arxiv: 2604.22305 · v1 · submitted 2026-04-24 · 📊 stat.AP

Recognition: unknown

Finite element model updating of building structures under seismic excitation: A parallelized latent space-based Bayesian framework

Kenzo Kodera, Minoru Matsubara, Sangwon Lee, Takeshi Ugata, Taro Yaoyama, Tatsuya Itoi

Pith reviewed 2026-05-08 09:13 UTC · model grok-4.3

classification 📊 stat.AP

keywords finite element model updatingBayesian inferencevariational autoencoderseismic excitationstructural dynamicsuncertainty quantificationsequential Monte CarloGPU parallelization

0 comments

The pith

A latent-space variational autoencoder combined with GPU-parallel sequential Monte Carlo sampling enables efficient Bayesian updating of finite-element building models from seismic response data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a Bayesian method for correcting finite-element models of buildings against real seismic measurements while rigorously tracking parameter uncertainties. High-dimensional time histories or frequency responses are first compressed by a multimodal variational autoencoder into a low-dimensional latent space that makes likelihood evaluation tractable. A population-based sequential Monte Carlo sampler then explores the posterior on GPUs, delivering amortized inference after a single training stage. Validation on both synthetic benchmarks and shaking-table experiments shows accurate recovery of stiffness and damping parameters together with well-calibrated uncertainty intervals, even when the data are sparse enough to produce multimodal posteriors.

Core claim

Projecting high-dimensional structural response data into a low-dimensional latent space via a multimodal variational autoencoder permits tractable likelihood evaluation in the original observation space, while a GPU-parallelized sequential Monte Carlo sampler performs robust posterior sampling without further simulator calls once the surrogate is trained.

What carries the argument

The multimodal variational autoencoder that maps seismic response time histories or frequency response functions into a low-dimensional latent representation, paired with a GPU-accelerated sequential Monte Carlo sampler for amortized Bayesian inference.

If this is right

Structural stiffness and damping parameters are recovered with quantified uncertainties from both numerical and experimental seismic data.
After training, posterior sampling requires no additional finite-element simulator evaluations.
GPU parallelization yields fast inference even when observations are sparse and induce multimodal posteriors.
The framework supports robust inference for building structures under seismic excitation without explicit modeling in the full observation space.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The amortized property could support online model updating as new seismic records arrive in structural health monitoring systems.
Similar latent-space compression might reduce costs for real-time fragility assessment of other civil infrastructure beyond nuclear facilities.
The approach opens a route to hybrid physics-informed surrogates that combine the variational autoencoder with additional sensor modalities.

Load-bearing premise

The latent-space projection must retain enough information from the original high-dimensional responses that the resulting approximate likelihood still yields reliable posterior distributions.

What would settle it

Direct high-dimensional Bayesian updating performed on the same shaking-table dataset without the latent-space compression should produce posterior parameter distributions and uncertainty widths that differ materially from those obtained by the proposed method.

Figures

Figures reproduced from arXiv: 2604.22305 by Kenzo Kodera, Minoru Matsubara, Sangwon Lee, Takeshi Ugata, Taro Yaoyama, Tatsuya Itoi.

**Figure 1.** Figure 1: Schematic of Multimodal VAE architecture adopted in this study. view at source ↗

**Figure 2.** Figure 2: The overview of the proposed framework: LSBI-SMC. view at source ↗

**Figure 3.** Figure 3: The target structure in the benchmarking example: a shear-type 4DOF building model. view at source ↗

**Figure 4.** Figure 4: Visualization of the latent space of the trained MVAE model in the benchmarking example. view at source ↗

**Figure 5.** Figure 5: The distribution of posterior samples in the benchmarking example ( view at source ↗

**Figure 6.** Figure 6: Posterior predictions of FRF in the benchmarking example ( view at source ↗

**Figure 7.** Figure 7: The finite element model of the target specimen in National Research Institute for Earth Science and Disaster Resilience view at source ↗

**Figure 8.** Figure 8: Observed frequency response functions in the experimental data (National Research Institute for Earth Science and view at source ↗

**Figure 9.** Figure 9: Inferred posterior samples in the experimental study for Case 1 (fully observed) and Case 2 (only the roof-top FRF view at source ↗

**Figure 10.** Figure 10: Predicted Q–d relationships for the 10%–scaled JMA Kobe excitation. observations for most stories, indicating that the proposed method captures the overall structural response even under reduced observability view at source ↗

**Figure 11.** Figure 11: Posterior distributions of time history responses for the 10%–scaled JMA Kobe excitation. view at source ↗

read the original abstract

Enhancing seismic fragility and risk assessment of nuclear power plants relies on accurate prediction of reactor building responses to seismic hazards, which can be further improved through dynamic analysis of high-fidelity finite element (FE) models. However, FE models often exhibit non-negligible discrepancies from actual structures due to various sources of uncertainty, necessitating FE model updating with rigorous quantification of associated uncertainties. This paper presents a GPU-accelerated latent space--based Bayesian framework for FE model updating of building structures. In the proposed framework, high-dimensional structural response data (e.g., time histories or frequency response functions) are projected into a low-dimensional latent space using a multimodal variational autoencoder (MVAE), thereby enabling efficient and tractable likelihood evaluation without explicit modeling in the original observation space. Once trained, the surrogate enables amortized inference, allowing posterior sampling to be performed without additional simulator evaluations. We specifically employ a sequential Monte Carlo (SMC) sampler, whose population-based formulation allows parallel evaluation of the approximate likelihood on GPUs, resulting in computational efficiency and robustness against multimodal and complex posterior distributions. The proposed framework is validated through both numerical benchmarking and experimental data from a shaking table test of a reinforced concrete building structure. The results demonstrate that the method accurately estimates structural parameters with well-quantified uncertainties, while achieving fast and efficient inference through GPU-based parallelization, and enabling robust inference even in the presence of sparse observations that induce multimodal and highly complex posterior distributions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper combines a multimodal VAE with GPU-parallel SMC to make Bayesian FE model updating feasible for high-dimensional seismic data on buildings, but the latent-space approximation lacks direct accuracy checks.

read the letter

The core contribution is a practical workflow that trains a multimodal variational autoencoder on structural responses like time histories or frequency functions, then runs amortized Bayesian inference via sequential Monte Carlo on GPUs. This lets the method avoid explicit high-dimensional likelihoods while handling multimodal posteriors from sparse observations. The abstract shows validation on both numerical benchmarks and a shaking-table test of a reinforced concrete structure, with claims of accurate parameter recovery and quantified uncertainty at reasonable compute cost. That combination for seismic FE updating is not just a rehash of existing VAE or SMC papers; it targets a real bottleneck in nuclear and building risk assessment where full-space modeling is too slow. The GPU parallelization and amortized aspect are clear engineering wins for repeated analyses. The weakest point is the unverified quality of the latent-space likelihood. The framework assumes the MVAE projection preserves enough information to keep posterior samples and uncertainty estimates reliable, yet the provided description gives no KL-divergence checks, moment comparisons, or reconstruction-error bounds against the original observation space on the same data. Without those, it's hard to know whether approximation bias creeps into the structural parameters or inflates uncertainties, especially under the sparse-data regimes the paper highlights. For a safety-critical application like reactor buildings, that gap is noticeable even if the numerical and experimental results look promising on the surface. This work is aimed at structural engineers and risk analysts who already use FE models and want uncertainty-aware updating without prohibitive compute. It is coherent on its own terms and shows honest engagement with the computational challenges, so it deserves a serious referee who can press on the approximation fidelity and reproducibility details. I would send it to review rather than desk-reject.

Referee Report

1 major / 3 minor

Summary. The manuscript presents a GPU-accelerated Bayesian framework for finite element model updating of building structures under seismic excitation. High-dimensional structural responses (time histories or frequency response functions) are projected into a low-dimensional latent space via a multimodal variational autoencoder (MVAE) to enable tractable likelihood evaluation. Posterior inference is performed with a sequential Monte Carlo (SMC) sampler that exploits GPU parallelization for efficiency and robustness to multimodal posteriors induced by sparse data. The approach is validated on numerical benchmarks and shaking-table experiments on a reinforced concrete structure, with claims of accurate structural parameter estimation and well-quantified uncertainties.

Significance. If the latent-space approximation preserves posterior fidelity, the framework offers a practical route to uncertainty-aware FE model updating for seismic fragility assessment of critical infrastructure such as nuclear power plants. The combination of amortized MVAE inference with parallel SMC is a clear computational strength, and the experimental validation on real shaking-table data adds relevance. However, the absence of direct fidelity diagnostics for the surrogate likelihood limits the strength of the accuracy claims.

major comments (1)

[Validation sections (numerical benchmarking and shaking-table experiments)] Validation sections (numerical benchmarking and shaking-table experiments): The central claim that the method 'accurately estimates structural parameters with well-quantified uncertainties' while remaining robust under sparse observations rests on the MVAE latent-space likelihood being a faithful surrogate. No quantitative fidelity check is reported (e.g., KL divergence between latent-space and full observation-space likelihoods, or differences in posterior moments/credible-interval coverage on identical datasets). This comparison is load-bearing for the accuracy and calibration assertions, especially given the paper's emphasis on multimodal posteriors.

minor comments (3)

[Abstract] Abstract: The phrase 'multimodal variational autoencoder (MVAE)' is introduced without a brief parenthetical description of its key architectural feature (joint encoding of multiple data modalities) that distinguishes it from a standard VAE.
[Methodology] Methodology: The latent dimension is treated as a free hyperparameter; a sensitivity study showing how posterior estimates and uncertainty quantification vary with this choice would clarify robustness.
[Figures and tables] Figures and tables: Posterior marginal plots should explicitly label whether they derive from the latent-space or full-space likelihood when both are feasible, to allow visual assessment of approximation quality.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their detailed and constructive review. The single major comment raises an important point about the need for explicit quantitative validation of the latent-space surrogate likelihood. We address this below and will incorporate the requested diagnostics in the revised manuscript.

read point-by-point responses

Referee: Validation sections (numerical benchmarking and shaking-table experiments): The central claim that the method 'accurately estimates structural parameters with well-quantified uncertainties' while remaining robust under sparse observations rests on the MVAE latent-space likelihood being a faithful surrogate. No quantitative fidelity check is reported (e.g., KL divergence between latent-space and full observation-space likelihoods, or differences in posterior moments/credible-interval coverage on identical datasets). This comparison is load-bearing for the accuracy and calibration assertions, especially given the paper's emphasis on multimodal posteriors.

Authors: We agree that direct quantitative fidelity diagnostics for the MVAE surrogate would strengthen the accuracy and calibration claims. The current manuscript validates the framework by showing that parameter estimates recover ground-truth values within credible intervals on numerical benchmarks and match independent experimental measurements on the shaking-table data, with the SMC sampler demonstrating robustness to the multimodal posteriors induced by sparse observations. However, we did not include explicit side-by-side comparisons of posteriors obtained with the full high-dimensional likelihood versus the latent-space approximation. In the revised manuscript we will add such a comparison on the numerical benchmark cases (where the full likelihood remains tractable for low-dimensional subsets of the data). We will report differences in posterior means and variances, credible-interval coverage rates, and, where feasible, estimated KL divergence between the two likelihoods. This addition will directly address the load-bearing nature of the surrogate for the stated claims. revision: yes

Circularity Check

0 steps flagged

No circularity detected; framework applies standard MVAE projection and SMC sampling to FE model updating

full rationale

The derivation chain consists of projecting high-dimensional response data into a latent space via a multimodal variational autoencoder, followed by amortized likelihood evaluation and GPU-parallelized sequential Monte Carlo sampling. These steps rely on established variational inference and Monte Carlo methods applied to a new application domain. Validation occurs through independent numerical benchmarks and shaking-table experiments on a reinforced concrete structure, with no equations or claims reducing the reported parameter estimates or uncertainty quantification to quantities defined by the same fitted inputs or self-citations. The central efficiency and robustness results therefore remain self-contained against external data.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, limiting visibility into specific hyperparameters and assumptions; the framework rests on standard variational inference and Monte Carlo sampling assumptions plus the domain-specific claim that latent-space approximation preserves sufficient information for likelihood evaluation.

free parameters (1)

latent dimension
The dimensionality of the MVAE latent space is a modeling choice that controls the trade-off between compression and fidelity of the structural response data.

axioms (1)

domain assumption The multimodal variational autoencoder provides a sufficiently accurate approximation to the true data-generating process for tractable likelihood evaluation in the latent space.
Invoked to justify replacing explicit high-dimensional likelihood with the surrogate without loss of posterior accuracy.

pith-pipeline@v0.9.0 · 5579 in / 1334 out tokens · 93603 ms · 2026-05-08T09:13:22.523089+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

5 extracted references · 2 canonical work pages

[1]

Accessed: 2026-04-11

ARKInformationSystems,Inc.,.TDAPIII:General-purpose3-DimensionalDynamicAnalysisProgramfor Civil Engineering and Architectural Use.https://www.ark-info-sys.co.jp/jp/product/tdap/ english/. Accessed: 2026-04-11. Beck, J.L., Katafygiotis, L.S.,

2026
[2]

A kernel two-sample test. J. Mach. Learn. Res. 13, 723–773. Hoffman,M.D.,Gelman,A.,2014. TheNo-U-Turnsampler: AdaptivelysettingpathlengthsinHamiltonian Monte Carlo. Journal of Machine Learning Research 15, 1593–1623. Huang, Y., Shao, C., Wu, B., Beck, J.L., Li, H.,

2014
[3]

Shake Table Experiment of Downsized 6-story RC Building with Earthquake Resisting Wall Frame

E-Defense Experimental Data Archive (ASEBI): Project Name "Shake Table Experiment of Downsized 6-story RC Building with Earthquake Resisting Wall Frame" Experiment No. E201401. doi:10.17598/NIED.0020-E201401. Neal, R.M.,

work page doi:10.17598/nied.0020-e201401
[4]

A survey of multimodal deep generative models. Adv. Robot. 36, 261–278. Wang,T.,Bi,S.,2026. Stochasticmodelupdatingusingconditionaldiffusion-basedprobabilisticgenerative models. Mech. Syst. Signal Process. 246, 113891. Wang, T., Bi, S., Zhao, Y., Dinh, L., Mottershead, J.,

2026
[5]

Data-driven stochastic model updating and damage detection with deep generative model. Mech. Syst. Signal Process. 232, 112743. Yaoyama,T.,Itoi,T.,Iyama,J.,2024. Probabilisticmodelupdatingofsteelframestructuresusingstrainand accelerationmeasurements: Amultitasklearningframework. StructuralSafety108,102442. doi:https: //doi.org/10.1016/j.strusafe.2024.1024...

work page doi:10.1016/j.strusafe.2024.102442 2024