PDE-constrained optimization for virtual sensing in structural dynamics: Full-field displacement and force recovery from sparse sensors

Minjae Kim, Jaehwan Jeong, Jaemin Kim · 2026

Read PDE-constrained optimization for virtual sensing in structural dynamics: Full-field displacement and force recovery from sparse sensors by Minjae Kim, Jaehwan Jeong, Jaemin Kim (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-29349.txt.

Introduction

Kim and two colleagues reconstruct the full vibration of a structure from only a few sensors. Virtual sensing matters for structural health monitoring, where knowing displacement and force everywhere, from sparse measurements, is a basic challenge. They formulate the task as optimisation constrained by the elastodynamic equation, the physics law of elastic motion: the unknown force distribution is the variable, and full-field displacement and force are recovered together.

Gradients come from reverse-mode automatic differentiation through the physics solver, and an L-BFGS optimiser iterates to the best state. The pipeline pairs an offline finite element stage with an online GPU stage. On three test cases of rising complexity, a cantilever plate, an elbow pipe and a reactor pressure vessel, the method beat modal expansion everywhere. It cut displacement errors by factors of two to seventeen. GPU acceleration gave up to a sixty-five fold speedup over CPU.

Catalog record

CREATORMinjae Kim, Jaehwan Jeong, Jaemin Kim
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE100 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.29349

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

PDE-constrained optimization for virtual sensing in structural dynamics: Full-field displacement and force recovery from sparse sensors Minjae Kim, Jaehwan Jeong, Jaemin Kim Abstract Virtual sensing — recovering full-field structural response from sparse sensor measurements — is a fundamental challenge in structural health monitoring. This study formulates virtual sensing as a PDE-constrained optimization (PDE-CO) problem, where the governing elastodynamic equation serves as an equality constraint, the applied force distribution is the optimization variable, and the full-field displacement and force are jointly recovered. Gradients of the Tikhonov-regularized objective are computed via reverse-mode automatic differentiation through the forward PDE solver, and L-BFGS iteratively finds the optimal state. The framework couples an offline FEniCSx stage for finite element assembly with an online GPU-accelerated JAX stage, and is verified on three examples of increasing complexity: a cantilever plate, a 90 ∘ 90^{\circ} elbow pipe, and a reactor pressure vessel (RPV) representative of a 300 MW pressurized nuclear reactor. PDE-CO consistently outperforms modal expansion across all three cases, reducing displacement errors by factors of 2 2 to 17 × 17\times with sub-percent accuracy on every example. Unlike modal expansion, where force is derived by back-calculation from truncated modal coordinates and is not jointly optimized, PDE-CO recovers displacement and force simultaneously through the PDE constraint; the increased computational cost is offset by GPU acceleration delivering up to 64.8 × 64.8\times speedup over CPU. Nomenclature Symbol Definition Symbol Definition \mathbf{K} Global stiffness matrix \mathbf{M} Global consistent mass matrix \mathbf{D}(\omega) Dynamic stiffness matrix; \mathbf{K}-\omega^{2}\mathbf{M} \mathbf{D}_{ff} Free-DOF block of \mathbf{D} \mathbf{P} Force mapping matrix \mathbf{S}_{f} Sensor selection matrix (Boolean) \mathbf{A} Forward operator; -\omega^{2}\mathbf{S}_{f}\mathbf{D}_{ff}^{-1}\mathbf{P} \bm{\Phi} Modal matrix (retained eigenmodes) \bm{\theta} Unknown force parameter vector \mathbf{q} Modal coordinate vector \hat{\mathbf{U}}_{f} Free-DOF displacement amplitude \hat{\mathbf{a}}_{\mathrm{obs}} Observed sensor acceleration \bm{\varepsilon} Small strain tensor \bm{\sigma} Cauchy stress tensor J(\bm{\theta}) Objective function \alpha Tikhonov regularization weight \omega Excitation circular frequency \omega_{i} i -th natural frequency E Young’s modulus \nu Poisson’s ratio \rho Mass density \mathbb{C} Elasticity tensor N Total DOFs N_{f} Number of free DOFs N_{\theta} Number of force parameters N_{s} Number of sensor DOFs N_{d} Number of retained modes \Omega Physical domain \Gamma_{D} Dirichlet boundary \Gamma_{N} Neumann boundary 1 Introduction Virtual sensing — the recovery of full-field displacement, strain, and acceleration fields across an entire structure from sparse sensor measurements — is a key capability for structural health monitoring (SHM) ( 23 ; 17 ) , fatigue assessment, and damage detection ( 8 ) (see Figure 1 ). The structural dynamics community has developed several approaches for virtual sensing, including modal least-squares regression ( 16 ) , Kalman filter methods ( 20 ; 18 ) , reduced-order modeling approaches ( 22 ; 13 ; 12 ) , and physics-guided machine learning approaches ( 10 ) . Among these, modal expansion is the most widely adopted: the displacement is represented as a linear combination of the N_{d} lowest-frequency (dominant) eigenmodes, and the modal coordinates are fitted to sensor data via least-squares. However, modal expansion suffers from fundamental limitations. First, retaining only N_{d} modes introduces modal truncation error, and N_{d} is bounded by the sensor count ( N_{d}\leq N_{s}-1 for an overdetermined system), so the sparse sensor data cannot constrain a higher-order modal basis. Second, the governing PDE is never enforced: the sensor data determines only N_{s} degrees of freedom, and no physical law propagates this information to the remaining N_{f}-N_{s} DOFs. When N_{s}\ll N_{f} , this purely kinematic approach lacks the information to reconstruct the response accurately away from sensor locations. The scientific computing community, in parallel, has developed differentiable simulation frameworks ( 26 ; 3 ) that formulate inverse problems as PDE-constrained optimization (PDE-CO), with automatic differentiation handling all sensitivity computations. PDE-CO treats the governing equation as an equality constraint, ensuring the recovered solution is physically consistent at every optimization iteration. By enforcing the PDE, sparse sensor information is propagated to the entire domain through the physics, filling the information gap that modal expansion cannot bridge. The PDE-CO framework offers several key advantages over modal expansion for virtual sensing. First, it eliminates modal truncation error. PDE-CO uses the full dynamic stiffness matrix \mathbf{D}_{ff}\in\mathbb{R}^{N_{f}\times N_{f}} rather than a truncated modal basis \bm{\Phi}\in\mathbb{R}^{N_{f}\times N_{d}} , so the solution space is the entire N_{f} -dimensional FE space. Second, the PDE constraint provides physics-based regularization. Every trial solution is consistent with the governing equation, and the inverse operator \mathbf{D}_{ff}^{-1} acts as a smoothing operator that attenuates high-frequency spatial components. Third, gradients are computed via reverse-mode AD and supplied to L-BFGS, a unified paradigm that extends identically to nonlinear forward models, material identification, and topology optimization. Fourth, the full-field displacement emerges directly from the primal PDE solve at the optimal force, so both the displacement field and the force distribution are recovered simultaneously within a single optimization framework without auxiliary post-processing steps. Figure 1 : Schematic of virtual sensing via physics-based digital twin. Left (physical twin): a reactor pressure vessel instrumented with sparse accelerometers (red dots). Right (digital twin): a physics-based finite element model reconstructs the full-field displacement distribution over the entire domain from the sparse sensor measurements. Despite their conceptual similarity, the two threads — modal-based virtual sensing and PDE-constrained optimization — have not converged for full-field inference in structural dynamics. The adjoint-based PDE-CO framework has been applied to static inverse problems such as elasticity imaging ( 21 ) , structural weakness identification ( 14 ) , and constitutive inverse design ( 1 ) . 26 developed JAX-FEM, a GPU-accelerated differentiable FE solver that computes the adjoint gradient via reverse-mode AD through the forward solver and performs optimization with L-BFGS, demonstrating PDE-CO for Poisson and static elasticity problems. In structural dynamics, 7 employed the adjoint method for sizing optimization under transient dynamic loads, but this is a design optimization problem rather than an inverse sensing problem. To the best of the authors’ knowledge, no prior work has applied PDE-CO to recover full-field displacement from sparse vibration sensors in structural dynamics. Existing virtual sensing methods rely on modal expansion ( 11 ) , Kalman filters ( 18 ) , physics-guided machine learning ( 10 ) , or transfer function inversion, but do not exploit the full PDE constraint in a differentiable optimization framework. This study addresses this gap through the following contributions. Virtual sensing under harmonic excitation is formulated as a PDE-CO problem, where the dynamic stiffness equation \mathbf{D}(\omega)\hat{\mathbf{U}}_{f}=\mathbf{P}\bm{\theta} serves as the equality constraint and the force parameter \bm{\theta}\in\mathbb{R}^{N_{\theta}} is the optimization variable. The force on the candidate loaded boundary \Gamma_{N} is parameterized via a Boolean selection matrix \mathbf{P} , and the gradient is computed via JAX reverse-mode AD; L-BFGS then iteratively recovers the optimal force distribution and the corresponding full-field displacement. It is demonstrated analytically and numerically that PDE-CO achieves superior displacement recovery compared to modal expansion, because the PDE constraint eliminates modal truncation error and propagates sensor information through the exact governing equation. The framework is verified on three examples of increasing complexity — a cantilever plate ( \epsilon_{u}=0.45\% , 4 sensors), a 90^{\circ} elbow pipe ( \epsilon_{u}=0.28\% , 8 sensors), and a double-loop cylindrical RPV representative of a 300 MW PWR ( \epsilon_{u}=0.885\% , 36 sensors) — consistently outperforming modal expansion ( \epsilon_{u} : 7.69\% , 1.11\% , 1.86\% ), with GPU acceleration on an NVIDIA RTX

Read the complete text