querying, and exchanging GCS data (both historic and live sensor data) from multiple sources and in heterogeneous formats. Incorporate a CEP engine for detecting abnormal situations by seamlessly combining expert knowledge, rule-based reasoning, and machine learning. Enable uncertainty quantification and predictive analytics using a combination of coupled-process modeling, AI/ML methods, and reduced-order modeling, and integrate and demonstrate the system’s capabilities with both real and simulated data. As far as we know, this is one of the first projects aimed to develop intelligent monitoring systems (IMS) targeting the GCS. Under this project, the team had developed a large number of web applications and scientific algorithms that contribute the main theme of intelligent monitoring. The team has published more than a dozen peer reviewed papers and disseminated the research results at multiple technical meetings.
This report details the results of a three-fold investigation of sensitivity analysis (SA) for machine learning (ML) explainability (MLE): (1) the mathematical assessment of the fidelity of an explanation with respect to a learned ML model, (2) quantifying the trustworthiness of a prediction, and (3) the impact of MLE on the efficiency of end-users through multiple users studies. We focused on the cybersecurity domain as the data is inherently non-intuitive. As ML is being using in an increasing number of domains, including domains where being wrong can elicit high consequences, MLE has been proposed as a means of generating trust in a learned ML models by end users. However, little analysis has been performed to determine if the explanations accurately represent the target model and they themselves should be trusted beyond subjective inspection. Current state-of-the-art MLE techniques only provide a list of important features based on heuristic measures and/or make certain assumptions about the data and the model which are not representative of the real-world data and models. Further, most are designed without considering the usefulness by an end-user in a broader context. To address these issues, we present a notion of explanation fidelity based on Shapley values from cooperative game theory. We find that all of the investigated MLE explainability methods produce explanations that are incongruent with the ML model that is being explained. This is because they make critical assumptions about feature independence and linear feature interactions for computational reasons. We also find that in deployed, explanations are rarely used due to a variety of reason including that there are several other tools which are trusted more than the explanations and there is little incentive to use the explanations. In the cases when the explanations are used, we found that there is the danger that explanations persuade the end users to wrongly accept false positives and false negatives. However, ML model developers and maintainers find the explanations more useful to help ensure that the ML model does not have obvious biases. In light of these findings, we suggest a number of future directions including developing MLE methods that directly model non-linear model interactions and including design principles that take into account the usefulness of explanations to the end user. We also augment explanations with a set of trustworthiness measures that measure geometric aspects of the data to determine if the model output should be trusted.