Reliable Real-Time Solution of Parametrized Partial Differential Equations: Reduced-Basis Output Bound Methods
Summary (2 min read)
1 Introduction
- The optimization, control, and characterization of an engineering component or system requires the prediction of certain ‘‘quantities of interest,’’ or performance metrics, which the authors shall denote outputs—for example deflections, maximum stresses, maximum temperatures, heat transfer rates, flowrates, or lift and drags.
- The parameters, which the authors shall denote inputs, serve to identify a particular ‘‘configuration’’of the component: these inputs may represent design or decision variables, such as geometry—for example, in optimization studies; control variables, such as actuator power—for example in real-time applications; or characterization variables, such as physical properties—for example in inverse problems.
- The authors thus arrive at an implicit input-output relationship, evaluation of which demands solution of the underlying partial differential equation.
- The authors goal is the development of computational methods that permit rapid and reliable evaluation of this partial-differentialequation-induced input-output relationship in the limit of many queries—that is, in the design, optimization, control, and characterization contexts.
- In Section 3 the authors describe, for coercive symmetric problems and ‘‘compliant’’ outputs, the reduced-basis approximation; and in Section 4 they present the associated a posteriori error estimation procedures.
3 Reduced-Basis Approach
- Fortunately, the convergence rate is not too sensitive to point selection: the theory only requires a log ‘‘on the average’’ distribution ~Maday et al. @19#!; and, in practice, l̄ need not be a sharp upper bound.
- In the off-line stage, the authors compute the u(mn) and form the AI N q and FI N : this requires N ~expensive! ‘‘a’’ finite element solutions and O(QN2) finite-element-vector inner products.
4 A Posteriori Error Estimation: Output Bounds
- From Section 3 the authors know that, in theory, they can obtain sN(m) very inexpensively: the on-line computational effort scales as O(2/3 N3)1O(QN2); and N can, in theory, be chosen quite small.
- The second set of disadvantages relates to the computational expense— the O(Q) off-line and the O(Q2) on-line scaling induced by ~24! and ~25!, respectively.
5. Extensions
- To wit, for the primal problem, the authors find uN(m)PWN such that a(uN(m),v;m) 5 f (v), ;vPWN ; and for the adjoint problem, they define ~though, unless otherwise indicated, do not compute!.
- Will only involve the symmetric part of a.
- In particular, well-posedness is now ensured only by the inf-sup condition: there exists positive b0 , b~m!, such that 0,b0<b~m!5 inf wPX sup vPX a~w ,v;m! iwiXiviX , ;mPD. (58) Two numerical difficulties arise due to this ‘‘weaker’’ stability condition.
Acknowledgments
- The authors would like to thank Mr. Thomas Leurent ~formerly! of MIT for his many contributions to the work described in this paper; thanks also to Shidrati Ali of the Singapore-MIT Alliance and Yuri Solodukhov of MIT for very helpful discussions.
- The authors would also like to acknowledge their longstanding collaborations with Professor Jaime Peraire of MIT and Professor Einar Rønquist of the Norwegian University of Science and Technology.
Did you find this useful? Give us your feedback
Citations
1,695 citations
Cites background from "Reliable Real-Time Solution of Para..."
...Its success is limited to the problems of linear elliptic parabolic PDEs with affine parameters or low-order polynomial nonlinearities [32, 25, 26, 41, 29]....
[...]
1,265 citations
1,230 citations
1,090 citations
Cites background or methods from "Reliable Real-Time Solution of Para..."
...The SCM is more efficient and general than earlier proposals [142,121,97]; also the SCM is much more easily implemented [64]....
[...]
...Much current effort is thus devoted to development of (i) a posteriori error estimation procedures and in particular rigorous error bounds for outputs of interest [121], and (ii) effective sampling strategies in particular for higher (than one) dimensional parameter domains [33,32,97,138,153]....
[...]
...In the context of affine parameter dependence, in which the operator is expressible as the sum of Q products of parameter-dependent functions and parameterindependent operators, the Offline-Online idea is quite self-apparent and indeed has been re-invented often [15, 66,70,114]; however, application of the concept to a posteriori error estimation — note the Online complexity of both the output and the output error bound calculation must be independent of N — is more involved and more recent [64,121,122]....
[...]
...In Section 9 we present rigorous and relatively sharp a posteriori output error bounds [3,23,108] for RB approximations [121,142]; in Section 10 we develop the coercivity-constant lower bounds [64] required by the a posteriori error estimation procedures....
[...]
...2 for non-compliant problems we discuss (primal-dual [117]) RB Galerkin projection [121] and optimality; in Section 7....
[...]
831 citations
References
2,089 citations
927 citations
428 citations
414 citations
Related Papers (5)
Frequently Asked Questions (9)
Q2. What is the problem with the inf-sup parameter?
Loss of stability can, in turn, lead to poor approximations—the inf-sup parameter enters in the denominator of the a priori convergence result.
Q3. What are the advantages of a nonintegrated approach?
there are significant computational and conditioning advantages associated with a ‘‘nonintegrated’’ approach, in which the authors introduce separate primal (u(mn)) and dual (c(mn)) approximation spaces for u(m) and c~m!, respectively.
Q4. What is the second numerical difficulty of the inf-sup parameter?
The second numerical difficulty is estimation of the inf-sup parameter, which for noncoercive problems plays the role of g(m) in Method The authora posteriori error estimation techniques.
Q5. What is the effectivity of the two?
If the primal and dual errors are a-orthogonal, or become increasingly orthogonal as N increases, then the effectivity will not, in fact, be bounded as N→` .
Q6. What is the on-line procedure for calculating the bounds for a given new ?
In the on-line stage, for any given new m, the authors first form AI N(m), FI N and AI 2N(m), FI 2N , then solve for uI N(m) and uI 2N(m), and finally evaluate sN ,2N6 (m): this requires O(4QN2)1O(16/3 N3) operations and O(4QN2) storage.
Q7. What is the method to estimate error?
Note that WN has good approximation properties both for the first and second lowest eigenfunctions, and hence eigenvalues; this is required by the Method The authorerror estimator to be presented below.
Q8. What is the cost of evaluating sN(m)?
In the on-line stage, for any given new m, the authors first form AI N from ~15!, then solve ~14! for uI N(m), and finally evaluate sN(m)5FI NT uI N(m): this requires O(QN2)1O(2/3 N3) operations and O(QN2) storage.
Q9. What is the new ingredient in the reduced-basis approximation and error estimator?
The essential new ingredient is the presence of the time variable, t.The reduced-basis approximation and error estimator procedures are similar to those for noncompliant nonsymmetric problems, except that the authors now include the time variable as an additional parameter.