Source-linked AI summary
Progressive construction of a parametric reduced-order model for PDE-constrained optimization
Matthew J. Zahr, Charbel Farhat
TL;DR
PDE-constrained optimization is costly because high-dimensional PDE solutions and sensitivities are repeatedly required, while fixed ROM bases may lack robustness across parameter variations. The paper introduces a progressively trained ROM framework with sensitivity-aware bases and a residual-based trust region. In aerodynamic shape optimization, it reduces HDM queries by a factor of 4 and recovers the optimum to machine precision, while hyperreduction remains necessary for CPU speedups.
Problem
Repeated high-dimensional PDE and sensitivity solves make PDE-constrained optimization expensive, and a single static ROM basis may be inaccurate across nonlinear parameter variations.
Method
The framework progressively trains ROMs from previous optimization solutions, uses state and sensitivity snapshots, and restricts iterates with a residual-based nonlinear trust region.
Results
A factor of 4 fewer HDM queries was observed in aerodynamic shape optimization, where the optimal solution was recovered to machine precision.
Takeaways & Limitations
Progressive ROM optimization avoids constructing a single globally accurate basis and avoids sampling parameter-space regions not visited during optimization.
Takeaways & Limitations
The presented method retains high-dimensional operations and requires hyperreduction to translate its query reduction into a CPU-time advantage.
Abstract
from arXiv · showhide
An adaptive approach to using reduced-order models as surrogates in PDE-constrained optimization is introduced that breaks the traditional offline-online framework of model order reduction. A sequence of optimization problems constrained by a given Reduced-Order Model (ROM) is defined with the goal of converging to the solution of a given PDE-constrained optimization problem. For each reduced optimization problem, the constraining ROM is trained from sampling the High-Dimensional Model (HDM) at the solution of some of the previous problems in the sequence. The reduced optimization problems are equipped with a nonlinear trust-region based on a residual error indicator to keep the optimization trajectory in a region of the parameter space where the ROM is accurate. A technique for incorporating sensitivities into a Reduced-Order Basis (ROB) is also presented, along with a methodology for computing sensitivities of the reduced-order model that minimizes the distance to the corresponding HDM sensitivity, in a suitable norm. The proposed reduced optimization framework is applied to subsonic aerodynamic shape optimization and shown to reduce the number of queries to the HDM by a factor of 4-5, compared to the optimization problem solved using only the HDM, with errors in the optimal solution far less than 0.1%.
1. INTRODUCTION
PDE-constrained optimization is expensive because each design-cycle parameter value requires high-fidelity PDE solutions and sensitivities. The paper therefore develops a progressively trained ROM surrogate that replaces a fixed offline-online workflow with adaptive sampling during optimization.
- Industry-scale PDE-constrained optimization requires many expensive PDE solutions and, for gradient-based methods, sensitivities.
- Traditional offline-online model reduction is artificial for many-query design optimization because total wall-time, rather than online time alone, determines efficiency.
- A single static low-dimensional basis may lose accuracy across nonlinear parameter variations and an optimization trajectory that is unknown in advance.
- Adaptive ROM approaches solve a sequence of reduced optimization problems and use their solutions to enrich the basis for subsequent problems.
- The proposed framework trains each ROM from prior optimization solutions, uses a residual-based nonlinear trust region, and incorporates state and sensitivity snapshots.
- The method is formulated within the NAND discretize-then-differentiate setting but is intended to generalize to other PDE-constrained optimization frameworks and basis-construction methods.
2. PDE-CONSTRAINED OPTIMIZATION
PDE-constrained optimization treats the PDE state as implicitly determined by parameters, so every parameter change requires a state solve. The section describes discretization, sensitivity computation, and the direct–adjoint trade-off for objective and constraint gradients.
- The continuous formulation defines a state space, parameter domain, objective, additional constraints, and nonlinear differential operator before discretization.
- The paper uses discretize-then-differentiate, in which the PDE is discretized before deriving sensitivity equations.
- NAND treats the PDE state as an implicit function of the parameters and enforces the discrete PDE at every optimization iteration.
- State sensitivities with respect to PDE parameters are used to compute objective and constraint gradients.
- The direct method solves n_p linear systems with the PDE Jacobian, whereas the adjoint method solves one transposed-Jacobian system per functional.The direct method is preferred when there are many functionals relative to parameters; the adjoint method has the opposite trade-off.
3. MODEL REDUCTION
Model reduction approximates high-dimensional PDE states in a low-dimensional affine subspace and projects the governing equations onto reduced bases. The paper emphasizes POD-based bases containing both state and sensitivity information, while noting that nonlinear speedups require hyperreduction.
- Model order reduction seeks storage and CPU savings by reducing the degrees of freedom and nonlinear equations of the high-dimensional model.
- The reduced approximation assumes the state lies approximately in a low-dimensional affine subspace, with sensitivities represented in the same subspace.
- Projection-based ROMs substitute the affine approximation into the high-dimensional model and enforce the nonlinear equations against a left reduced basis.
- Galerkin and LSPG projections provide alternative reduced formulations, with LSPG associated with a minimum-residual property.The minimum-residual property is described as preventing added basis information from degrading the solution approximation under stated assumptions.
- The derivation is for projection-based ROMs, while hyperreduction extensions are deferred because nonlinear high-dimensional operations remain in the computation.
- POD and the method of snapshots compress high-dimensional simulation data into a reduced-order basis from samples at parameter configurations.
- The proposed basis construction applies POD separately to state and sensitivity snapshots before combining them, avoiding heuristic weighting and allowing separate dimension control.
- Minimum-residual state and minimum-error sensitivity formulations are used so that adding sensitivity information does not degrade state approximation and adding state information does not adversely affect sensitivities.
4. OPTIMIZATION USING A PROGRESSIVE ROM
The paper develops a progressive ROM-constrained optimization framework with residual-based trust regions, sensitivity-aware reduced models, and basis updates for approximating PDE-constrained optimization.
- Progressive ROM framework: The framework uses progressively constructed ROMs as surrogates for the HDM and formulates a sequence of ROM-constrained optimization problems.The reduced problems are used to approximate the original PDE-constrained optimization problem.
- Scope and limitations: The method prioritizes demonstrating ROB-based accuracy; large-dimensional quantities still appear in reduced computations because hyperreduction is reserved for future work.The paper identifies hyperreduction as the step needed to demonstrate the intended speedup.
- Sensitivity computation: Reduced sensitivities can be approximated by minimizing their distance to corresponding HDM sensitivities in a selected positive-definite norm.This minimum-error formulation is developed alongside gradient computations for Galerkin and LSPG ROMs.
- Sensitivity computation: The minimum-error sensitivity formulation approaches the true reduced sensitivity where the ROM residual is small, especially near HDM sample points.The LSPG sensitivity formulation obtains this property by dropping second-order terms, but gradient consistency can be problematic when the ROM is inaccurate.
- Sensitivity computation: When sensitivity-basis truncation is skipped, reconstructed reduced sensitivities exactly match HDM sensitivities at training points.This follows because the desired sensitivities remain in the span of the ROB and the minimum-error formulation is used.
- Progressive ROM framework: The progressive procedure uses a residual-based nonlinear trust region and updates the ROB with new HDM snapshots, including low-rank SVD updates to reduce basis-update cost.Sampling a poor point while starting the next problem from the best point is intended to reduce the risk that an aggressive trust-region tolerance harms convergence.
5. APPLICATIONS
The study applies the progressively constructed ROM framework to subsonic inverse design, recovering the Cub-RAE2822 airfoil from NACA0012 using pressure-discrepancy minimization. The ROM approach matches the target shape and pressure distribution while substantially reducing HDM queries.
- Problem setup: The aerodynamic test recovers the Cub-RAE2822 airfoil from the NACA0012 airfoil by minimizing subsonic pressure-distribution discrepancy.The flow conditions are Mach 0.5 and zero angle of attack.
- Shape parametrization: The shape is parameterized with SDESIGN cubic design elements, whose control nodes deform the airfoil surface and surrounding body-fitted CFD mesh.The control nodes move vertically, and a structural-analogy mesh-motion algorithm propagates the deformation through the CFD mesh.
- ROM construction: Each HDM sample contributes 9 snapshots: one steady-state flow solution and 8 sensitivities for the 8 shape parameters.These snapshots are used to construct the reduced-order basis.
- HDM reference: The HDM reference optimization reduces the initial objective by 9 orders of magnitude before numerical difficulties terminate the run.The reference solution requires 24 optimization iterations and 29 HDM queries.
- ROM results: Using 7 HDM samples, the progressive ROM reduces the initial pressure discrepancy by 18 orders of magnitude to essentially machine zero, requiring 4 times fewer HDM queries than HDM optimization.The final shape and pressure coefficient match the target closely.
- Accuracy control: The residual-based nonlinear trust region controls ROM accuracy, while the objective function is mostly sufficient to keep this inverse-design trajectory near accurate regions.The trust-region bound is reached only once in this example; other objectives such as drag or downforce may require it more strongly.
6. CONCLUSIONS
The paper proposes a progressive ROM-based optimization framework that updates its basis during optimization and avoids constructing one globally accurate basis. It also introduces minimum-error reduced sensitivities and demonstrates fewer HDM queries in aerodynamic shape optimization.
- Progressive ROM optimization: The adaptive framework breaks the traditional offline-online model-reduction workflow by repeatedly updating the ROB from HDM samples at reduced optimization solutions.Each updated ROB defines the next reduced optimization problem.
- Reduced sensitivities: The minimum-error sensitivity framework returns the best HDM-sensitivity approximation available in the given ROB under a chosen norm.For Least-Squares Petrov-Galerkin projection, reduced sensitivities approach the true reduced sensitivities as ROM error decreases.
- Computational outcome: A factor of 4 fewer HDM queries was observed in aerodynamic shape optimization, while the optimal solution was recovered to machine precision.The approach avoids sampling unvisited parameter regions and avoids constructing a globally accurate ROB.
A. LOW-RANK SVD UPDATE ALGORITHMS
The low-rank SVD update algorithm appends new vectors to an existing factorization and computes the updated decomposition through QR and SVD operations.
- Algorithm inputs and output: Brand’s algorithm updates a rank-r data matrix factorization by appending a vector or block of vectors.The input includes X = UΣV^T and an appended matrix Y with k columns.
- Algorithm inputs and output: The updated factorization is obtained as the SVD of the enlarged data matrix.The algorithm outputs the SVD after incorporating the appended data matrix.
- Update steps: The update computes a QR decomposition of P̄ = PR_A before forming the smaller matrix K.
- Update steps: The final reduced computation takes the SVD of K = CSD^T, whose dimension is (r+k) × (r+k).