Abstract:This work introduces an efficient methodology for: (i) predicting dynamic responses across a broad frequency band for large-scale, highly complex structures, and (ii) forecasting their high-frequency response using associated low-frequency information. Structures of interest are characterized by a large number of degrees of freedom (DOFs) and numerous local vibration modes that couple, within the frequency range of interest, with long-wavelength global elastic modes. Given this complexity, high-fidelity numerical models present a significant computational burden, and a reduced-order modeling (ROM) approach was developed by the authors using substructuring techniques. However, despite its enhanced efficiency, this ROM remains prohibitive for tasks such as uncertainty quantification (UQ). Moreover, given the limitations on the number of training samples used for machine learning tasks, the "small-data" challenge is not easily addressed by traditional machine learning paradigms. Another obstacle for this class of problems is the large number of input parameters and response quantities of interest. This high-dimensional setting exacerbates the small-data challenge for UQ and, more generally, for the development of predictive models. This paper presents a novel implementation of the Probabilistic Learning on Manifold (PLoM) method, which is adept at handling small datasets for both supervised and unsupervised learning. Even with scarce data, PLoM can identify the low-dimensional manifold inherent in the training dataset. Subsequently, projected Itô sampling—a stochastic sampling method—facilitates the generation of samples constrained to the manifold, thereby augmenting the training data. Nonparametric conditioning is then used to infer the joint distribution of a subset of the observables, conditional on another subset having been observed. The focus of this paper is particularly on a multi-level spent nuclear fuel canister, and we infer the statistical properties of high-frequency dynamics, encoded as a frequency response function (FRF), from low-frequency observations of the FRF.

Manifold Learning for Organizing Unstructured Sets of Process Observations

Deep Manifold Computing and Visualization

Manifold Learning with Contracting Observers for Data-driven Time-series Analysis

Interactive Pattern Discovery in High-Dimensional, Multimodal Data Using Manifolds

Probabilistic Learning on Manifolds (PLoM) for cross-scale diagnostics in structural dynamics

Manifold Learning in Atomistic Simulations: A Conceptual Review

A Self-Organising Multi-Manifold Learning Algorithm

An integrated manifold learning approach for high-dimensional data feature extractions and its applications to online process monitoring of additive manufacturing

From snapshots to manifolds - A tale of shear flows

Local manifold learning and its link to domain-based physics knowledge

Data-driven discovery of intrinsic dynamics

Manifold Learning for Accelerating Coarse-Grained Optimization

Hierarchical Structuring Of Data On Manifolds

Modeling of High Dimensional Blast Furnace System by Manifold Learning

Self-organized manifold learning and heuristic charting via adaptive metrics

Manifold Learning for the Emulation of Spatial Fields from Computational Models.

Manifold-based learning and synthesis.

From partial data to out-of-sample parameter and observation estimation with Diffusion Maps and Geometric Harmonics

Actuation manifold from snapshot data

Overview of Manifold Learning

On Learning what to Learn: heterogeneous observations of dynamics and establishing (possibly causal) relations among them