Non-intrusive data-driven model reduction for differential–algebraic equations derived from lifting transformations
Introduction
Model reduction is effective in reducing the computational cost of simulating complex systems but remains a challenging task when the governing physics exhibit nonlinear dynamics that are not readily amenable to low-dimensional approximations. Variable transformations combined with data-driven learning of the reduced-order model (ROM) operators have emerged as one strategy to address the challenges of nonlinear model reduction [1], [2]; however, for a large class of systems, including those that arise in reacting flow and phase-field models, the desired variable transformations lead to systems of partial differential–algebraic equations (PDAEs). This paper considers the form of the PDAEs that arise in lifting a nonlinear system to polynomial form and exploits that structure to extend the Operator Inference (OpInf) approach of [3] to these lifted PDAE systems.
As a driving application, we consider a solidification process in metal additive manufacturing. Additive manufacturing is a process during which a three-dimensional part is built via the layer-by-layer deposition of material according to its digital model. Additive manufacturing’s layer-wise process adds value by allowing for the manufacturing of components with complex geometries that are either infeasible or difficult to build by conventional manufacturing processes. However, the additive manufacturing process takes place over a wide range of length scales and time scales, making numerical simulations computationally expensive. Further, uncertainty quantification is essential since the structure and properties of the resulting components are sensitive to process parameter variations [4]. Thus, ROMs are key enablers to making control, optimization, and uncertainty quantification computationally feasible for additive manufacturing. ROMs can be used in multiscale modeling of additively manufactured parts to reduce the computational costs of part-scale simulations while maintaining the desired properties at microscale [5].
Our target problem poses several challenges for existing model reduction methods. First, the transport-dominated physics of the solidification interface result in highly localized changes in the state solution with time. Classical projection-based model reduction methods that seek approximations of the state in a linear subspace (see e.g., [6], [7], [8], [9]) require many modes to achieve accuracy in a problem such as this one, rendering the resulting ROMs inefficient. Second, the forward solidification model, a coupled system of nonlinear partial differential equations (PDEs) comprising a phase-field equation and a heat equation, has a strong nonlinear dependence on the process parameters. Classical projection-based model reduction methods that use hyper-reduction methods (such as the Empirical Interpolation Method [10] and the Discrete Empirical Interpolation Method [11]) will require many interpolation points to approximate the nonlinear terms, again rendering the resulting ROMs inefficient.
Methods based on variable transformations are becoming an effective alternative for model reduction of nonlinear systems of PDEs. These approaches draw upon the notion that the introduction of auxiliary variables (often referred to as “lifting”) can lead to a reformulation of the governing equations in a structured form. For example, [12] shows how general nonlinear ordinary differential equations (ODEs) can be written as so-called “polynomial ordinary differential systems” through the introduction of additional variables. In biology, variable transformations called “recasting” are used to transform nonlinear ODEs to the so-called S-system form, a polynomial form that is faster to solve numerically [13]. Approaches based on the Koopman operator lift a nonlinear dynamical system to an infinite-dimensional space in which the dynamics are linear [14], [15]. Ref. [16] introduced the idea of reformulating nonlinear dynamical systems in quadratic form for model reduction and showed that the number of auxiliary variables needed to lift a system to quadratic-bilinear form is linear in the number of elementary nonlinear functions in the original state equations. The work in [16] shows that a large class of nonlinear terms that appear in engineering systems (including monomial, sinusoidal, and exponential terms) may be lifted to quadratic form. Lifting has been extended to model reduction of problems governed by PDEs and shown to be a competitive alternative to hyper-reduction methods [1], [17]. Yet, for many practical applications it is neither feasible nor desirable to explicitly transform the high-fidelity PDE solver, which motivates the use of non-intrusive data-driven model reduction.
Following the definitions in [18], a non-intrusive model reduction method computes the ROM using outputs of the high-fidelity simulation but without access to the full-order operators (or to their action on a vector). This is in contrast to a black-box method, which computes the ROM without using a priori or explicit knowledge of the form of the high-fidelity problem definition or its numerical implementation. Ref. [18] notes that black-box methods are non-intrusive, but not all non-intrusive methods are necessarily black-box: a non-intrusive method can exploit knowledge of the high-fidelity problem definition and the corresponding problem structure, even though it does not access the full-order operators themselves. Gray-box methods combine some knowledge of the underlying problem with a data-driven component. Gray-box methods may or may not be non-intrusive. The advantage of non-intrusive approaches is that they compute the ROM directly from simulation data, without needing access to the high-fidelity operators. Methods such as dynamic mode decomposition (DMD) [19], [20], [21], [22], [23] and OpInf [3], [18], [24] operate on snapshot data (i.e., simulated state solutions), as do approaches that directly build surrogate models of the proper orthogonal decomposition (POD) modal coefficients [25], [26], [27], [28], [29]. Other non-intrusive methods, such as those based on the Loewner framework [30], [31], [32], [33], [34], require only compressed data (e.g., input–output measurements or transfer function measurements), and are input-invariant, in contrast to OpInf where the reduced model depends on the chosen snapshots. Recent work has recognized the advantages of non-intrusive reduction methods when it comes to exploiting the power of variable transformations as discussed above. In particular, the Lift & Learn method of [2] combines lifting of a nonlinear PDE with data-driven learning of the ROM via the OpInf method of [3], so that variable transformations are applied only to snapshot data and not to the high-fidelity PDE solver itself.
For several classes of nonlinear PDEs, the particular form of the nonlinear terms means that lifting will lead to a system of PDAEs. For example, this is the case for the Arrhenius reaction terms in the tubular reactor example of [1]. It is also the case for the nonlinear thermodynamic dependencies in the solidification model considered in this paper. It is well known that reduction of PDAEs, and, in the semi-discrete case, DAEs, is challenging and that the algebraic equations require special treatment [35], [36]. In some applications the DAEs can be reformulated as a system of differential equations and model reduction techniques are applied to the index-reduced ODEs [37], but these approaches often lead to stiff systems [38]. Another approach taken in the literature is index-aware model reduction in which the nonlinear DAE is linearized about a stationary solution, the linearized DAE is decoupled into the differential and algebraic parts, and model reduction is applied to each part individually [38], [39]. These existing DAE model reduction approaches are intrusive; here we formulate a non-intrusive data-driven approach. When the algebraic equations arise through the lifting process, they take on a particular structured form. In this paper we elicit that structured form and we exploit it to learn the resulting ROM via non-intrusive operator inference. In particular, we show that the reduction of the lifted algebraic equations can be computed explicitly using only manipulations of the low-dimensional basis vectors, while the reduction of the differential equations follows the OpInf approach of [3].
Section 2 of this paper presents the lifting of a nonlinear system of PDEs to polynomial form and discusses the form of the algebraic equations that arise. Section 3 develops the proposed non-intrusive operator inference approach for the lifted system of DAEs. Section 4 presents application of the approach to solidification of a pure metal. Finally, concluding remarks are presented in Section 5.
Section snippets
Nonlinear model reduction via lifting to polynomial form
This section first presents the general projection-based reduction of a nonlinear system and discusses its computational challenges. We then discuss lifting of nonlinear systems to systems with polynomial terms with particular attention to the differential–algebraic structure that arises for several classes of nonlinear equations. We derive the form of the ROM of the lifted DAE system.
Operator inference for lifted differential–algebraic equations
This section presents our approach for learning the operators of the ROM defined by (6), (7). The differential equations (6) use the standard OpInf approach from [3], as discussed in Section 3.1. In Section 3.2, we develop a tailored approach for the algebraic equations (7), which exploits the particular structure of the lifted system.
Application: Solidification process in additive manufacturing
We test the efficacy of the proposed method for a solidification process in metal additive manufacturing. The additive manufacturing process takes place over a wide range of length scales and time scales, and it is sensitive to variations in process parameters. Numerical simulations of the additive manufacturing process are computationally expensive, making it a challenge to achieve control, optimization, and uncertainty quantification. Reduced-order modeling is thus a critical enabler for
Concluding remarks
This paper has proposed and demonstrated a non-intrusive data-driven model reduction method that addresses the PDAE structure arising in lifting nonlinear systems to polynomial form. The approach provides a new alternative for model reduction of highly nonlinear systems for which more classical hyper-reduction techniques may be ineffective. The approach is effective for the studied solidification problem; however, the relatively slow decay of the POD singular values points to the inefficiencies
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
This work has been supported in part by the U.S. Department of Energy AEOLUS MMICC center under award DE-SC0019303, program manager W. Spotz. The authors acknowledge helpful discussions with Y. Bao, G. Biros, S. DeWitt and B. Radhakrishnan in creating the solidification model implementation.
References (64)
- et al.
Lift & Learn: Physics-informed machine learning for large-scale nonlinear dynamical systems
Physica D
(2020) - et al.
Data-driven operator inference for nonintrusive projection-based model reduction
Comput. Methods Appl. Mech. Engrg.
(2016) - et al.
An ‘empirical interpolation’ method: application to efficient reduced-basis discretization of partial differential equations
C. R. Math.
(2004) - et al.
Recasting nonlinear differential equations as S-systems: A canonical nonlinear form
Math. Biosci.
(1987) - et al.
Linear predictors for nonlinear dynamical systems: Koopman operator meets model predictive control
Automatica
(2018) - et al.
Operator inference for non-intrusive model reduction of systems with non-polynomial nonlinear terms
Comput. Methods Appl. Mech. Engrg.
(2020) - et al.
Modeling and control of physical processes using proper orthogonal decomposition
Math. Comput. Modelling
(2001) - et al.
A fully adaptive nonintrusive reduced-order modelling approach for parametrized time-dependent problems
Comput. Methods Appl. Mech. Engrg.
(2021) - et al.
Projection-based model reduction: Formulations for physics-based machine learning
Comput. & Fluids
(2019) - et al.
Non-linear model reduction for the Navier–Stokes equations using residual DEIM method
J. Comput. Phys.
(2014)
Modeling and numerical simulations of dendritic crystal growth
Physica D
Nonlinear model order reduction via lifting transformations and proper orthogonal decomposition
AIAA J.
Phase field simulations of microstructure evolution in IN718 using a surrogate Ni–Fe–Nb alloy during laser powder bed fusion
Metals
A POD–EIM reduced two-scale model for crystal growth
Adv. Comput. Math.
Approximation of Large-Scale Dynamical Systems
A survey of projection-based model reduction methods for parametric dynamical systems
SIAM Rev.
Model Reduction and Approximation: Theory and Algorithms
Model Reduction of Parametrized Systems
Nonlinear model reduction via discrete empirical interpolation
SIAM J. Sci. Comput.
Universal formats for nonlinear ordinary differential systems
J. Math. Phys.
Analysis of fluid flows via spectral properties of the Koopman operator
Annu. Rev. Fluid Mech.
QLMOR: A projection-based nonlinear model order reduction approach using quadratic-linear representation of nonlinear systems
IEEE Trans. Comput.-Aided Des. Integr. Circuits Syst.
Two-sided projection methods for nonlinear model order reduction
SIAM J. Sci. Comput.
Learning physics-based models from data: perspectives from inverse problems and model reduction
Acta Numer.
Dynamic mode decomposition of numerical and experimental data
J. Fluid Mech.
Compressed sensing and dynamic mode decomposition
J. Comput. Dyn.
Dynamic Mode Decomposition: Data-Driven Modeling of Complex Systems
Ergodic theory, dynamic mode decomposition, and computation of spectral properties of the koopman operator
SIAM J. Appl. Dyn. Syst.
Toward fitting structured nonlinear systems by means of dynamic mode decomposition
Reduced-order modeling of parameterized PDEs using time–space-parameter principal component analysis
Internat. J. Numer. Methods Engrg.
Greedy nonintrusive reduced order model for fluid dynamics
AIAA J.
Data-driven parametrized model reduction in the Loewner framework
SIAM J. Sci. Comput.
Cited by (17)
Preserving Lagrangian structure in data-driven reduced-order modeling of large-scale dynamical systems
2024, Physica D: Nonlinear PhenomenaLagrangian operator inference enhanced with structure-preserving machine learning for nonintrusive model reduction of mechanical systems
2024, Computer Methods in Applied Mechanics and EngineeringCanonical and noncanonical Hamiltonian operator inference
2023, Computer Methods in Applied Mechanics and EngineeringgLaSDI: Parametric physics-informed greedy latent space dynamics identification
2023, Journal of Computational PhysicsOperator inference with roll outs for learning reduced models from scarce and low-quality data
2023, Computers and Mathematics with ApplicationsMelting behavior prediction of latent heat storage materials: A multi-pronged solution
2023, Journal of Energy Storage