Bulletin of the American Physical Society
71st Annual Meeting of the APS Division of Fluid Dynamics
Volume 63, Number 13
Sunday–Tuesday, November 18–20, 2018; Atlanta, Georgia
Session G32: Machine Learning and Data Driven Models II |
Hide Abstracts |
Chair: Alireza Yazdani, Brown University Room: Georgia World Congress Center B404 |
Monday, November 19, 2018 10:35AM - 10:48AM |
G32.00001: Will it flood? Classifying entrainment outcomes via machine learning Lachlan R Mason, Indranil Pan, Aditya Karnik, Assen Batchvarov, Richard V Craster, Omar K Matar Multiphase flow simulations are advancing to the extent at which high-fidelity results can guide engineering decision-making. Accurate simulations, however, carry a high computational cost, often due to resolution constraints and the inclusion of complex, physics-driven models: question-to-answer iteration times are routinely on the order of weeks. In this study, we accelerate an engineering analysis of a benchmark flow: that of a falling film reactor. We investigate how an injected gas stream drives droplet entrainment, with the intent of predicting the certainty at which this harmful process occurs. We first train a machine learning (ML) classifier via a low-fidelity, though sufficiently representative, volume-of-fluid solver, thus mapping the class boundary demarcating flooding. Additional high-fidelity simulations along the class boundary are then used to improve the ML classifier. We quantify the savings in computational time versus prediction accuracy using this two-step ML-augmented approach, as opposed to a conventional full parameter sweep with the high-fidelity model. |
Monday, November 19, 2018 10:48AM - 11:01AM |
G32.00002: Prediction of Effective Thermal Conductivity for Lithium-Ion Battery Electrodes Using Machine Learning Techniques Fazlolah Mohaghegh, Jayathi Murthy This research investigates the effectiveness of implementing machine learning techniques to predict the effective thermal conductivity of the lithium-ion battery electrodes. It uses scanned images of the electrode to construct the discretized computational domain. An image analysis determines the position of active particles in the porous medium. Using a uniform grid, a conservative finite volume method finds the effective thermal conductivity of the porous medium based on different conductivities of each material. To perform the machine learning, the number of training samples is set up to be one order of magnitude more than the total number of grid points in the representative elemental volume. The training and testing groups are formed by sampling over random places in the image. Then, a deep learning network is trained to predict the effective thermal conductivity of the medium based on the geometry i.e. position and size of the active particles. The predictions are within 4% of the simulation results showing the accuracy of the machine learning method. Moreover, we show that the proposed new approach in which an image is taken as the input and the related effective thermal conductivity is obtained from the available trained network is more efficient than the simulation. |
Monday, November 19, 2018 11:01AM - 11:14AM |
G32.00003: LAT-NET++: Compressing Fluid Simulations using Deep Neural Networks Oliver Hennigh, Michael Chertkov We present extensions and improvements of our previous work Lat-Net [0], a deep learning based method to emulate Lattice Boltzmann fluid simulations for reduced computation and memory usage. Our first improvement is to add active learning in the training process which allows intelligent sampling of the train set. Second, we decouple Lat-Net from the Lattice Boltzmann Method allowing our method to be used in conjunction with other flow solvers. Third, we conduct rigorous tests of our method by looking at various statistical properties of the predicted flow. In addition to this, we present a method to optimize parameters of large eddy simulations such as the Smagorinsky constant. Following a similar structure as Lat-Net, we treat these constants as trainable parameters and optimize them with gradient descent. This approach can be viewed either as heavily constraining Lat-Net with the underlying physics of the flow solver or a data driven method to optimize parameters of sub-grid scale models.
[0] https://arxiv.org/abs/1705.09036 |
Monday, November 19, 2018 11:14AM - 11:27AM |
G32.00004: CFD simulations of a data center to train an artificial neural network model Jayati Athavale, Minami Yoda, Yogendra Kumar Joshi Data centers, large facilities that host computing and networking equipment for dealing with large volumes of data, are the physical manifestation of the “cloud.” This study presents an experimentally validated room-level computational fluid dynamics (CFD) simulation of a raised-floor data center configuration consisting of one cold aisle with six racks on each side, and three computer room air conditioning units around the room periphery. Predictions from the finite-volume software package Future Facilities 6SigmaDCX, employing a pressure-based solver, are in good agreement with experimental measurements of total air flow rate and rack inlet temperatures, with average discrepancies less than 4% and 1.7 °C, respectively. The numerical predictions using this approach over a variety of operating conditions are used to train an artificial neural network (ANN)-based model to predict temperature and airflow distributions in near real time. The ANN model, with its rapid prediction capability, can then be used to develop a control framework to minimize power consumption in data centers, which accounts for more than 2% of total American electricity consumption. |
Monday, November 19, 2018 11:27AM - 11:40AM |
G32.00005: Data-driven prediction of unsteady flow over a circular cylinder using deep learning Sangseung Lee, Donghyun You Unsteady flow fields over a circular cylinder are predicted using deep learning networks. Deep learning networks construct nonlinear mappings that allows prediction of flow fields at future occasions based on flow fields at past occasions. Deep learning networks equipped with four different loss-function configurations are trained using flow fields at ReD = 100, 200, 300, and 400. Two networks are trained using different loss-function sets: with and without loss functions for mass and momentum conservation, and two other networks are trained using loss-function sets: with and without loss functions for mass and momentum conservation both with a loss function for adversarial training. The trained networks are employed to predict flow fields at ReD = 500, and 3000, at which Reynolds numbers, the networks are not exposed to flow fields a priori. Results predicted by each network are compared and analyzed to identify effects of the configuration of loss functions and the use of adversarial training on the predictive performance. |
Monday, November 19, 2018 11:40AM - 11:53AM |
G32.00006: A Machine Learning Model for Unsteady Wake Dynamics Tharindu Pradeeptha Miyanawala, Rajeev Jaiman This work introduces a novel physical model for laminar wake-body interaction systems by learning low-dimensional approximation. Of particular interest is to predict a long time series of unsteady flow dynamics using a learned low-dimensional model. We use convolutional neural networks (CNN) to learn wake-body interaction dynamics, which assemble layers of linear convolutions with nonlinear activations to extract low-dimensional features. We first project the high-fidelity time series data from the finite element Navier-Stokes solver to a low-dimensional subspace using proper orthogonal decomposition (POD). The time-dependent coefficients of the POD subspace are mapped to the flow field via a CNN with nonlinear rectification, and the CNN is trained using stochastic gradient descent method to predict the POD time coefficients when a new flow field is fed to it. The mean flow field, POD basis vectors and the trained CNN are used to predict the long time series of the flow fields and results are compared with the full-order simulation results. POD-CNN predictions maintain a remarkable accuracy throughout the long time series for the entire fluid domain including a highly nonlinear near wake region. |
Monday, November 19, 2018 11:53AM - 12:06PM |
G32.00007: Deep Learning-aided Spectral Analysis of Fluid Flows Balaji Jayaraman, Shivakanth Chary Puligilla A linear time-invariant (LTI) representation of Markovian dynamics through the Koopman operator is often leveraged for characterizing the spectral behavior of the dynamical systems. The key to the success of the Koopman-based representation and the relevance of the learned spectral characteristics to the underlying system depends on the appropriateness of the map that transforms the input state to the 'observable'. In this work, we leverage deep learning algorithms to deduce transformation maps from training data and ultimately the LTI Koopman operator. In particularly, we present (i) a deep Koopman network (DKN) that learns the observable and the transition operator simultaneously in a single network and (ii) a deep autoencoder network (DAN) that accomplishes the same in two steps. We compare the efficacy of these architectures to the class of Dynamic Mode Decomposition (DMD)-based Koopman approximation techniques. While deep learning-based methods are shown to be good at capturing temporally evolving dynamics for systems that can be represented by limited data. In this talk, we will focus on the accuracy of the spectral information, like physically important eigenmodes and eigenvalues found using these methods from canonical PDE systems and fluids flows. |
Monday, November 19, 2018 12:06PM - 12:19PM |
G32.00008: Physics-Informed Generative Adversarial Networks by Incorporating Conservation Laws Yang Zeng, Jinlong Wu, Heng Xiao Recently, machine learning techniques have proven to be successful in many data-driven physical modeling tasks, including in mimicking distributions of processes in complex systems using a flavor of deep neural networks called generative adversarial networks (GANs). GANs have also been designed to generate solutions of PDEs governing complex systems without having to numerically solve these PDEs, by using existing high-fidelity simulations or experimental data as training data. In this work, we present a physics-informed GAN by enforcing constraints of conservation laws to improve the quality of the generated solutions of GANs. We show that this physics-informed GAN generates more realistic solutions of potential flows compared to traditional GANs without any physical constraints. These results suggest that the physics-informed GAN is more suitable for the task of physical modeling and has great potential in many areas where directly simulating the physics is usually expansive, e.g., turbulence. |
Monday, November 19, 2018 12:19PM - 12:32PM |
G32.00009: Hidden Fluid Mechanics: Navier-Stokes Informed Deep Learning from the Passive Scalar Transport Alireza Yazdani, Maziar Raissi, George Em Karniadakis Inspired by the recent developments in physics-informed deep learning framework, we propose a novel Navier-Stokes informed neural networks that encodes the governing equations of fluid motions i.e., mass, momentum and transport equations to infer hidden quantities of interest such as velocity and pressure fields merely from spatio-temporal visualizations of a passive scaler (e.g. dye or smoke) transported in arbitrarily complex domains. Our approach towards solving the aforementioned data assimilation problem is unique as we design an algorithm that is agnostic to the geometry or the initial and boundary conditions. This makes the algorithm highly flexible in choosing the spatio-temporal domain of interest for data acquisition as well as subsequent training and predictions. The proposed algorithm achieves accurate predictions of the pressure and velocity fields in both two and three dimensional flows for several benchmark problems motivated by real-world applications. Our results demonstrate that this relatively simple methodology can be used in physical and biomedical problems to extract valuable quantitative information (e.g. lift and drag forces or wall shear stresses in arteries) for which direct measurements may not be possible. |
Monday, November 19, 2018 12:32PM - 12:45PM |
G32.00010: Feature Engineering for Data Driven Applications in Physical Sciences Aashwin Mishra, Gianluca Iaccarino In data driven applications, features are attributes of the system used to represent the underlying problem to the algorithm. Feature engineering involves the transformation of raw data into descriptive and discriminative elements. In addition to improved performance from predictive models, this can be used to improve the interpretability of the model. Owing to its importance, approaches have been developed for feature construction, selection and transformation. However, the nature of data produced in physical science problems makes some of these approaches sub-optimal, while others may be rendered misleading. In this talk, we focus on the pertinence of different approaches to feature engineering in physical science applications. In particular we investigate the issue of multi-collinearity amongst groups of features of arbitrary sizes using illustrative datasets. Finally, we compare the performance of different approaches on a corpus of data from fluid flow simulations.
|
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. |
© 2025 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