arXiv subjects

Karthik Duraisamy

Publications and source records attributed to Karthik Duraisamy.

At least 55 records · Page 3Linked to original sources

arXiv2022

A unified understanding of scale-resolving simulations and near-wall modeling of turbulent flows using optimal finite element projections

The main objective of this work is to develop a unified framework that can be used as a lens to quantitatively assess and augment a wide range of coarse-grained models of turbulence, viz. large eddy simulations (LES), hybrid Reynolds-averaged/LES methods and wall-modeled (WM)LES. Taking a turbulent channel flow as an example, optimality is assessed in the wall-resolved limit, the hybrid RANS/LES limit and the WMLES limit, via projections at different resolutions suitable for these approaches. These optimal a priori estimates are shown to have similar characteristics to existing a posteriori solutions reported in the literature. Consistent accuracy metrics are developed for scale-resolving methods using the optimal solution as a reference, and evaluations are performed. We further characterize the slip velocity in WMLES in terms of the near-wall under-resolution and develop a universal scaling relationship. Insights from the a-priori tests are used to augment existing slip-based wall models. Various a posteriori tests reveal superior performance over the dynamic slip wall model. Guidance for the development of improved slip-wall models is provided, including a target for the dynamic procedure.

physics.flu-dyn↗

arXiv2022

A Non-intrusive Approach for Physics-constrained Learning with Application to Fuel Cell Modeling

A data-driven model augmentation framework, referred to as Weakly-coupled Integrated Inference and Machine Learning (IIML), is presented to improve the predictive accuracy of physical models. In contrast to parameter calibration, this work seeks corrections to the structure of the model by a) inferring augmentation fields that are consistent with the underlying model, and b) transforming these fields into corrective model forms. The proposed approach couples the inference and learning steps in a weak sense via an alternating optimization approach. This coupling ensures that the augmentation fields remain learnable and maintain consistent functional relationships with local modeled quantities across the training dataset. An iterative solution procedure is presented in this paper, removing the need to embed the augmentation function during the inference process. This framework is used to infer an augmentation introduced within a Polymer electrolyte membrane fuel cell (PEMFC) model using a small amount of training data (from only 14 training cases.) These training cases belong to a dataset consisting of high-fidelity simulation data obtained from a high-fidelity model of a first generation Toyota Mirai. All cases in this dataset are characterized by different inflow and outflow conditions on the same geometry. When tested on 1224 different configurations, the inferred augmentation significantly improves the predictive accuracy for a wide range of physical conditions. Predictions and available data for the current density distribution are also compared to demonstrate the predictive capability of the model for quantities of interest which were not involved in the inference process. The results demonstrate that the weakly-coupled IIML framework offers sophisticated and robust model augmentation capabilities without requiring extensive changes to the numerical solver.

cs.CE↗

arXiv2022

Discretization-independent surrogate modeling over complex geometries using hypernetworks and implicit representations

Numerical solutions of partial differential equations (PDEs) require expensive simulations, limiting their application in design optimization, model-based control, and large-scale inverse problems. Surrogate modeling techniques seek to decrease the computational expense while retaining dominant solution features and behavior. Traditional Convolutional Neural Network-based frameworks for surrogate modeling require lossy pixelization and data-preprocessing, and generally are not effective in realistic engineering applications. We propose alternative deep-learning based surrogate models for discretization-independent, continuous representations of PDE solutions, which can be used for learning and prediction over domains with complex, variable geometry and mesh topology. Three methods are proposed and compared; design-variable-coded multi-layer perceptron (DV-MLP), design-variable hypernetworks (DV-Hnet), and non-linear independent dual system (NIDS). Each method utilizes a main network which consumes pointwise spatial information to provide a continuous representation, allowing predictions at any location in the domain. Input features include a minimum-distance function evaluation to implicitly encode the problem geometry. The geometric design variables, which define and distinguish problem instances, are used differently by each method, appearing as additional main-network input features (DV-MLP), or as hypernetwork inputs (DV-Hnet and NIDS). The methods are applied to predict solutions around complex, parametrically-defined geometries on non-parametrically-defined meshes with model predictions obtained many orders of magnitude faster than the full order models. Test cases include a vehicle-aerodynamics problem with complex geometry and limited training data, with a design-variable hypernetwork performing best, with a competitive time-to-best-model despite a much greater parameter count.

physics.comp-ph↗

arXiv2022

Entropy-Stable Schemes in the Low-Mach-Number Regime: Flux-Preconditioning, Entropy Breakdowns, and Entropy Transfers

Entropy-Stable (ES) schemes, specifically those built from [Tadmor \textit{Math. Comput.} 49 (1987) 91], have been gaining interest over the past decade, especially in the context of under-resolved simulations of compressible turbulent flows using high-order methods. These schemes are attractive because they can provide stability in a global and nonlinear sense (consistency with thermodynamics). However, fully realizing the potential of ES schemes requires a better grasp of their local behavior. Entropy-stability itself does not imply good local behavior [Gouasmi \textit{et al.} \textit{J. Sci. Comp.} 78 (2019) 971, Gouasmi \textit{et al.} \textit{Comput. Methd. Appl. M.} 363 (2020) 112912]. In this spirit, we studied ES schemes in problems where \textit{global stability is not the core issue}. In the present work, we consider the accuracy degradation issues typically encountered by upwind-type schemes in the low-Mach-number regime [Turkel \textit{Annu. Rev. Fluid Mech.} 31 (1999) 285] and their treatment using \textit{Flux-Preconditioning} [Turkel \textit{J. Comput. Phys.} 72 (1987) 277, Miczek \textit{et al.} \textit{A \& A} 576 (2015) A50]. ES schemes suffer from the same issues and Flux-Preconditioning can improve their behavior without interfering with entropy-stability. This is first demonstrated analytically: using similarity and congruence transforms we were able to establish conditions for a preconditioned flux to be ES, and introduce the ES variants of the Miczek's and Turkel's preconditioned fluxes. This is then demonstrated numerically through first-order simulations of two simple test problems representative of the incompressible and acoustic limits, the Gresho Vortex and a right-moving acoustic wave. The results are overall consistent with previous studies [...]

physics.flu-dyn↗

arXiv2021

Non-intrusive Balancing Transformation of Highly Stiff Systems with Lightly-damped Impulse Response

Balanced truncation (BT) is a model reduction method that utilizes a coordinate transformation to retain eigen-directions that are highly observable and reachable. To address realizability and scalability of BT applied to highly stiff and lightly-damped systems, a non-intrusive data-driven method is developed for balancing discrete-time systems via the eigensystem realization algorithm (ERA). The advantage of ERA for balancing transformation makes full-state outputs tractable. Further, ERA enables balancing despite stiffness, by eliminating computation of balancing modes and adjoint simulations. As a demonstrative example, we create balanced ROMs for a one-dimensional reactive flow with pressure forcing, where the stiffness introduced by the chemical source term is extreme (condition number $10^{13}$), preventing analytical implementation of BT. We investigate the performance of ROMs in prediction of dynamics with unseen forcing inputs and demonstrate stability and accuracy of balanced ROMs in truly predictive scenarios whereas without ERA, POD-Galerkin and Least-squares Petrov-Galerkin projections fail to represent the true dynamics. We show that after the initial transients under unit impulse forcing, the system undergoes lightly-damped oscillations, which magnifies the influence of sampling properties on predictive performance of the balanced ROMs. We propose an output domain decomposition approach and couple it with tangential interpolation to resolve sharp gradients at reduced computational costs.

eess.SY↗

arXiv2021

Disentangling Generative Factors of Physical Fields Using Variational Autoencoders

The ability to extract generative parameters from high-dimensional fields of data in an unsupervised manner is a highly desirable yet unrealized goal in computational physics. This work explores the use of variational autoencoders (VAEs) for non-linear dimension reduction with the specific aim of {\em disentangling} the low-dimensional latent variables to identify independent physical parameters that generated the data. A disentangled decomposition is interpretable, and can be transferred to a variety of tasks including generative modeling, design optimization, and probabilistic reduced order modelling. A major emphasis of this work is to characterize disentanglement using VAEs while minimally modifying the classic VAE loss function (i.e. the Evidence Lower Bound) to maintain high reconstruction accuracy. The loss landscape is characterized by over-regularized local minima which surround desirable solutions. We illustrate comparisons between disentangled and entangled representations by juxtaposing learned latent distributions and the true generative factors in a model porous flow problem. Hierarchical priors are shown to facilitate the learning of disentangled representations. The regularization loss is unaffected by latent rotation when training with rotationally-invariant priors, and thus learning non-rotationally-invariant priors aids in capturing the properties of generative factors, improving disentanglement. Finally, it is shown that semi-supervised learning - accomplished by labeling a small number of samples ($O(1\%)$) - results in accurate disentangled latent representations that can be consistently learned.

physics.comp-ph↗

arXiv2021

Variational Multi-scale Super-resolution : A data-driven approach for reconstruction and predictive modeling of unresolved physics

The variational multiscale (VMS) formulation formally segregates the evolution of the coarse-scales from the fine-scales. VMS modeling requires the approximation of the impact of the fine scales in terms of the coarse scales. In linear problems, our formulation reduces the problem of learning the sub-scales to learning the projected element Green's function basis coefficients. For the purpose of this approximation, a special neural-network structure - the variational super-resolution N-N (VSRNN) - is proposed. The VSRNN constructs a super-resolved model of the unresolved scales as a sum of the products of individual functions of coarse scales and physics-informed parameters. Combined with a set of locally non-dimensional features obtained by normalizing the input coarse-scale and output sub-scale basis coefficients, the VSRNN provides a general framework for the discovery of closures for both the continuous and the discontinuous Galerkin discretizations. By training this model on a sequence of $L_2-$projected data and using the subscale to compute the continuous Galerkin subgrid terms, and the super-resolved state to compute the discontinuous Galerkin fluxes, we improve the optimality and the accuracy of these methods for the convection-diffusion problem, linear advection and turbulent channel flow. Finally, we demonstrate that - in the investigated examples - the present model allows generalization to out-of-sample initial conditions and Reynolds numbers. Perspectives are provided on data-driven closure modeling, limitations of the present approach, and opportunities for improvement.

physics.comp-ph↗

arXiv2021

Conditionally Parameterized, Discretization-Aware Neural Networks for Mesh-Based Modeling of Physical Systems

Simulations of complex physical systems are typically realized by discretizing partial differential equations (PDEs) on unstructured meshes. While neural networks have recently been explored for surrogate and reduced order modeling of PDE solutions, they often ignore interactions or hierarchical relations between input features, and process them as concatenated mixtures. We generalize the idea of conditional parameterization -- using trainable functions of input parameters to generate the weights of a neural network, and extend them in a flexible way to encode critical information. Inspired by discretized numerical methods, choices of the parameters include physical quantities and mesh topology features. The functional relation between the modeled features and the parameters is built into the network architecture. The method is implemented on different networks and applied to frontier scientific machine learning tasks including the discovery of unmodeled physics, super-resolution of coarse fields, and the simulation of unsteady flows with chemical reactions. The results show that the conditionally-parameterized networks provide superior performance compared to their traditional counterparts. The CP-GNet - an architecture that can be trained on very few data snapshots - is proposed as the first deep learning model capable of standalone prediction of reacting flows on irregular meshes.

cs.LG↗

arXiv2021

Model Reduction for Multi-Scale Transport Problems using Model-form Preserving Least-Squares Projections with Variable Transformation

A projection-based formulation is presented for non-linear model reduction of problems with extreme scale disparity. The approach allows for the selection of an arbitrary, but complete, set of solution variables while preserving the structure of the governing equations. Least-squares-based minimization is leveraged to guarantee symmetrization and discrete consistency with the full-order model (FOM). Two levels of scaling are used to achieve the conditioning required to effectively handle problems with extremely disparate physical phenomena, characterized by extreme stiffness in the system of equations. The formulation -- referred to as model-form preserving least-squares with variable transformation (MP-LSVT) -- provides global stabilization for both implicit and explicit time integration schemes. To achieve computational efficiency, a pivoted QR decomposition is used with oversampling, and adapted to the MP-LSVT method. The framework is demonstrated in representative two- and three-dimensional reacting flow problems, and the MP-LSVT is shown to exhibit improved stability and accuracy over standard projection-based ROM techniques. Physical realizability and local stability are promoted by enforcing limiters in both temperature and species mass fractions. These limiters are demonstrated to be important in eliminating regions of spurious burning, thus enabling the ROMs to provide accurate representations of the heat release rate and flame propagation speed. In the 3D application, it is shown that more than two orders of magnitude acceleration in computational efficiency can be achieved, while also providing reasonable future-state predictions. A key contribution of this work is the development and demonstration of a comprehensive ROM formulation that targets highly challenging multi-scale transport-dominated problems.

physics.comp-ph↗

arXiv2021

`Basic' Generalization Error Bounds for Least Squares Regression with Well-specified Models

This note examines the behavior of generalization capabilities - as defined by out-of-sample mean squared error (MSE) - of Linear Gaussian (with a fixed design matrix) and Linear Least Squares regression. Particularly, we consider a well-specified model setting, i.e. we assume that there exists a `true' combination of model parameters within the chosen model form. While the statistical properties of Least Squares regression have been extensively studied over the past few decades - particularly with {\bf less restrictive problem statements} compared to the present work - this note targets bounds that are {\bf non-asymptotic and more quantitative} compared to the literature. Further, the analytical formulae for distributions and bounds (on the MSE) are directly compared to numerical experiments. Derivations are presented in a self-contained and pedagogical manner, in a way that a reader with a basic knowledge of probability and statistics can follow.

math.ST↗

arXiv2021

Generalizable Physics-constrained Modeling using Learning and Inference assisted by Feature Space Engineering

This work presents a formalism to improve the predictive accuracy of physical models by learning generalizable augmentations from sparse data. Building on recent advances in data-driven turbulence modeling, the present approach, referred to as Learning and Inference assisted by Feature-space Engineering (LIFE), is based on the hypothesis that robustness and generalizability demand a meticulously-designed feature space that is informed by the underlying physics, and a carefully constructed features-to-augmentation map. The critical components of this approach are: (1) Maintaining consistency across the learning and prediction environments; (2) Tightly-coupled inference and learning by constraining the augmentation to be learnable throughout the inference process; (3) Identification of relevant physics-informed features in appropriate functional forms to enable significant overlap in feature space for a wide variety of cases to promote generalizability; (4) Maintaining explicit control over feature space to change the augmentation function behavior only in the vicinity of available datapoints. To demonstrate the viability of this approach, it is used in the modeling of bypass transition. The augmentation is developed on skin friction data from two flat plate cases from the ERCOFTAC dataset. Piecewise linear interpolation on a structured grid in feature-space is used as a sample functional form for the augmentation to demonstrate the capability of localized learning. The augmented model is then applied to a variety of flat plate cases which are characterized by different freestream turbulence intensities, pressure gradients, and Reynolds numbers. The predictive capability of the augmented model is also tested on single-stage high-pressure-turbine cascade cases, and the model performance is analyzed from the perspective of information contained in the feature space.

physics.flu-dyn↗

arXiv2021

Variational Encoders and Autoencoders : Information-theoretic Inference and Closed-form Solutions

This work develops problem statements related to encoders and autoencoders with the goal of elucidating variational formulations and establishing clear connections to information-theoretic concepts. Specifically, four problems with varying levels of input are considered : a) The data, likelihood and prior distributions are given, b) The data and likelihood are given; c) The data and prior are given; d) the data and the dimensionality of the parameters is specified. The first two problems seek encoders (or the posterior) and the latter two seek autoencoders (i.e. the posterior and the likelihood). A variational Bayesian setting is pursued, and detailed derivations are provided for the resulting optimization problem. Following this, a linear Gaussian setting is adopted, and closed form solutions are derived. Numerical experiments are also performed to verify expected behavior and assess convergence properties. Explicit connections are made to rate-distortion theory, information bottleneck theory, and the related concept of sufficiency of statistics is also explored. One of the motivations of this work is to present the theory and learning dynamics associated with variational inference and autoencoders, and to expose information theoretic concepts from a computational science perspective.

cs.IT↗

arXiv2021

A formal proof of the Lax equivalence theorem for finite difference schemes

The behavior of physical systems is typically modeled using differential equations which are too complex to solve analytically. In practical problems, these equations are discretized on a computational domain, and numerical solutions are computed. A numerical scheme is called convergent, if in the limit of infinitesimal discretization, the bounds on the discretization error is also infinitesimally small. The approximate solution converges to the "true solution" in this limit. The Lax equivalence theorem enables a proof of convergence given consistency and stability of the method. In this work, we formally prove the Lax equivalence theorem using the Coq Proof Assistant. We assume a continuous linear differential operator between complete normed spaces, and define an equivalent mapping in the discretized space. Given that the numerical method is consistent (i.e., the discretization error tends to zero as the discretization step tends to zero), and the method is stable (i.e., the error is uniformly bounded), we formally prove that the approximate solution converges to the true solution. We then demonstrate convergence of the difference scheme on an example problem by proving both its consistency and stability, and then applying the Lax equivalence theorem. In order to prove consistency, we use the Taylor-Lagrange theorem by formally showing that the discretization error is bounded above by the nth power of the discretization step, where n is the order of the truncated Taylor polynomial.

math.NA↗

arXiv2021

Sub-grid scale characterization and asymptotic behavior of multi-dimensional upwind schemes for the vorticity transport equations

We study the sub-grid scale characteristics of a vorticity-transport-based approach for large-eddy simulations. In particular, we consider a multi-dimensional upwind scheme for the vorticity transport equations and establish its properties in the under-resolved regime. The asymptotic behavior of key turbulence statistics of velocity gradients, vorticity, and invariants is studied in detail. Modified equation analysis indicates that dissipation can be controlled locally via non-linear limiting of the gradient employed for the vorticity reconstruction on the cell face such that low numerical diffusion is obtained in well-resolved regimes and high numerical diffusion is realized in under-resolved regions. The enstrophy budget highlights the remarkable ability of the truncation terms to mimic the true sub-grid scale dissipation and diffusion. The modified equation also reveals diffusive terms that are similar to several commonly employed sub-grid scale models including tensor-gradient and hyper-viscosity models. Investigations on several canonical turbulence flow cases show that large-scale features are adequately represented and remain consistent in terms of spectral energy over a range of grid resolutions. Numerical dissipation in under-resolved simulations is consistent and can be characterized by diffusion terms discovered in the modified equation analysis. A minimum state of scale separation necessary to obtain asymptotic behavior is characterized using metrics such as effective Reynolds number and effective grid spacing. Temporally-evolving jet simulations, characterized by large-scale vortical structures, demonstrate that high Reynolds number vortex-dominated flows are captured when criteria is met and necessitate diffusive non-linear limiting of vorticity reconstruction be employed to realize accuracy in under-resolved simulations.

physics.flu-dyn↗

arXiv2021

Perspectives on Machine Learning-augmented Reynolds-averaged and Large Eddy Simulation Models of Turbulence

This work presents a review and perspectives on recent developments in the use of machine learning (ML) to augment Reynolds-averaged Navier--Stokes (RANS) and Large Eddy Simulation (LES) models of turbulent flows. Different approaches of applying supervised learning to represent unclosed terms, model discrepancies and sub-filter scales are discussed in the context of RANS and LES modeling. Particular emphasis is placed on the impact of the training procedure on the consistency of ML augmentations with the underlying physical model. Techniques to promote model-consistent training, and to avoid the requirement of full fields of direct numerical simulation data are detailed. This is followed by a discussion of physics-informed and mathematical considerations on the choice of the feature space, and imposition of constraints on the ML model. With a view towards developing generalizable ML-augmented RANS and LES models, outstanding challenges are discussed, and perspectives are provided. While the promise of ML-augmented turbulence modeling is clear, and successes have been demonstrated in isolated scenarios, a general consensus of this paper is that truly generalizable models require model-consistent training with careful characterization of underlying assumptions and imposition of physically and mathematically informed priors and constraints to account for the inevitable shortage of data relevant to predictions of interest. Thus, machine learning should be viewed as one tool in the turbulence modeler's toolkit. This modeling endeavor requires multi-disciplinary advances, and thus the target audience for this paper is the fluid mechanics community, as well as the computational science and machine learning communities.

physics.flu-dyn↗

arXiv2021

Sparsity-promoting algorithms for the discovery of informative Koopman invariant subspaces

Koopman decomposition is a non-linear generalization of eigen-decomposition, and is being increasingly utilized in the analysis of spatio-temporal dynamics. Well-known techniques such as the dynamic mode decomposition (DMD) and its linear variants provide approximations to the Koopman operator, and have been applied extensively in many fluid dynamic problems. Despite being endowed with a richer dictionary of nonlinear observables, nonlinear variants of the DMD, such as extended/kernel dynamic mode decomposition (EDMD/KDMD) are seldom applied to large-scale problems primarily due to the difficulty of discerning the Koopman invariant subspace from thousands of resulting Koopman eigenmodes. To address this issue, we propose a framework based on multi-task feature learning to extract the most informative Koopman invariant subspace by removing redundant and spurious Koopman triplets. In particular, we develop a pruning procedure that penalizes departure from linear evolution. These algorithms can be viewed as sparsity promoting extensions of EDMD/KDMD. Further, we extend KDMD to a continuous-time setting and show a relationship between the present algorithm, sparsity-promoting DMD, and an empirical criterion from the viewpoint of non-convex optimization. The effectiveness of our algorithm is demonstrated on examples ranging from simple dynamical systems to two-dimensional cylinder wake flows at different Reynolds numbers and a three-dimensional turbulent ship air-wake flow. The latter two problems are designed such that very strong nonlinear transients are present, thus requiring an accurate approximation of the Koopman operator. Underlying physical mechanisms are analyzed, with an emphasis on characterizing transient dynamics. The results are compared to existing theoretical expositions and numerical approximations.

math.DS↗

arXiv2020

On the Structure of Time-delay Embedding in Linear Models of Non-linear Dynamical Systems

This work addresses fundamental issues related to the structure and conditioning of linear time-delayed models of non-linear dynamics on an attractor. While this approach has been well-studied in the asymptotic sense (e.g. for infinite number of delays), the non-asymptotic setting is not well-understood. First, we show that the minimal time-delays required for perfect signal recovery are solely determined by the sparsity in the Fourier spectrum for scalar systems. For the vector case, we provide a rank test and a geometric interpretation for the necessary and sufficient conditions for the existence of an accurate linear time delayed model. Further, we prove that the output controllability index of a linear system induced by the Fourier spectrum serves as a tight upper bound on the minimal number of time delays required. An explicit expression for the exact linear model in the spectral domain is also provided. From a numerical perspective, the effect of the sampling rate and the number of time delays on numerical conditioning is examined. An upper bound on the condition number is derived, with the implication that conditioning can be improved with additional time delays and/or decreasing sampling rates. Moreover, it is explicitly shown that the underlying dynamics can be accurately recovered using only a partial period of the attractor. Our analysis is first validated in simple periodic and quasi-periodic systems, and sensitivity to noise is also investigated. Finally, issues and practical strategies of choosing time delays in large-scale chaotic systems are discussed and demonstrated on 3D turbulent Rayleigh-Bénard convection.

math.DS↗

arXiv2020

Multi-level Convolutional Autoencoder Networks for Parametric Prediction of Spatio-temporal Dynamics

A data-driven framework is proposed towards the end of predictive modeling of complex spatio-temporal dynamics, leveraging nested non-linear manifolds. Three levels of neural networks are used, with the goal of predicting the future state of a system of interest in a parametric setting. A convolutional autoencoder is used as the top level to encode the high dimensional input data along spatial dimensions into a sequence of latent variables. A temporal convolutional autoencoder (TCAE) serves as the second level, which further encodes the output sequence from the first level along the temporal dimension, and outputs a set of latent variables that encapsulate the spatio-temporal evolution of the dynamics. The use of dilated temporal convolutions grows the receptive field exponentially with network depth, allowing for efficient processing of long temporal sequences typical of scientific computations. A fully-connected network is used as the third level to learn the mapping between these latent variables and the global parameters from training data, and predict them for new parameters. For future state predictions, the second level uses a temporal convolutional network to predict subsequent steps of the output sequence from the top level. Latent variables at the bottom-most level are decoded to obtain the dynamics in physical space at new global parameters and/or at a future time. Predictive capabilities are evaluated on a range of problems involving discontinuities, wave propagation, strong transients, and coherent structures. The sensitivity of the results to different modeling choices is assessed. The results suggest that given adequate data and careful training, effective data-driven predictive models can be constructed. Perspectives are provided on the present approach and its place in the landscape of model reduction.

physics.comp-ph↗