Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “engine control”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,135 records · Page 63

The Belgian Research and Development Feasibility Programme for the Geological Disposal of High-Level and Long-Lived Radioactive Waste - 12338

ONDRAF/NIRAS, the Belgian Agency for Radioactive Waste and Enriched Fissile Materials, considers geological disposal in the poorly indurated Boom Clay as the reference solution for the long-term management of high-level and/or long-lived radioactive waste. To develop a safety concept and design for geological disposal, ONDRAF/NIRAS follows an iterative process demonstrating that the repository will be both safe and feasible to implement. This process is called the safety strategy. A part of the safety strategy is the feasibility programme which aims at demonstrating, at a conceptual level, that the proposed geological disposal system can be constructed, operated and progressively closed. The followed methodology is based on the substantiation of a hierarchy of feasibility statements. These statements cover all activities from the removal of primary waste packages from interim storage buildings to the closure of the disposal site and a period of institutional control. They focus on engineering practicability, health and safety and environmental considerations, costs and quality assurance issues. A 4 year research project to support the R and D feasibility programme was launched in 2009 with several international partners coordinated by ONDRAF/NIRAS. It aims at confirming that there are no fundamental flaws or show-stoppers in the feasibility of building and operating the facilities for geological disposal in the Boom Clay. (authors)

12 MANAGEMENT OF RADIOACTIVE AND NON-RADIOACTIVE W↗

SUMMARY REPORT FOR ZINC 65 CONTAMINATION CONTROL

Radioactive zinc, {sup 65}Zn, was detected after extraction of 215 TPBARs in from TVA reactor fuel cycle 6. A team consisting of Tritium Engineering, Tritium Operations, Tritium Radiation Control, and Savannah River National Laboratory personnel evaluated the risk and response and developed short, medium and long term goals for contamination control. One of the goals was incorporated into site Performance Based Incentive CO 3.4, to optimize the filter geometry and operating conditions for the Tritium Extraction Facility. This goal included a scoping study to determine if the contamination could be contained within the high radiation environment of the furnace module as well. In order to optimize the filters studies were conducted to independently evaluate the effect of pore size on pumping efficiency and zinc trapping efficiency (1). A study was also conducted to evaluate the effect of temperature on the trapping efficiency and adhesion (2). In addition, the potential for chemically trapping zinc in the lithium trap was evaluated using a thermodynamic study (3) followed by preliminary experimental testing (4). Based on the work that was completed it is determined that a 20 {mu}m filter heated to between 120 and 200 C will act as an effective physical trap for zinc vapors. It may be possible to chemically react zinc with copper or cobalt to form zinc intermetallic compounds or alloys but additional work under more prototypic conditions are required.

11 NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗

Force Control and Nonlinear Master-Slave Force Profile to Manage an Admittance Type Multi-Fingered Haptic User Interface

Natural movements and force feedback are important elements in using teleoperated equipment if complex and speedy manipulation tasks are to be accomplished in remote and/or hazardous environments, such as hot cells, glove boxes, decommissioning, explosives disarmament, and space to name a few. In order to achieve this end the research presented in this paper has developed an admittance type exoskeleton like multi-fingered haptic hand user interface that secures the user’s palm and provides 3-dimensional force feedback to the user’s fingertips. Atypical to conventional haptic hand user interfaces that limit themselves to integrating the human hand’s characteristics just into the system’s mechanical design this system also perpetuates that inspiration into the designed user interface’s controller. This is achieved by manifesting the property differences of manipulation and grasping activities as they pertain to the human hand into a nonlinear master-slave force relationship. The results presented in this paper show that the admittance-type system has sufficient bandwidth that it appears nearly transparent to the user when the user is in free motion and when the system is subjected to a manipulation task, increased performance is achieved using the nonlinear force relationship compared to the traditional linear scaling techniques implemented in the vast majority of systems.

42 ENGINEERING↗

Effect of skin flexibility on aerodynamic performance of flexible skin flapping wings for micro air vehicles

As part of the ongoing research on micro air vehicles, the present work focuses on the effect of membrane flexibility on the aerodynamic performance of flexible latex flapping wings. Wings with membrane thicknesses 0.37, 0.28, and 0.13 mm are chosen, which are named as least flexible (A), flexible (B), and most flexible (C), respectively. The experiments are performed in an air chamber of size 1.5 m × 1.5 m × 1.5 m, facilitated with wind velocities up to 15 m/s. The time-averaged lift and drag as functions of flapping frequency, forward flight velocity, the angles of attack (AoA), and advance ratio (J). The novel electronic control system developed previously is used to monitor and measure the flapping frequency. It is found that the effect of flexibility on the aerodynamic performance mainly depends on the range of flight speed; at 7200 ≤ Re ≤ 18,000, the lift and drag increase with increase of flexibility, and at 18,000 ≤ Re ≤ 25,200, the lift decreases and drag increases with increase of flexibility. Hence latex compliant (Wing C) wings are advantageous in the low Re range, while the least flexible wing (Wing A) is preferable for higher range.

42 ENGINEERING↗

Tests of cosmic ray radiography for power industry applications

In this report, we assess muon multiple scattering tomography as a non-destructive inspection technique in several typical areas of interest to the nuclear power industry, including monitoring concrete degradation, gate valve conditions, and pipe wall thickness. This work is motivated by the need for imaging methods that do not require the licensing, training, and safety controls of x-rays, and by the need to be able to penetrate considerable overburden to examine internal details of components that are otherwise inaccessible, with minimum impact on industrial operations. In some scenarios, we find that muon tomography may be an attractive alternative to more typical measurements.

36 MATERIALS SCIENCE↗

An 802pW 93% Peak Efficiency Buck Converter with 5.5×10 6 Dynamic Range Featuring Fast DVFS and Asynchronous Load-Transient Control

Here this paper presents a buck converter with sub-nW quiescent power, high efficiency, and a wide dynamic range for ultra-low-power (ULP) IoT SoCs. To optimize the SoC power consumption, the buck converter supports fast dynamic voltage and frequency scaling (DVFS) and enables fast load-transient response (FLTR) through asynchronous control. In addition, the buck converter is fully self-contained with all features integrated on chip including a proposed adaptive deadtime controller. Fabricated in 65nm CMOS, measurement results show the buck converter has an 802pW quiescent power at 1.5V input voltage and a 93% peak efficiency. The measured dynamic range is from 0.5nW to 2.75mW, which is over 6 orders of magnitude. The measured voltage droop is 54mV for a 45nA-to-1mA load current step thanks to the asynchronous load-transient detector. The buck converter achieves the highest efficiency and widest dynamic range among all the state-of-the-art sub-nW switching voltage regulators, which makes it well suited for power management in ULP SoCs.

42 ENGINEERING↗

Liquid balance monitoring inside conventional, Retrofit, and bio-reactor landfill cells

Highlights: • The Retrofit, Control, and As-Built cells received 48, 14, and 213 L Mg{sup −1} (liters of liquids per metric ton of waste). • The leachate collection system yielded 60, 57 and 198 L Mg{sup −1} from the Retrofit, Control, and As-Built cells. • The head on liner in all cells was below regulatory limits. • Measured moisture content of the waste samples was consistent with that calculated from accumulated liquid by balance. • The in-place saturated hydraulic conductivity of the MSW was calculated to be in the range of 10{sup −8} to 10{sup −7} m s{sup −1}. - Abstract: The Outer Loop landfill bioreactor (OLLB) in Louisville, KY, USA has been the site of a study to evaluate long-term bioreactor performance at a full-scale operational landfill. Three types of landfill units were studied including a conventional landfill (Control cell), a new landfill area that had an air addition and recirculation piping network installed as waste was being placed (As-Built cell), and a conventional landfill that was modified to allow for liquids recirculation (Retrofit cell). During the monitoring period, the Retrofit, Control, and As-Built cells received 48, 14, and 213 L Mg{sup −1} (liters of liquids per metric ton of waste), respectively. The leachate collection system yielded 60, 57 and 198 L Mg{sup −1} from the Retrofit, Control, and As-Built cells, respectively. The head on liner in all cells was below regulatory limits. In the Control and As-Built cells, leachate head on liner decreased once waste placement stopped. The measured moisture content of the waste samples was consistent with that calculated from the estimate of accumulated liquid by the liquid balance. Additionally, measurements on excavated solid waste samples revealed large spatial variability in waste moisture content. The degree of saturation in the Control cells decreased from 85% to 75%. The degree of saturation increased from 82% to 83% due to liquids addition in the Retrofit cells and decreased back to 80% once liquid addition stopped. In the As-Built cells, the degree of saturation increased from 87% to 97% during filling activities and then started to decrease soon after filling activities stopped to reach 92% at the end of the monitoring period. The measured leachate generation rates were used to estimate an in-place saturated hydraulic conductivity of the MSW in the range of 10{sup −8} to 10{sup −7} m s{sup −1} which is lower than previous reports. In the Control and Retrofit cells, the net loss in liquids, 43 and 12 L Mg{sup −1}, respectively, was similar to the measured settlement of 15% and 5–8% strain, respectively (Abichou et al., 2013). The increase in net liquid volume in the As-Built cells indicates that the 37% (average) measured settlement strain in these cells cannot be due to consolidation as the waste mass did not lose any moisture but rather suggests that settlement was attributable to lubrication of waste particle contacts, softening of flexible porous materials, and additional biological degradation.

12 MANAGEMENT OF RADIOACTIVE AND NON-RADIOACTIVE W↗

Molecular scale insight of pore morphology relation with mechanical properties of amorphous silica using ReaxFF

Porous materials are typically heterogeneous and they contain large variations of micro-/nano-pore structures, causing complicated behaviors. In continuum models, most mechanical properties of porous materials are estimated based on porosity, while the variations of micro/nano structures are ignored. That could be problematic as the microscopic heterogeneity may affect the mechanical response of porous materials. Thus, understanding micro/nano heterogeneity impact has been the focus in many scientific and engineering subjects. In the present study, we investigated the effect of nanopore structure (including pore shape and orientation) as well as porosity on mechanical properties of amorphous silica (a-SiO 2 ). The pore sizes in our simulations are comparable to the corresponding ones observed in a-SiO 2 based materials. We found that the existing of nanopores strongly influences Young’s modulus (E) and critical energy release rate (G IC ). These properties decrease with increasing porosity. Importantly, the impact of nanopores was characterized by structural parameters of porous materials. In addition to dependency on porosity, Young’s modulus also was found to vary as a function of potential energy per atom, which highly depends on nanopore shape. Furthermore, critical energy release rate was found to increase with increasing ligament length (also known as pore wall thickness). Finally, the results highlighted the importance of nanopore structures, which must be taken into account when studying fracture mechanisms in porous materials. Based on our findings, it was proposed that mechanical properties of porous materials can be controlled by nano-engineering pore structures.

36 MATERIALS SCIENCE↗

Multitarget control models for building thermal comfort and indoor air quality - A critical review

The control techniques in buildings contribute significantly to thermal comfort and indoor air quality (IAQ). However, the gaps are existing for multitarget controls considering both thermal comfort and IAQ. They are: (1) both experimental and modeling control studies were conducted for thermal comfort, focusing on temperature and humidity. (2) All three (physical, grey-box, black-box) modeling approaches were investigated for temperature control. (3) Physical and grey-box modeling approach were adopted for humidity control. (4) physical models were developed for VOCs and CO2 control. (5) grey-box and black-box models were lacking for VOCs and CO2. (6) Multi-target controls were lacking for temperature, humidity and CO2s. (7) Limited studies are available for multi-target controls for temperature, humidity, and VOCs. (8) Multi-target controls are not available yet for temperature, humidity, VOCs, and CO2.

42 ENGINEERING↗

A Distance Measure Comparison to Improve Crowding in Multi-Modal Problems.

Solving multi-modal optimization problems are of interest to researchers solving real world problems in areas such as control systems and power engineering tasks. Extensions of simple Genetic Algorithms, particularly types of crowding, have been developed to help solve these types of problems. This paper examines the performance of two distance measures, Mahalanobis and Euclidean, exercised in the processing of two different crowding type implementations against five minimization functions. Within the context of the experiments, empirical evidence shows that the statistical based Mahalanobis distance measure when used in Deterministic Crowding produces equivalent results to a Euclidean measure. In the case of Restricted Tournament selection, use of Mahalanobis found on average 40% more of the global optimum, maintained a 35% higher peak count and produced an average final best fitness value that is 3 times better.

97 MATHEMATICS AND COMPUTING↗

EVALUATION OF CONTROLLED END GAS AUTO IGNITION WITH EXHAUST GAS RECIRCULATION IN A STOICHIOMETRIC, SPARK IGNITED, NATURAL GAS ENGINE

Many stationary and heavy-duty on-road natural gas fueled engines today operate under stoichiometric conditions with a three-way catalyst. The disadvantage of stoichiometric natural gas engines compared to lean-burn natural gas and diesel engines is lower efficiency, resulting primarily from lower power density and compression ratio. Exhaust gas recirculation (EGR) coupled with advanced combustion controls can enable operation with higher compression ratio and power density, which yields higher efficiency. This also results in engine operation between the limits of knock and misfire. Operation between these limits has been named controlled end gas auto-ignition (C-EGAI) and can be used to improve the brake efficiency of the engine. Various methods of cylinder pressure-based knock quantification were explored to implement C-EGAI. The indicated quantification methods are used for the implementation of a control scheme for C-EGAI with a relation to the fractional heat release due to auto-ignition. A custom EGR system was built and the effect of EGR on the performance of a stochiometric, spark ignited, natural gas engine is evaluated. C-EGAI is implemented and the optimal parameters are determined for peak performance under EGR and C-EGAI conditions. In this study, knock detection is used for the recognition, magnitude, and location of the auto-ignition events. Cylinder pressure-based knock detection was the primary method for determining the occurrence and location of knock but was also used for implementing the ignition control scheme for controlled end gas auto-ignition. The combustion intensity metric (CIM) enabled parametric ignition timing control which allowed for the creation of a relationship between fractional heat release due to auto-ignition and CIM. Many stationary and heavy-duty on-road natural gas fueled engines today operate under stoichiometric conditions with a three-way catalyst. The disadvantage of stoichiometric natural gas engines compared to lean-burn natural gas and diesel engines is lower efficiency, resulting primarily from lower power density and compression ratio. Exhaust gas recirculation (EGR) coupled with advanced combustion controls can enable operation with higher compression ratio and power density, which yields higher efficiency. This also results in engine operation between the limits of knock and misfire. Operation between these limits has been named controlled end gas auto-ignition (C-EGAI) and can be used to improve the brake efficiency of the engine. Various methods of cylinder pressure-based knock quantification were explored to implement C-EGAI. The indicated quantification methods are used for the implementation of a control scheme for C-EGAI with a relation to the fractional heat release due to auto-ignition. A custom EGR system was built and the effect of EGR on the performance of a stochiometric, spark ignited, natural gas engine is evaluated. C-EGAI is implemented and the optimal parameters are determined for peak performance under EGR and C-EGAI conditions. In this study, knock detection is used for the recognition, magnitude, and location of the auto-ignition events. Cylinder pressure-based knock detection was the primary method for determining the occurrence and location of knock but was also used for implementing the ignition control scheme for controlled end gas auto-ignition. The combustion intensity metric (CIM) enabled parametric ignition timing control which allowed for the creation of a relationship between fractional heat release due to auto-ignition and CIM. Both exhaust gas recirculation and controlled end gas auto-ignition were analyzed with a cooperative fuel research (CFR) engine modified for boosted fuel/air intake. The data was interpreted to provide a proper evaluation of unique analytical methods to quantify the results of C_EGAI and characterize the live auto-ignition events. The control variables for this method of C-EGAI were optimized with EGR conditions to generate the point of peak performance on the CFR engine under stoichiometric, spark ignited, natural gas conditions.

Bayliff, Scott Michael↗

Biochemical and Structural Studies of NADH-Dependent FabG Used To Increase the Bacterial Production of Fatty Acids under Anaerobic Conditions

Major efforts in bioenergy research have focused on producing fuels that can directly replace petroleum-derived gasoline and diesel fuel through metabolic engineering of microbial fatty acid biosynthetic pathways. Typically, growth and pathway induction are conducted under aerobic conditions, but for operational efficiency in an industrial context, anaerobic culture conditions would be preferred to obviate the need to maintain specific dissolved oxygen concentrations and to maximize the proportion of reducing equivalents directed to biofuel biosynthesis rather than ATP production. A major concern with fermentative growth conditions is elevated NADH levels, which can adversely affect cell physiology. The purpose of this study was to identify homologs of Escherichia coli FabG, an essential reductase involved in fatty acid biosynthesis, that display a higher preference for NADH than for NADPH as a cofactor. Four potential NADH-dependent FabG variants were identified through bioinformatic analyses supported by crystallographic structure determination (1.3- to 2.0-Å resolution). In vitro assays of cofactor (NADH/NADPH) preference in the four variants showed up to ~35-fold preference for NADH, which was observed with the Cupriavidus taiwanensis FabG variant. In addition, FabG homologs were overexpressed in fatty acid- and methyl ketone-overproducing E. coli host strains under anaerobic conditions, and the C. taiwanensis variant led to a 60% higher free fatty acid titer and 75% higher methyl ketone titer relative to the titers of the control strains. With further engineering, this work could serve as a starting point for establishing a microbial host strain for production of fatty acid-derived biofuels (e.g., methyl ketones) under anaerobic conditions.

59 BASIC BIOLOGICAL SCIENCES↗

Recovery of Natural Gas Equipment Emissions into Gas Compression Engines for the Reduction of Potential Greenhouse Gas Emissions

Since the turn of the millennium, the United States (U.S.) oil and natural gas (ONG) industry has nearly doubled its natural gas production rate. As a result, the ONG industry has recently come under increasing scrutiny for its contributions to greenhouse gas (GHG) emissions. Consequently, various solutions to this problem have been proposed and formulated to reduce the impacts of GHG emissions on the environment. West Virginia University (WVU) have found it important to research the impacts of recovering vented gas streams into prime-mover engines. The U.S. Department of Energy (DOE) and National Energy Technology Laboratory (NETL) have granted WVU funding to research and develop a “Methane Mitigator” (M2) - a “Scalable Vent Mitigation Strategy to Simultaneously Reduce Methane Emissions and Fuel Consumption from the Compression Industry.” One of the main areas of interest for this research was the collection of emissions from natural gas equipment into a Caterpillar G3508J natural gas compression engine. The parameters being analyzed from the engine were brake-specific emissions and power output. The emissions sources considered for this research were pneumatic controllers (PCs), reciprocating compressor vents, and the engine’s open crankcase breather. The compressor vent and PC emissions were simulated using a mass flow controller (MFC) and flowed into the engine using two separate methods: (1) directly into the air intake, and (2) through a retrofitted closed crankcase ventilation system (CCV), serving as a buffer volume. The crankcase emissions were quantified without the CCV, and the impact on exhaust emissions from circulating the crankcase gases into the intake was measured. The simulated compressor vent and PC flows from the MFC had limited effect on the steady state operation of the engine and resulting performance. When the simulated flows were fed directly into the engine’s air intake, the changes within the engine’s continuous performance and emission parameters were larger but lasted for shorter durations. Conversely, when the simulated flows were fed into the CCV before entering the air intake, the changes in the engine’s performance and emission parameters were less pronounced for continuous analysis but lasted for longer durations. In either case, the continuous emission changes in both emissions and performance varied in size depending on the test scenario being run, but the cycle average changes in emissions and performance showed little impact overall compared to the engine’s baseline operation. As a result, the inclusion of a CCV shows a decrease in baseline carbon dioxide equivalent (CO2-eq.) engine emissions (from combined exhaust and open crankcase) of almost 4%. Likewise, the CCV inclusion reduced baseline total methane (CH4) from combined exhaust and open crankcase by upwards of 16%. These atmospheric emissions only decreased further with the inclusions of collected PC and compressor vent flows. The resulting changes in time-averaged rated exhaust behavior (or lack thereof) prove that the proposed M2 system could likely be deployed at sites with modern lean-burn natural gas engines as a viable option for reducing and eliminating potential GHG sources that would have otherwise been unutilized as energy sources.

03 NATURAL GAS↗

Influence of Spindle Wobble in Turning on the Workpiece’s Surface Topology

The shaping trajectory of the cutting tool relative to the workpiece consists of the trajectories of the machine tool’s working components, trajectories unrelated to control of perturbations (including spindle wobble), and deformational-displacement trajectories of the tool and workpiece relative to the supporting system. In the present work, the perturbation considered is wobble, which depends on the precision and state of the machine tool. Spindle wobble in a lathe leads to the formation of various attractive sets associated with deformational displacement of the tool. The result is to change the shaping trajectory. That affects the geometric topology of the machined surface, which depends not only on the shaping motion but also on independent processes that accompany cutting, such as plastic deformation, thermodynamic phenomena, and dissipative structures formed in the cutting zone. The present work examines how the attractive sets associated with the deformational displacement produced by spindle wobble are related to the surface topology. The contribution of the wobble to the surface topology is assessed.

42 ENGINEERING↗

Influence of Axial Power Distributions on Rod-to-wall Gap

In a scaled-down (3 x 3, 4 x 4, 5 x 5 or 6 x 6) rod bundle thermal hydraulic experiment, channel walls were often applied to surround the experimental rod bundle and provide a hydraulic flow boundary for the experiments. However, in the actual PWR (Pressurized Water Reactor), there are no walls surrounding the rod bundle fuel assembly. These non-representative walls can lead to non-prototypical conditions in the rod bundle experiments. There are several researches about the wall effects in rod bundle experiments. Based on hydraulic consideration, Zang developed a wall effect factor by subchannel analysis method to investigate the wall effects of different scale rod bundles. He did not pay attention to the wall effects in thermal aspects. Li studied the wall effects of different scale rod bundles of tight arrangement spiral-fin fuel assembly by CFD. The existing researches mostly focus on the wall effects of different scale rod bundles. There are very few, if any at all, studies focus on how to avoid or minimize wall effects through proper experimental design by adjusting the rod-to-wall gap (the gap between the channel wall and outer rod wall) for rod bundles of various scales. In addition, the axial power distribution is different in different periods of the life cycle in the PWR core. For example, it usually is outlet peak power distribution in the axial direction at the beginning of life span. However, it may turn to the inlet peak power distribution when the control rods are pulled out from the high burn up of fuel in the reactor core. Both flow dynamics and heat transfer mechanism change depending on various axial power distributions. Subsequently, it is important to study the impacts of the wall effects on flow and heat transfer with different axial power distributions in rod bundle experiments. That is, in the evaluation of rod-to-wall gap in rod bundle experiments, the influence of axial power distributions should be examined. In this paper, the rod-to-wall gap was calculated in axially uniform and non-uniform power, including inlet peak, cosine, and exit peak heat flux distributions. The rod-to-wall gap calculation results were then compared with respect to two different criterion: the average fluid temperature and the maximum wall temperature criterion, to study the influence of axial heat flux profile on radial thermal distribution.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗

Metrics required for Power System Resilient Operations and Protection

Today’s complex grid involves many interdependent systems. Various layers of hierarchical control and communication systems are coordinated, both spatially and temporally to achieve gird reliability. As new communication network based control system technologies are being deployed, the interconnected nature of these systems is becoming more complex. Deployment of smart grid concepts promises effective integration of renewable resources, especially if combined with energy storage. However, without a philosophical focus on resilience, a smart grid will potentially lead to higher magnitude and/or duration of disruptive events. The effectiveness of a resilient infrastructure depends upon its ability to anticipate, absorb, adapt to, and/or rapidly recover from a potentially catastrophic event. Future system operations can be enhanced with a resilient philosophy through architecting the complexity with state awareness metrics that recognize changing system conditions and provide for an agile and adaptive response. The starting point for metrics lies in first understanding the attributes of performance that will be qualified. In this paper, we will overview those attributes and describe how they will be characterized by designing a distributed agent that can be applied to the power grid.

24 POWER TRANSMISSION AND DISTRIBUTION↗

Layer-Parallel Training of Deep Residual Neural Networks

Residual neural networks (ResNets) are a promising class of deep neural networks that have shown excellent performance for a number of learning tasks, e.g., image classification and recognition. Mathematically, ResNet architectures can be interpreted as forward Euler discretizations of a nonlinear initial value problem whose time-dependent control variables represent the weights of the neural network. Hence, training a ResNet can be cast as an optimal control problem of the associated dynamical system. For similar time-dependent optimal control problems arising in engineering applications, parallel-in-time methods have shown notable improvements in scalability. This paper demonstrates the use of those techniques for efficient and effective training of ResNets. The proposed algorithms replace the classical (sequential) forward and backward propagation through the network layers with a parallel nonlinear multigrid iteration applied to the layer domain. This adds a new dimension of parallelism across layers that is attractive when training very deep networks. From this basic idea, we derive multiple layer-parallel methods. The most efficient version employs a simultaneous optimization approach where updates to the network parameters are based on inexact gradient information in order to speed up the training process. Finally, using numerical examples from supervised classification, we demonstrate that the new approach achieves a training performance similar to that of traditional methods, but enables layer-parallelism and thus provides speedup over layer-serial methods through greater concurrency.

97 MATHEMATICS AND COMPUTING↗