Bulletin of the American Physical Society
APS March Meeting 2022
Volume 67, Number 3
Monday–Friday, March 14–18, 2022; Chicago
Session S49: Extreme-Scale Computational Science Discovery in Fluid Dynamics and Related Disciplines IIFocus Recordings Available
|
Hide Abstracts |
Sponsoring Units: DCOMP DFD Chair: Daniel Livescu, LANL Room: McCormick Place W-471B |
Thursday, March 17, 2022 8:00AM - 8:36AM |
S49.00001: Parallel Physics-Informed Neural Networks Invited Speaker: George E Karniadakis We develop a distributed framework for the physics-informed neural networks (PINNs) based on two recent extensions, namely conservative PINNs (cPINNs) and extended PINNs (XPINNs), which employ domain decomposition in space and in time-space, respectively. This domain decomposition endows cPINNs and XPINNs with several advantages over the vanilla PINNs, such as parallelization capacity, large representation capacity, efficient hyperparameter tuning, and is particularly effective for multi-scale and multi-physics problems. Here, we present a parallel algorithm for cPINNs and XPINNs constructed with a hybrid programming model described by MPI + X, where X ∈{CPUs, GPUs}. The main advantage of cPINN and XPINN over the more classical data and model parallel approaches is the flexibility of optimizing all hyperparameters of each neural network separately in each subdomain. We compare the performance of distributed cPINNs and XPINNs for various forward problems, using both weak and strong scalings. Our results indicate that for space domain decomposition, cPINNs are more efficient in terms of communication cost but XPINNs provide greater flexibility as they can also handle time-domain decomposition for any differential equations, and can deal with any arbitrarily shaped complex subdomains. To this end, we also present an application of the parallel XPINN method for solving an inverse diffusion problem with variable conductivity on the United States map, using ten regions as subdomains. |
Thursday, March 17, 2022 8:36AM - 8:48AM |
S49.00002: Machine Learning to Discover Interpretable Models in Fluids and Plasmas Alan Kaptanoglu, Jared Callaham, Christopher J Hansen, Steven L Brunton Many tasks in fluid and plasma physics, such as design optimization and control, are challenging because fluids and plasmas are nonlinear and exhibit a large range of scales in both space and time. This range of scales necessitates exceedingly high-dimensional measurements and computational discretization to resolve all relevant features, resulting in vast data sets and time-intensive computations. Indeed, fluid dynamics is one of the original big data fields, and many high-performance computing architectures, experimental measurement techniques, and advanced data processing and visualization algorithms were driven by decades of research in fluid mechanics. Machine learning constitutes a growing set of powerful techniques to extract patterns and build models from fluid and plasma data, complementing existing theoretical, numerical, and experimental efforts. The sparse identification of nonlinear dynamics (SINDy) algorithm is one such method that identifies a minimal dynamical system model while balancing model complexity with accuracy, avoiding overfitting. This approach tends to promote models that are interpretable and generalizable, capturing the essential physics of the system. We discuss recent advances with the SINDy method, including the incorporation of physical constraints from global conservation laws, promoting global stability, solving for weak-formulation differential equations, and more. These advances have been consolidated into the open-source PySINDy code, enabling anyone with access to measurement data to engage in scientific model discovery. |
Thursday, March 17, 2022 8:48AM - 9:00AM |
S49.00003: Adjoint-based optimization of large-scale reacting turbulent flow simulations Ali Kord, JESSE S CAPECELATRO Despite recent progress in extreme-scale computing of fluid dynamic systems, its utility in design optimization remains challenging. The 'curse of dimensionality' generally precludes the use of high-fidelity simulations within a brute-force trial-and-error approach when seeking optimal design parameters. The adjoint of the perturbed and linearized flow equations presents a promising means of leveraging high-fidelity simulations for optimization. In this talk, we present a discrete adjoint-based method for measuring local sensitivity in direct numerical simulations (DNS) of turbulent reacting flows. We introduce a novel adaptive dissipation scheme compatible with a discrete adjoint formulation that preserves scalar boundedness while retaining high-order accuracy. A flamelet-progress variable approach is employed to handle chemical reactions using tabulated chemistry. This allows for discrete-exact sensitivity to be computed efficiently for arbitrary chemical mechanisms. Putting this together, DNS of a three-dimensional reacting round jet are performed at moderate Reynolds numbers. Sensitivity obtained from the adjoint solution is used to study flame and chemistry responses to turbulent flow actuation. |
Thursday, March 17, 2022 9:00AM - 9:12AM |
S49.00004: Machine-Learning-Enabled Prediction of Spatiotemporal Boundary Conditions in Multiphase Flow Simulations Gina M Magnotti, Sudeepta Mondal, Roberto Torelli, Bethany A Lusch Simulating fuel injection and spray breakup remains a computational challenge due to the multi-physics and multi-scale nature of the problem. Static coupling approaches have been adopted whereby simulation of the internal flow is replaced with a spatiotemporal boundary condition that defines the injection profile at the injector orifice exit. However, the generation of these data is a computationally expensive task due to fine temporal and spatial resolution requirements. This presentation summarizes our work in developing a machine-learning-based emulator to learn efficient surrogate models for spatiotemporal boundary conditions. An interpretable Bayesian learning strategy is employed to understand the effect of design parameters on the learned spatiotemporal fields. Autoencoders are utilized for efficient dimensionality reduction of the flowfields. Gaussian process (GP) models are then used to predict the spatiotemporal flowfields at the injector exit for test design conditions not seen during training. The emulation framework can predict the spatiotemporal boundary conditions within a few seconds, thus achieving a speed-up factor of up to 38 million over the traditional simulation-based approach. |
Thursday, March 17, 2022 9:12AM - 9:24AM |
S49.00005: Predicting Boundary-Layer Transition (BLT) using Artificial Intelligence (AI) Causality Inference Arturo Rodriguez, Andres Enriquez, Jose Terrazas, Daniel Villanueva, Brandon Paez, Nicholas Dudu, Rafael Baez, Christopher Harris, Vinod Kumar The transition of the boundary layer has been studied through stability analysis and Direct Numerical Simulations; it has also been characterized by methods such as N-Factor and different modes of instabilities. The methods cannot be generalized to all kinds of cases where the failure grows in intermittency and uncertainty. In intermittency, a vital role is played in releasing accumulated energy in fluid flows in unpredictable times and the uncertain generation of chaos. To mitigate the failure of conventional and deterministic methods where chaos always exists based on Lorenz's studies. I propose to enter probabilistic methods where the conditional probability with Do-Calculus probability can be exchanged based on graph operations where the solution based on stochastic processes can be found. The graph can show cause-and-effect relationships based on events that explain why the transition from the boundary layer to higher excitations leads to turbulence. These studies will show previous boundary layer studies, assumptions, the different forms of instabilities, causes, and leading effects to turbulence. In addition, I will provide a better understanding of the current metrics of how to characterize turbulence, the benefits of these studies, where it guides us, and how to generate graphs based on causality inference, which can show us why the flow becomes turbulent based on various causes where we can answer the ultimate question of why. |
Thursday, March 17, 2022 9:24AM - 9:36AM |
S49.00006: Physics-informed Machine Learning for Reduced-order Modeling of Lagrangian Turbulence Yifeng Tian, Michael Woodward, Mikhail Stepanov, Chris Fryer, Criston M Hyett, Michael Chertkov, Daniel Livescu Fully resolving turbulent flows in physical sciences and engineering applications using Direct Numerical Simulation (DNS) is generally prohibitively expensive due to the wide range of scales and their non-linear interactions. This challenge has motivated the development of efficient reduced-order models of turbulence dynamics, which has seen a remarkable new boost from disparate fields of Machine Learning. The talk represents an overview of our efforts in developing reduced-order models describing various aspects of Lagrangian turbulent dynamics under the Physics-Informed Machine Learning (PIML) paradigm. We inject physical constraints into the construction of Neural Network models for turbulence dynamics at both coarse-grained scale (based on Smoothed Particle Hydrodynamics) and Kolmogorov scale (based on Velocity Gradient Tensor dynamics). A large Lagrangian dataset is extracted from high-Reynolds number DNS and used to train the PIML models. Through a series of diagnostic tests, we show that the trained PIML models are capable of producing the correct flow structures and turbulence statistics in homogeneous isotropic turbulence. In addition, the Lagrangian framework shows good promise for extrapolating the PIML models outside the training data, e.g. to higher Reynolds numbers. |
Thursday, March 17, 2022 9:36AM - 9:48AM |
S49.00007: Sparse identification of multiphase turbulence closures for strongly-coupled gas-particle flows Sarah Beetham, Rodney O Fox, JESSE S CAPECELATRO In this talk, we will present a data-driven framework for model closure of the multiphase Reynolds Average Navier—Stokes (RANS) equations. To date, the majority of RANS closures are based on extensions of single-phase turbulence models, which fail to capture complex two-phase flow dynamics across dilute and dense regimes, especially when two-way coupling between the phases is important. This eliminates the augmentation of existing models as an option for solving the multiphase closure problem. We will focus on gas-solid flows at moderate volume fractions and Reynolds numbers, such that strong coupling between the phases gives rise to large-scale heterogeneity (clusters) that drive the underlying turbulence. Data generated from highly resolved simulations are used in a sparse regression method for model closure that ensures form invariance. We will demonstrate how the sparse regression methodology identifies compact, algebraic models from large-scale simulation data. |
Thursday, March 17, 2022 9:48AM - 10:00AM |
S49.00008: Computational Investigation of the Effects of Chemistry on Mars Retropropulsion Environments eric j nielsen, gabriel c nastac, aaron c walden, ashley m korzun, chris stone, patrick j moran The effects of real-gas chemistry on a human-scale Mars lander concept are evaluated using scale-resolving computational fluid dynamics. Ground testing of such vehicle concepts requires significant compromises on physical scale, instrumentation, configuration, and environments. The effects of real-gas chemistry are often neglected due to constraints on test articles and facilities; most experiments use inert simulant gases at low temperatures. Instead, a strong reliance on high-fidelity computational analyses is required to expand the understanding of retropropulsion aerodynamics. In this work, simulations are performed on thousands of Graphics Processing Units using the Summit system available at the Oak Ridge Leadership Computing Facility, resulting in game-changing computational performance. An overview of the computational approach is presented and results are compared with those obtained in a previous perfect gas campaign. |
Thursday, March 17, 2022 10:00AM - 10:12AM |
S49.00009: A Discrete Ion Stochastic Continuum Overdamped Solvent Algorithm for Modeling Electrolytes Daniel R Ladiges, J. Galen Wang, Ishan Srivastava, Sean P Carney, Andrew J Nonaka, Katherine Klymko, Guy C Moore, Alejandro L Garcia, Sachin R Natesh, Aleksandar Donev, John B Bell In this talk we present a methodology for the mesoscale simulation of fluid/particle systems such as strong electrolytes. This is an extension of the Fluctuating Immersed Boundary (FIB) approach that treats a solute as discrete Lagrangian particles that interact with Eulerian hydrodynamic and electrostatic fields. In both cases the Immersed Boundary (IB) method of Peskin is used for particle-field coupling. Hydrodynamic interactions are taken to be overdamped, with thermal noise incorporated using the fluctuating Stokes equation, including a "dry diffusion" Brownian motion to account for scales not resolved by the coarse-grained model of the solvent. Long range electrostatic interactions are computed by solving the Poisson equation, with short range corrections included using a novel immersed-boundary variant of the classical Particle-Particle Particle-Mesh (P3M) technique. This approach is designed to enable scaling to large problems which are difficult to tackle using many existing mesoscale methods. It has been implemented using the AMReX framework for use on large scale HPC systems, including heterogeneous CPU+GPU architectures. |
Thursday, March 17, 2022 10:12AM - 10:24AM Withdrawn |
S49.00010: Coupling melt-pool fluid dynamics to microstructure development to process-aware material model in the Exascale Additive Manufacturing project James F Belak, John Coleman, Robert Carson, Matthew Rolchigo, Samuel T Reeve The Exascale Additive Manufacturing (AM) project (ExaAM) is focused on metal AM process modeling at the fidelity of the microstructure. To date, ExaAM has instantiated exascale-ready continuum simulations of the AM build and the re-solidification process from fluid dynamics simulations of the laser scan. The calibrated continuum fluid dynamics simulations are used to drive microstructure development simulations from which properties are calculated. Using the AMBench-01 problem as a test case, we generate an ensemble of microstructures and properties from which a continuum material model is determined. While potentially unique to the AMBench-01 build, this process-aware material model enables build simulations closer to AM conditions and part-scale qualification simulations in as-built material conditions. Comparison will be made to experiments of microstructure, properties and build. |
Thursday, March 17, 2022 10:24AM - 10:36AM |
S49.00011: A Quantum Inspired Approach to Exploit Turbulence Structures Peyman Givi, Nikita Gourianov, Martin Kiffner, Michael Lubasch, Hessam Babaee, Sergey Dolgov, Quincy van den Berg, Dieter Jaksch We introduce a new paradigm for analyzing the structure of turbulent flows by quantifying correlations between different length scales using methods inspired from quantum many-body physics. We present results for interscale correlations of two paradigmatic flow examples, and use these insights along with tensor network theory to design a structure-resolving algorithm for simulating turbulent flows. With this algorithm, we find that the incompressible Navier-Stokes equations can be accurately solved even when reducing the number of parameters required to represent the velocity field by more than one order of magnitude compared to direct numerical simulation. Our quantum-inspired approach provides a pathway towards conducting computational fluid dynamics on quantum computers. |
Thursday, March 17, 2022 10:36AM - 10:48AM |
S49.00012: High-Fidelity Numerical Simulations of Fire Using Adaptive Mesh Refinement Peter Hamlington, Caelan B Lapointe, Michael Meehan, Sam Simons-Wellin, Nicholas T Wimer, Jeffrey F Glusman High-fidelity computational predictions of fire in natural and built environments are constrained by the difficulty of modeling complex physics across wide temporal and spatial scale ranges. Fire structure and dynamics are governed by coupled nonlinear interactions between turbulence, buoyancy-driven flow, and flame chemistry, and these multi-physics phenomena typically span an enormous range of spatial and temporal scales. In this talk, we outline recent efforts to use adaptive mesh refinement (AMR), where the grid is resolved at small scales only in regions of high dynamical and physical significance, for the study of fire structure, dynamics, and evolution in a range of contexts. We will focus on two areas in particular: (i) the structure and dynamics of non-reacting buoyancy driven flows relevant to pool fires and (ii) fire spread in natural and built environments. Physical understanding resulting from these studies will allow refinement of prevailing theories as well as provide guidance on improvements to subgrid-scale models for large eddy simulations of practical fire problems. Challenges faced in the application of AMR to fire simulations are outlined, and future research directions, with a focus on outer-loop processes (e.g., optimization), are also highlighted. |
Thursday, March 17, 2022 10:48AM - 11:00AM |
S49.00013: Stability of a phase boundary with heat and mass fluxes Snezhana I Abarzhi, D.V. Ilyin This work focuses on the long-standing problem of stability of a phase boundary with heat and mass fluxes across it. Our theory resolves challenges not addressed before, including boundary conditions for thermal heat flux, structure of perturbation waves, and dependence of waves coupling on the system parameters. We find the interface stability in a broad range of parameters; discover new fluid instabilities in the advection, diffusion and low Mach regimes; and discuss the theory outcomes for experiments and simulations and for multiphase flows in nature and technology. |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700