-
GDP: Generalized Device Placement for Dataflow Graphs
Authors:
Yanqi Zhou,
Sudip Roy,
Amirali Abdolrashidi,
Daniel Wong,
Peter C. Ma,
Qiumin Xu,
Ming Zhong,
Hanxiao Liu,
Anna Goldie,
Azalia Mirhoseini,
James Laudon
Abstract:
Runtime and scalability of large neural networks can be significantly affected by the placement of operations in their dataflow graphs on suitable devices. With increasingly complex neural network architectures and heterogeneous device characteristics, finding a reasonable placement is extremely challenging even for domain experts. Most existing automated device placement approaches are impractica…
▽ More
Runtime and scalability of large neural networks can be significantly affected by the placement of operations in their dataflow graphs on suitable devices. With increasingly complex neural network architectures and heterogeneous device characteristics, finding a reasonable placement is extremely challenging even for domain experts. Most existing automated device placement approaches are impractical due to the significant amount of compute required and their inability to generalize to new, previously held-out graphs. To address both limitations, we propose an efficient end-to-end method based on a scalable sequential attention mechanism over a graph neural network that is transferable to new graphs. On a diverse set of representative deep learning models, including Inception-v3, AmoebaNet, Transformer-XL, and WaveNet, our method on average achieves 16% improvement over human experts and 9.2% improvement over the prior art with 15 times faster convergence. To further reduce the computation cost, we pre-train the policy network on a set of dataflow graphs and use a superposition network to fine-tune it on each individual graph, achieving state-of-the-art performance on large hold-out graphs with over 50k nodes, such as an 8-layer GNMT.
△ Less
Submitted 28 September, 2019;
originally announced October 2019.
-
Transcritical Mixing and Auto-Ignition of n-dodecane Liquid Fuel using a Diffused Interface Method
Authors:
Luis Bravo,
Peter C. Ma,
Matthias Ihme
Abstract:
High-fidelity simulations of transcritical spray mixing and auto ignition in a combustion chamber are performed at high pressure and temperature conditions using a recently developed finite rate LES solver. The simulation framework is based on a diffused-interface (DI) method that solves the compressible multi-species conservation equations along with the Peng Robinson state equation and real-flui…
▽ More
High-fidelity simulations of transcritical spray mixing and auto ignition in a combustion chamber are performed at high pressure and temperature conditions using a recently developed finite rate LES solver. The simulation framework is based on a diffused-interface (DI) method that solves the compressible multi-species conservation equations along with the Peng Robinson state equation and real-fluid transport properties. A finite volume approach with entropy stable scheme is employed to accurate simulate the non-linear real fluid flow. LES analysis is performed for non-reacting and reacting spray conditions targeting the ECN Spray A configuration at chamber conditions with a pressure of 60 bar and temperatures between 800 K and 1200 K to investigate effects of the real-fluid environment and low-temperature chemistry. Comparisons with measurements in terms of global spray parameters and mixture fraction distributions demonstrates the accuracy in modeling the turbulent mixing behavior. Good overall agreement of the auto-ignition process is obtained from simulation results at different ambient temperature conditions and the formation of intermediate species is captured by the simulations, indicating that the presented numerical framework adequately reproduces the corresponding low-and-high-temperature ignition processes under high-pressure conditions that are relevant to realistic diesel fuel injection systems.
△ Less
Submitted 23 September, 2018;
originally announced September 2018.
-
Computational Modeling and Analysis of Diesel-fuel Injection and Autoignition at Transcritical Conditions
Authors:
Matthias Ihme,
Peter C. Ma,
Luis Bravo
Abstract:
The need for improved engine efficiencies has motivated the development of high-pressure combustion systems, in which operating conditions achieve and exceed critical conditions. Associated with these conditions are strong variations in thermo-transport properties as the fluid undergoes phase transition, and two-stage ignition with low-temperature combustion. Accurately simulating these physical p…
▽ More
The need for improved engine efficiencies has motivated the development of high-pressure combustion systems, in which operating conditions achieve and exceed critical conditions. Associated with these conditions are strong variations in thermo-transport properties as the fluid undergoes phase transition, and two-stage ignition with low-temperature combustion. Accurately simulating these physical phenomena at real-fluid environments remains a challenge. By addressing this issue, a high-fidelity LES-modeling framework is developed to conduct simulations of transcritical fuel spray mixing and auto-ignition at high-pressure conditions. The simulation is based on a recently developed diffused interface method that solves the compressible multi-species conservation equations along with a Peng-Robinson state equation and real-fluid transport properties. LES analysis is performed for non-reacting and reacting spray conditions targeting the ECN Spray A configuration at chamber conditions with a pressure of 60 bar and temperatures between 900 K and 1200 K to investigate effects of the real-fluid environment and low-temperature chemistry. Comparisons with measurements in terms of global spray parameters (i.e., liquid and vapor penetration lengths) are shown to be in good agreement. Analysis of the mixture fraction distributions in the dispersed spray region demonstrates the accuracy in modelling the turbulent mixing behavior. Good agreement of the ignition delay time and the lift-off length is obtained from simulation results at different ambient temperature conditions and the formation of intermediate species is captured by the simulations, indicating that the presented numerical framework adequately reproduces the corresponding low- and high-temperature ignition processes under high-pressure conditions, which are relevant to realistic diesel-fuel injection systems.
△ Less
Submitted 23 September, 2018;
originally announced September 2018.
-
Lyapunov exponent and Wasserstein metric as validation tools for assessing short-time dynamics and quantitative model evaluation of large-eddy simulation
Authors:
Hao Wu,
Peter C. Ma,
Yu Lv,
Matthias Ihme
Abstract:
In this work, methods for the evaluation of LES-quality and LES-accuracy are presented, which include the Lyapunov exponent for the analysis of short-time predictability of LES-calculation and the Wasserstein metric for the quantitative assessment of simulation results. Both methods are derived and evaluated in application to the Volvo test case. Both the non-reacting and reacting cases are calcul…
▽ More
In this work, methods for the evaluation of LES-quality and LES-accuracy are presented, which include the Lyapunov exponent for the analysis of short-time predictability of LES-calculation and the Wasserstein metric for the quantitative assessment of simulation results. Both methods are derived and evaluated in application to the Volvo test case. Both the non-reacting and reacting cases are calculated. For the non- reacting cases, good agreement with the experimental data is achieved by solvers at high numerical resolution. The reacting cases are more challenging due to the small length scale of the flame and the suppression of sinuous mode of absolute instability by the density ratio. The analysis of the turbulent simulation data using the concept of the Lyapunov exponent and the Wasserstein metric provides a more quantitative approach to assess the mesh dependency of the simulation results. The convergence of the Lyapunov exponent is shown to be a more sensitive and stronger indication of mesh-independence. Though grid convergence for the reacting cases cannot be reached with the chosen resolutions, the Lyapunov exponents and the Wasserstein metric are shown to be capable of identifying quantity-specific sensitivities with respect to the numerical resolution, while requiring significantly less computational resources than acquiring profiles of conventional turbulent statistics.
△ Less
Submitted 9 January, 2018;
originally announced January 2018.
-
Numerical analysis on mixing processes for transcritical real-fluid simulations
Authors:
Peter C. Ma,
Hao Wu,
Daniel T. Banuti,
Matthias Ihme
Abstract:
The accurate and robust simulation of transcritical real-fluid flows is crucial for many engineering applications. Diffused interface methods are frequently employed and several numerical schemes have been developed for simulating transcritical flows. These schemes can be categorized into two types, namely fully conservative and quasi-conservative schemes. An adaptive scheme which is a hybrid of t…
▽ More
The accurate and robust simulation of transcritical real-fluid flows is crucial for many engineering applications. Diffused interface methods are frequently employed and several numerical schemes have been developed for simulating transcritical flows. These schemes can be categorized into two types, namely fully conservative and quasi-conservative schemes. An adaptive scheme which is a hybrid of the two is developed in this study. By considering several numerical test cases, it is shown that different schemes predict distinctly different mixing behaviors. It is shown that the mixing processes follow the isobaric-adiabatic and isobaric-isochoric mixing models for fully and quasi-conservative schemes, respectively, and the adaptive scheme yields a mixing behavior that spans both models. The distinct mixing behaviors are a consequence of numerical diffusion instead of physical diffusion and can be attributed to insufficient numerical spatial resolution. This work provides a better understanding on the interpretation of numerical simulation results and the mixing models that are commonly used to study transcritical flows.
△ Less
Submitted 6 December, 2017;
originally announced December 2017.
-
Direct numerical simulations of turbulent channel flow under transcritical conditions
Authors:
Peter C. Ma,
Xiang I. A. Yang,
Matthias Ihme
Abstract:
Turbulent flows under transcritical conditions are present in regenerative cooling systems of rocker engines and extraction processes in chemical engineering. The turbulent flows and the corresponding heat transfer phenomena in these complex processes are still not well understood experimentally and numerically. The objective of this work is to investigate the turbulent flows under transcritical c…
▽ More
Turbulent flows under transcritical conditions are present in regenerative cooling systems of rocker engines and extraction processes in chemical engineering. The turbulent flows and the corresponding heat transfer phenomena in these complex processes are still not well understood experimentally and numerically. The objective of this work is to investigate the turbulent flows under transcritical conditions using DNS of turbulent channel flows. A fully compressible solver is used in conjunction with a Peng-Robinson real-fluid equation of state to describe the transcritical flows. A channel flow with two isothermal walls is simulated with one heated and one cooled boundary layers. The grid resolution adopted in this study is slightly finer than that required for DNS of incompressible channel flows. The simulations are conducted using both fully (FC) and quasi-conservative (QC) schemes to assess their performance for transcritical wall-bounded flows. The instantaneous flows and the statistics are analyzed and compared with the canonical theories. It is found that results from both FC and QC schemes qualitatively agree well with noticeable difference near the top heated wall, where spurious oscillations in velocity can be observed. Using the DNS data, we then examine the usefulness of Townsend attached eddy hypothesis in the context of flows at transcritical conditions. It is shown that the streamwise energy spectrum exhibits the inverse wavenumber scaling and that the streamwise velocity structure function follows a logarithmic scaling, thus providing support to the attached eddy model at transcritical conditions.
△ Less
Submitted 6 December, 2017;
originally announced December 2017.
-
Application of Pareto-efficient combustion modeling framework to large eddy simulations of turbulent reacting flows
Authors:
Hao Wu,
Peter C. Ma,
Thomas Jaravel,
Matthias Ihme
Abstract:
In the application of the combustion models based on low-dimensional manifolds (for instance flamelet models) to large-eddy simulation (LES) of reactive turbulent flows, the modeling simplifications of the combustion process is a critical source of uncertainty in addition to those due to the turbulent closure model and numerical discretization. The ability to quantitatively assess this uncertainty…
▽ More
In the application of the combustion models based on low-dimensional manifolds (for instance flamelet models) to large-eddy simulation (LES) of reactive turbulent flows, the modeling simplifications of the combustion process is a critical source of uncertainty in addition to those due to the turbulent closure model and numerical discretization. The ability to quantitatively assess this uncertainty in absence of the reference result is vital to the reliable and predictive simulations of practical combustion devices.In the present study, the Pareto-efficient combustion (PEC) framework is extended to adaptive LES combustion simulations of turbulent flames. The key component of the PEC framework is the so-called manifold drift term. Its extension LES is proposed to make such assessment by examining the compliance of a particular combustion model in describing a quantity of interest with respect to the underlying flow field representation. With the focus on improving predictions of CO emissions of flamelet-based combustion models, this work employs PEC to augment the flamelet/progress variable (FPV) model through local sub-model assignment of the finite-rate chemistry (FRC) model. To this end, a series of LES-PEC calculations are performed on a piloted partially-premixed dimethyl ether flame (DME-D), using a combination of FPV and FRC models. The drift term is utilized in the PEC framework to the estimate the model related error for quantities of interest. The PEC approach is demonstrated to be capable of significantly improving the prediction of CO emissions compared with the monolithic FPV simulation. The improved accuracy is achieved by enriching the FPV model with FRC in regions where the lower-order model is determined insufficient through the evaluation of drift terms.
△ Less
Submitted 4 December, 2017;
originally announced December 2017.
-
Efficient time step** for reactive turbulent simulations with stiff chemistry
Authors:
Hao Wu,
Peter C. Ma,
Matthias Ihme
Abstract:
A combination of a steady-state preserving operator splitting method and a semi-implicit integration scheme is proposed for efficient time step** in simulations of unsteady reacting flows, such as turbulent flames, using detailed chemical kinetic mechanisms. The operator splitting is based on the Simpler balanced splitting method, which is constructed with improved stability properties and reduc…
▽ More
A combination of a steady-state preserving operator splitting method and a semi-implicit integration scheme is proposed for efficient time step** in simulations of unsteady reacting flows, such as turbulent flames, using detailed chemical kinetic mechanisms. The operator splitting is based on the Simpler balanced splitting method, which is constructed with improved stability properties and reduced computational cost. The method is shown to be capable of stable and accurate prediction of ignition and extinction for reaction-diffusion systems near critical conditions. The ROK4E scheme is designed for semi-implicit integration of spatially independent chemically reacting systems. Being a Rosenbrock-Krylov method, ROK4E utilizes the low-rank approximation of the Jacobian to reduce the cost for integrating the system of ODEs that have relative few stiff components. The efficiency of the scheme is further improved via the careful choice of coefficients to require three right-hand-side evaluations over four stages. Combing these two methods, efficient calculation is achieved for large-scale parallel simulations of turbulent flames.
△ Less
Submitted 4 December, 2017;
originally announced December 2017.
-
MVP-Workshop Contribution: Modeling of Volvo Bluff Flame Experiment and Comparison of Finite-Volume and Discontinuous-Galerkin Schemes
Authors:
Hao Wu,
Peter C. Ma,
Yu Lv,
Matthias Ihme
Abstract:
The Volvo burner features the canonical configuration of a bluff-body stabilized premixed flame. This configuration was studied experimentally under the Volvo Flygmotor AB program. Two cases are considered in this study: a non-reacting case with an inlet flow speed of 16.6 m/s and a reacting case with equilibrium ratio of 0.65 and inflow speed of 17.3 m/s. The characteristic vortex shedding in the…
▽ More
The Volvo burner features the canonical configuration of a bluff-body stabilized premixed flame. This configuration was studied experimentally under the Volvo Flygmotor AB program. Two cases are considered in this study: a non-reacting case with an inlet flow speed of 16.6 m/s and a reacting case with equilibrium ratio of 0.65 and inflow speed of 17.3 m/s. The characteristic vortex shedding in the wake behind the bluff body is present in the non-reacting case, while two oscillation modes are intermittently present in the reacting case. A series of large-eddy simulations are performed on this configuration using two solvers, one using a high-resolution finite-volume (FV) scheme and the other featuring a high-order discontinuous-Galerkin (DG) discretization. The FV calculations are conducted on hexahedral meshes with three different resolution (4mm, 2mm, and 1mm). The DG calculations are performed using two different polynomial orders on the same tetrahedral mesh. For the non-reacting cases, good agreement with respect to the experimental data is achieved by both solvers at high numerical resolution. The reacting cases are calculated using a two-step global mechanism in combination with the thickened-flame model. Reasonable agreement with experiments is obtained by both solvers at higher resolution. Models for combustion-turbulence interaction are necessary for the reacting case as it contains the length scale of the flame, which is smaller than the grid resolution in all calculations. The impact of such models on the flame stability and flow/flame dynamics is the subject of future research.
△ Less
Submitted 29 August, 2017; v1 submitted 18 July, 2017;
originally announced July 2017.
-
Modeling and Simulation of Diesel Injection at Transcritical Conditions
Authors:
Peter C. Ma,
Matthias Ihme,
Luis Bravo
Abstract:
The need for improved engine efficiencies has motivated the development of high-pressure combustion systems, in which operating conditions achieve and exceed critical conditions. Associated with these conditions are large thermodynamic gradients and strong variations in transport properties as the fluid undergoes mixing and phase transition. Accurately simulating these real-fluid environments rema…
▽ More
The need for improved engine efficiencies has motivated the development of high-pressure combustion systems, in which operating conditions achieve and exceed critical conditions. Associated with these conditions are large thermodynamic gradients and strong variations in transport properties as the fluid undergoes mixing and phase transition. Accurately simulating these real-fluid environments remains a main challenge. Different modeling approaches have been employed, which can be categorized as diffused and sharp interface methods. The objective of this study is to examine the diffused interface method for simulating diesel-fuel injection at conditions related to the supercritical regime. To this end, a recently developed compressible real-fluid solver for transcritical conditions is employed. Simulations of an ECN-relevant diesel-fuel injector are performed and predictions for instantaneous and statistical flow-field results are compared against available measurements. It is expected that results from this analysis will be useful in identifying limitations of current modeling techniques and in improving physical and numerical models for high-pressure injection systems.
△ Less
Submitted 19 May, 2017;
originally announced May 2017.
-
Numerical framework for transcritical real-fluid reacting flow simulations using the flamelet progress variable approach
Authors:
Peter C. Ma,
Daniel T. Banuti,
Jean-Pierre Hickey,
Matthias Ihme
Abstract:
An extension to the classical FPV model is developed for transcritical real-fluid combustion simulations in the context of finite volume, fully compressible, explicit solvers. A double-flux model is developed for transcritical flows to eliminate the spurious pressure oscillations. A hybrid scheme with entropy-stable flux correction is formulated to robustly represent large density ratios. The ther…
▽ More
An extension to the classical FPV model is developed for transcritical real-fluid combustion simulations in the context of finite volume, fully compressible, explicit solvers. A double-flux model is developed for transcritical flows to eliminate the spurious pressure oscillations. A hybrid scheme with entropy-stable flux correction is formulated to robustly represent large density ratios. The thermodynamics for ideal-gas values is modeled by a linearized specific heat ratio model. Parameters needed for the cubic EoS are pre-tabulated for the evaluation of departure functions and a quadratic expression is used to recover the attraction parameter. The novelty of the proposed approach lies in the ability to account for pressure and temperature variations from the baseline table. Cryogenic LOX/GH2 mixing and reacting cases are performed to demonstrate the capability of the proposed approach in multidimensional simulations. The proposed combustion model and numerical schemes are directly applicable for LES simulations of real applications under transcritical conditions.
△ Less
Submitted 9 April, 2017;
originally announced April 2017.
-
Numerical methods to prevent pressure oscillations in transcritical flows
Authors:
Peter C. Ma,
Yu Lv,
Matthias Ihme
Abstract:
The accurate and robust simulation of transcritical real-fluid effects is crucial for many engineering applications, such as fuel injection in internal combustion engines, rocket engines and gas turbines. For example, in diesel engines, the liquid fuel is injected into the ambient gas at a pressure that exceeds its critical value, and the fuel jet will be heated to a supercritical temperature befo…
▽ More
The accurate and robust simulation of transcritical real-fluid effects is crucial for many engineering applications, such as fuel injection in internal combustion engines, rocket engines and gas turbines. For example, in diesel engines, the liquid fuel is injected into the ambient gas at a pressure that exceeds its critical value, and the fuel jet will be heated to a supercritical temperature before combustion takes place. This process is often referred to as transcritical injection. The largest thermodynamic gradient in the transcritical regime occurs as the fluid undergoes a liquid-like to a gas-like transition when crossing the pseudo-boiling line (Yang 2000, Oschwald et al. 2006, Banuti 2015). The complex processes during transcritical injection are still not well understood. Therefore, to provide insights into high-pressure combustion systems, accurate and robust numerical simulation tools are required for the characterization of supercritical and transcritical flows.
△ Less
Submitted 29 May, 2017; v1 submitted 9 April, 2017;
originally announced April 2017.
