Output-based adaptive aerodynamic simulations using convolutional neural networks
Introduction
Thanks to fast-paced increases in computing power and highly-developed numerical methods, computational fluid dynamics (CFD) has become commonplace in aerospace design and analysis over the last few decades. Although CFD simulations are now routinely carried out in aerospace applications, the resulting CFD solutions often come with low reliability, without active quantification of the numerical errors. The two main categories of numerical errors in CFD simulations are modeling errors due to assumptions or simplifications of the actual physics, and discretization errors induced by the finite-dimensional discretization of the continuous physical model. Both types of errors significantly affect the numerical solutions of CFD simulations, which can often lead to non-negligible errors in the outputs of interest such as drag and lift.
Typically, an appropriate model is chosen based on the best knowledge or the experience of the practitioner. This task is often highly problem-dependent and generally non-trivial for non-expert users. In order to reduce the error associated with physical models, calibration can be performed based on the data from experiments or direct numerical simulations, which remains an active research area [1], [2], [3], [4]. In this paper, we focus on the error caused by finite-dimensional discretizations of the continuous well-selected model, i.e., the governing equations are assumed to be accurate. Commonly used a priori meshes in CFD runs, even when generated with best practice guidelines, cannot guarantee accurate solutions [5]. Quantifying the uncertainty due to discretization errors is thus essential for the reliable use of CFD in practice. However, this liability cannot be managed easily for complex flow-fields, even by experienced practitioners.
Luckily, adjoint-based error estimation, also known as the dual-weighted residual method, provides a robust and effective approach to quantify the effect of discretization error on a chosen output of interest [6], [7], [8]. The adjoint variables weight the local readily-available flow residual to form an error measure of the output, which can be used to provide error bounds or a pure signed correction for the output. The key feature of the adjoint-based error estimation is the ability to localize the output error and to identify the regions important for accurate output prediction. Solution-adaptive methods via adjoint-based output error estimation have dramatically improved the accuracy and efficiency of CFD [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19]. Despite the great success in CFD applications, the additional computational cost and implementation complexity associated with adjoint-based methods cannot be neglected. On the one hand, adjoint-based methods require solving a dual linear system, i.e., the adjoint equation set, which is of the same size as the flow problem or even larger when solving on an enriched space. This additional cost can be mitigated in problems where the adjoint solutions on the current space are solved regardless, such as gradient-based optimization with error estimation and mesh adaptation [20], [21], [22], [23], [24]; yet for problems such as unsteady simulations, uncertainty quantification or gradient-free optimization, the extra cost associated with adjoint solutions compound and can be accumulated when the adjoint is repeatedly solved. On the other hand, the implementation of adjoint methods often requires the transpose of the residual Jacobian matrix, which is not always available in explicit solvers or Jacobian-free methods [25]. In these circumstances, either the continuous adjoint equations should be derived and directly discretized [26] or special implementation efforts are required [27], adding considerable costs and efforts in the development. Moreover, adjoint consistency [20], [28], [29], which is critical for effective adaptation, requires special treatment and imposes additional difficulties on the adjoint implementation. The additional computational costs associated with the adjoint solves, in addition to the implementation efforts, has largely hindered the effective use of adjoint-based error estimation and the corresponding adaptation techniques in practice.
In the past decade, error surrogate models based on machine learning techniques have received much attention, largely because of their non-intrusive nature and fast online evaluations. Several contributions have been made in error modeling for parameterized reduced-order models (ROMs) [30], [31], and the ideas have been extended to estimates of discretization-induced errors [32]. Efforts have also been devoted to predicting the errors in flow solutions and the outputs of interest obtained on coarse computational meshes [33], [34], and the models have also been used to guide the selection of a set of a priori meshes [35]. Nonetheless, in these studies, no output error indicator is provided to perform mesh adaptation. Manevitz et al. used neural networks to predict the solution gradients in time-dependent problems, which then provided an indicator to drive the mesh adaptation [36]. However, these feature-based adaptive indicators are generally not as effective as adjoint-based indicators, especially for functional outputs and problems with discontinuities [19], [37]. Furthermore, these works rely on a set of user-selected local features (feature engineering) to construct the model, requiring either expert knowledge or fine-tuning. Moreover, due to the local nature of the selected features (although some neighboring information comes in with the gradient features), these models either largely ignore the error transport such that they are not expected to be effective for convection-dominated problems, or still require the adjoint variables to bring in the global sensitivity information.
In this paper, we focus on inferring the output error for a CFD simulation, as well as the corresponding localized error indicator field to drive mesh adaptation, directly from the solution field. The latter task is more challenging as both the flow state field and the output error indicator field can be high-dimensional. Moreover, effective error indicators must take the error transport into account, especially for convection-dominated systems. Without solving for the output adjoint variables, we seek other approaches to discover the global output sensitivity accounting for the error transport. Formally, the adjoint solution can be regarded as a generalized Green’s function, which convolve the residual perturbation to produce an output perturbation (error estimate) [38]. In order to emulate the adjoint operator, we introduce convolutional neural networks (CNNs) to construct the surrogate error model. In particular, a set of discrete linear convolution operators is trained to approximate the generalized Green’s function which convolves the discretized solution field to produce the corresponding output error with respect to a refined space. In other words, the network can be regarded as an approximate adjoint-weighted-residual operator applied to the solution field, which produces the whole error indicator field as well as the total output error. On the other hand, the convolution operators preserve the spatial locality and are shared for the input solution field. As a result, the dimension of the free parameters in the network model scales well for large scale problems, making it well-suited for the high-dimensional map between the input solution and the output error indicator fields.
A CNN architecture that is especially efficient for this type of tasks is the encoder-decoder. It has shown excellent performance for image semantic segmentation and feature extraction in computer vision tasks [39], [40], [41], [42], [43], and has recently been popularized in physical modeling applications [44], [45], [46], [47] as well. The network is composed of two subnetworks: an encoder CNN that extracts a low-dimensional representation (code) from the input data, i.e., the solution field, followed by a decoder CNN that reconstructs the high-dimensional output field, i.e., the adaptive error indicator field. The ability of a CNN to automatically learn internal invariant features and multi-scale feature hierarchies alleviates the need for a tedious, hand-crafted feature engineering process, making this approach more flexible and robust. Instead of using the network output field to obtain the total output error, we connect the codes (low-dimensional representations) extracted from the input field to a fully-connected network (FCN) to predict the total output error. The network training is supervised by both the adaptive error indicator field and the total output error to capture both the local and global features related to the numerical error. Since the two regression tasks are trained simultaneously, separate models and additional training costs are avoided.
The remainder of this paper proceeds as follows. We introduce the standard adjoint-based output error estimation and mesh adaptation in Section 2. Section 3 presents the details of the proposed CNN model and the training procedure. The primary results are shown in Section 4. Section 5 concludes the present work and discusses potential future work.
Section snippets
Parameterized governing equations
In this work, we consider parameterized steady-state flow governing equations in a fully-discretized form,where is a vector of parameters sampled from the parameter space characterizing the physics of the system, e.g., initial and boundary conditions, material properties, or shape parameters in a design optimization problem; denotes the flow state vector, uniquely defining the continuous flow state field where is the approximation space defined by a
Surrogate model as a regression problem
The error surrogate model can be treated as two regression problems: given the input solution vector from a CFD simulation we would like to predict the scalar output error as well as the adaptive error indicator field over the entire mesh. Here we omit the subscript for simpler exposition. The output and input dimensions can be very different. For example in a finite-element simulation, the state vector can be post-processed into several state components of the same dimension
Results
We test our proposed model in inviscid transonic aerodynamic flow simulations over airfoils, which involve geometries and irregular computational domains. The Euler equations govern the flow, and these are discretized using a discontinuous Galerkin finite element method [64], [65]. An element-based artificial viscosity [66] is adopted for shock capturing. We use the Roe approximate Riemann solver [67] for the inviscid flux, and the second form of Bassi and Rebay treatment for the viscous flux
Conclusion
Output error quantification and mesh adaptation are essential for the reliable use of CFD. However, they are in general non-trivial tasks, even for experienced users. Adjoint-based methods provide a robust approach to estimate and reduce the output error through effective error estimation and mesh adaptation. However, the reliance on the output adjoint solutions imposes both implementation and cost challenges in practice. We propose a new method to manage this liability with machine learning
CRediT authorship contribution statement
Guodong Chen: Conceptualization, Methodology, Software, Investigation, Data curation, Writing - original draft. Krzysztof J. Fidkowski: Conceptualization, Methodology, Writing - review & editing, Supervision, Project administration, Funding acquisition.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
The authors acknowledge support from the Department of Energy under grant DE-FG02-13ER26146/DE-SC0010341, and from The Boeing Company, with technical monitor Dr. Mori Mani. Guodong Chen’s work was also supported by the Michigan Institute for Computational Discovery and Engineering Fellowship, and the Rackham Graduate Student Research Grant from the University of Michigan.
References (71)
- et al.
Verification and validation in computational fluid dynamics
Prog Aerosp Sci
(2002) - et al.
Bayesian calibration of the constants of the turbulence model for a CFD model of street canyon flow
Comput Methods Appl Mech Eng
(2014) - et al.
A paradigm for data-driven predictive modeling using field inversion and machine learning
J Comput Phys
(2016) - et al.
Adaptive discontinuous Galerkin finite element methods for the compressible Euler equations
J Comput Phys
(2002) - et al.
Grid adaptation for functional outputs: application to two-dimensional inviscid flows
J Comput Phys
(2002) - et al.
A triangular cut-cell adaptive method for high-order discretizations of the compressible Navier–Stokes equations
J Comput Phys
(2007) - et al.
Adjoint-based adaptive discontinuous Galerkin methods for the 2D compressible Euler equations
J Comput Phys
(2009) - et al.
Fully anisotropic goal-oriented mesh adaptation for 3d steady Euler equations
J Comput Phys
(2010) - et al.
An optimization-based framework for anisotropic simplex mesh adaptation
J Comput Phys
(2012) - et al.
An optimization-based framework for anisotropic hp-adaptation of high-order discretizations
J Comput Phys
(2018)
Adjoint-based airfoil optimization with discretization error control
Int J Numer Methods Fluids
Jacobian-free Newton–Krylov methods: a survey of approaches and applications
J Comput Phys
A comparison of the continuous and discrete adjoint approach to automatic aerodynamic optimization
38th aerospace sciences meeting and exhibit, AIAA paper 2000-667
Adjoint consistency analysis of discontinuous Galerkin discretizations
SIAM J Numer Anal
Predicting goal error evolution from near-initial-information: a learning algorithm
J Comput Phys
Machine-learning based error prediction approach for coarse-grid computational fluid dynamics (CG-CFD)
Prog Nucl Energy
A data-driven framework for error estimation and mesh-model optimization in system-level thermal-hydraulic simulation
Nucl Eng Des
Neural network time series forecasting of finite-element mesh adaptation
Neurocomputing
Comparison of adjoint-based and feature-based grid adaptation for functional outputs
Int J Numer Methods Fluids
Prediction of aerodynamic flow fields using convolutional neural networks
Comput Mech
Aerodynamic design via control theory
J Sci Comput
Sub-cell shock capturing for discontinuous Galerkin methods
44th AIAA aerospace sciences meeting and exhibit, AIAA paper 2006-0112
The Characteristics of 78 Related Airfoil Sections from Tests in the Variable-Density Wind Tunnel
Tech. Rep.
Turbulence modeling in the age of data
Annu Rev Fluid Mech
Data summary from the first AIAA computational fluid dynamics drag prediction workshop
J Aircr
An optimal control approach to a posteriori error estimation in finite element methods
Acta Numer
Adjoint recovery of superconvergent functionals from PDE approximations
SIAM Rev
Adjoint methods for PDEs: a posteriori error analysis and postprocessing by duality
Acta Numer
Adjoint-based, three-dimensional error prediction and grid adaptation
AIAA J
Adjoint error estimation and adaptive refinement for embedded-boundary cartesian meshes
18th AIAA computational fluid dynamics conference, AIAA paper 2007–4187
Adjoint-based adaptive mesh refinement for complex geometries
46th AIAA aerospace sciences meeting and exhibit
Review of output-based error estimation and mesh adaptation in computational fluid dynamics
AIAA J
An a Posteriori Error Control Framework for Adaptive Precision Optimization Using Discontinuous Galerkin Finite Element Method
Output error estimates and mesh refinement in aerodynamic shape optimization
51st AIAA aerospace sciences meeting including the new horizons forum and aerospace exposition, AIAA paper 2013-865
PDE-constrained optimization with error estimation and control
J Comput Phys
Cited by (12)
DynAMO: Multi-agent reinforcement learning for dynamic anticipatory mesh optimization with applications to hyperbolic conservation laws
2024, Journal of Computational PhysicsSuperAdjoint: Super-resolution neural networks in adjoint-based error estimation
2024, Journal of Computational and Applied MathematicsMesh optimization using an improved self-organizing mechanism
2023, Computers and FluidsQuasi-optimal hp-finite element refinements towards singularities via deep neural network prediction
2023, Computers and Mathematics with ApplicationsLEARNING ROBUST MARKING POLICIES FOR ADAPTIVE MESH REFINEMENT
2024, SIAM Journal on Scientific ComputingMachine learning mesh-adaptation for laminar and turbulent flows: applications to high-order discontinuous Galerkin solvers
2024, Engineering with Computers