Adaptive Reduced Basis Trust Region Methods Accelerate Complex Parameter Identification

Researchers Michael Kartmann, Tim Keil, Mario Ohlberger, Stefan Volkwein, and Barbara Kaltenbacher have introduced an adaptive reduced basis trust region method to solve inverse parameter identification problems in elliptic partial differential equations, addressing severe computational bottlenecks caused by infinite-dimensional parameter spaces. In physical and mathematical modeling, forward simulations require exact coefficients, yet unmeasurable distributed quantities must routinely be determined from indirect, noisy measurements. This process generates ill-posed inverse problems where reconstructed parameters do not depend continuously on observations, rendering direct inversion unstable. Although iterative schemes like the iteratively regularized Gauß-Newton method stabilize solutions via Tikhonov penalties, they demand numerous costly forward numerical solves. Standard model order reduction techniques rely on offline greedy sampling to construct reduced-order models, but this offline procedure collapses when parameter dimensions are infinite or exceptionally large. The researchers solved this problem by adaptively constructing a reduced parameter space during the online optimization process while simultaneously enforcing certified state space model order reduction.

The mathematical formulation targets the identification of an unknown distributed parameter from noisy observations of an exact state governed by a linear elliptic partial differential equation. The governing bilinear form is continuous, coercive, and linear with respect to the parameter, ensuring unique forward solutions through the Lax-Milgram theorem. Because the forward observation operator is continuously Fréchet-differentiable but not continuously invertible, the inverse problem remains severely ill-posed. To achieve stable parameter recovery, the authors formulate the task as minimizing a discrepancy functional between model predictions and noisy measurements within an admissible parameter set. The iteratively regularized Gauß-Newton method addresses this minimization by sequentially linearizing the forward operator around current parameter estimates and solving a regularized quadratic subproblem penalized by a Tikhonov functional. Regularization parameters are calibrated at each step through a backtracking strategy satisfying linearized discrepancy bounds, while overall iteration halts under Morozov’s discrepancy principle to prevent noise amplification from corrupting the reconstructed coefficient fields.

Applying classical reduced basis reduction to this iterative setting reveals severe computational roadblocks rooted in parameter dimensionality. When an infinite-dimensional parameter space is discretized with standard piecewise bilinear finite element basis functions, the parameter-affine expansion of the differential operator breaks into tens of thousands of discrete affine components. In traditional reduced basis frameworks, this high dimensionality induces two prohibitive computational penalties: projecting each affine component onto reduced subspaces becomes exceedingly cumbersome, and assembling residual-based error estimators becomes computationally infeasible because each reduced basis vector requires calculating full-order Riesz representatives across every affine component. To eliminate this dual barrier, the authors established a structured two-stage reduction methodology. The first stage adaptively compresses the infinite-dimensional parameter manifold into a low-dimensional subspace during the online iterative search. The second stage applies certified reduced basis approximations to both primal and dual state spaces directly on top of the reduced parameter manifold, eliminating the combinatorial proliferation of affine operator terms.

The foundational parameter reduction stage exploits the first-order optimality conditions of the linearized Tikhonov subproblem, which demonstrate that optimal parameter updates align naturally with the dual observation operator acting on the linearized adjoint state. Building on this structural property, the algorithm adaptively enriches the reduced parameter space using full-order model gradients of the discrepancy functional evaluated at successive iterates. Because the gradient represents the steepest descent direction, incorporating it guarantees sufficient objective decrease, while successive gradients accumulated across optimization steps capture approximate curvature information analogous to Quasi-Newton methods. At each outer iteration, the method solves a low-dimensional minimization problem restricted to the active reduced parameter space using an inner iteratively regularized Gauß-Newton solver. The newly obtained gradient is subsequently incorporated into the reduced parameter space through Gram-Schmidt orthonormalization. By substituting thousands of finite element unknowns with a small set of basis functions, the number of affine components governing state equations contracts dramatically.

Once the parameter manifold is restricted to a compact representation, standard reduced basis techniques efficiently project both primal and adjoint state spaces onto low-dimensional subspaces. To rigorously certify state approximations without executing full-order solves at each step, the authors derived an a posteriori error bound for the discrepancy functional based on evaluated primal and dual residual norms divided by the coercivity constant. This error estimator is embedded into an adaptive error-aware trust region framework designed specifically for iterative regularization. In every outer iteration, the algorithm calculates an approximated generalized Cauchy point via a Landweber-type gradient step satisfying an Armijo sufficient decrease condition and a trust region radius constraint. This Cauchy point provides an initial guess for solving a parameter- and state-reduced trust region subproblem. The inner solver executes regularized Newton steps constrained by a relative error bound until meeting either a reduced discrepancy stopping criterion or the boundary of the trust region.

After the inner subproblem generates a candidate parameter trial step, the algorithm determines whether to accept or reject the iterate using computationally inexpensive surrogate conditions derived from the a posteriori error bound. A candidate step is accepted if the reduced objective value plus its certified error bound is strictly less than the objective value at the Cauchy point, whereas it is rejected if the reduced objective minus the bound exceeds that threshold. If these inexpensive bounds prove inconclusive, the algorithm evaluates the full-order discrepancy functional directly. When a step is accepted, the trust region radius expands if the ratio of actual full-order reduction to predicted reduced-order reduction surpasses a prescribed acceptance parameter, while rejected steps retain current parameters and contract the trust region radius. Accepted iterates trigger full-order solves for primal states, adjoint states, and discrepancy gradients, which are orthonormalized to enrich both the reduced parameter and state spaces simultaneously.

A major computational challenge in the combined reduction algorithm involves assembling the a posteriori error estimator, which requires solving full-order systems for Riesz representatives as basis dimensions expand. To prevent the certification process from dominating runtime, the authors introduced an adaptive switching condition that monitors algorithmic effort. The framework tracks the number of full-order solves needed to update Riesz representatives between iterations and compares it directly against the full-order solves required to evaluate residual norms online during Armijo backtracking line searches. When the assembly cost surpasses the online evaluation expense in subsequent iterations, the algorithm halts offline Riesz assembly and computes residual norms directly online. In a dedicated case study, the mixed strategy switched from offline assembly to online residual computation at iteration eight. This automated transition avoided assembling unneeded Riesz representatives once the reduced basis had attained high accuracy, significantly reducing total full-order solves and accelerating overall computation time.

The authors validated their framework through comprehensive numerical experiments implemented using the pyMOR library on a two-dimensional unit square domain discretized with 90,601 piecewise bilinear finite element degrees of freedom under an observational noise level of 10^-5. In the first test scenario, the researchers reconstructed a distributed reaction coefficient within an elliptic partial differential equation featuring homogeneous Dirichlet boundary conditions, benchmarked against full-order and pure parameter-reduced implementations. Pure parameter reduction decreased operator coupling applications by a factor of 100, yet solving full-order state subproblems prevented net runtime reductions. Conversely, the combined parameter and state space reduced trust region method achieved an overall computational speedup factor of approximately 3.5 compared to the full-order method. The combined approach reached the discrepancy stopping threshold with 148 full-order solves, of which 124 served error certification. Reconstructed coefficients closely matched target Gaussian distributions, attaining relative L2 parameter errors of roughly 6 percent using only 9 to 14 reduced parameter basis functions.

Subsequent experiments investigated the reconstruction of distributed diffusion coefficients governed by divergence-form elliptic operators, where regularization within an H1 Sobolev space prevented high-amplitude numerical artifacts observed under L2 penalties. To enhance reconstruction fidelity, the researchers smoothed parameter basis snapshots by applying the inverse discretized Neumann operator to adjoint coupling vectors. In discontinuous diffusion test problems, the combined reduced method decreased full-order partial differential equation solves by a factor of 60 relative to the full-order algorithm, producing a tenfold computational speedup while employing only 29 parameter basis functions. Relative L2 parameter errors remained at approximately 2 percent, with pointwise relative errors staying below 15 percent across most domain regions. Additional test cases incorporating composite targets with Gaussian shapes, hat functions, and discontinuous obstacles demonstrated consistent speedup factors between 3 and 4 for reaction problems and roughly 10 for diffusion problems, confirming the algorithm’s adaptability across contrasting spatial profiles.

The study demonstrates that the combined parameter and state space reduction framework guarantees global convergence with respect to iteration counts, as every accepted step performs at least as effectively as a gradient descent step with Armijo step length. However, the authors noted specific structural limitations regarding parameter recovery near domain boundaries: because Morozov’s discrepancy principle assesses data misfit exclusively within the range of the forward operator, homogeneous Dirichlet boundary conditions suppress boundary state sensitivities, leaving low-dimensional parameter bases less accurate along domain perimeters. Furthermore, while local convergence and regularization properties under vanishing noise levels are anticipated based on established iteratively regularized Gauß-Newton theory, formal regularization proofs for the combined reduction setting remain reserved for future research. Overall, the findings confirm that adaptively coupling discrepancy gradient parameter enrichment with certified error-aware trust region state reduction provides an efficient, rigorous path for resolving large-scale and infinite-dimensional inverse parameter identification problems.

Subject of Research: Model order reduction and adaptive trust region regularization for inverse problems

Article Title: Adaptive reduced basis trust region methods for parameter identification problems

Article References: Kartmann, M., Keil, T., Ohlberger, M., Volkwein, S., & Kaltenbacher, B. (2024). Adaptive reduced basis trust region methods for parameter identification problems. Computational Science and Engineering, 1(1), Article 3. https://doi.org/10.1007/s44207-024-00002-z

Image Credits: AI Generated

DOI: 10.1007/s44207-024-00002-z

Keywords: parameter identification, inverse problems, model order reduction, reduced basis method, trust region methods, iteratively regularized Gauss-Newton, partial differential equations, numerical analysis, Adaptive, reduced, basis, trust