Model order reduction (MOR) has become one of the most widely used tools to create efficient surrogate models for time‐critical applications. For nonlinear models, however, linear MOR approaches are only practicable to a limited extent. Nonlinear approaches, on the contrary, often require intrusive manipulations of the used simulation code. Hence, nonintrusive MOR approaches using classic model order reduction along with machine learning (ML) algorithms can provide remedy. Such approaches have drawn a lot of attention in the recent years. They rely on the idea to learn the dynamics not in a high dimensional but in a reduced space, that is, they predict the discrete sequence of reduced basis' coefficients. Open questions are the suitability of such methods in the field of structural dynamics and the best choice of the used ML algorithm. Both are addressed in this article in addition to the integration of the methodology into a modular and flexible framework that can effortless be adapted to various requirements. By applying the methodology to a dynamic mechanical system, accurate surrogate models are received, which can speed up the simulation time significantly, while still providing high‐quality state approximations.
Over the last decades, computer modeling has evolved from a supporting tool for engineering prototype design to an ubiquitous instrument in non-traditional fields such as medical rehabilitation. This area comes with unique challenges, e.g. the complex modeling of soft tissue or the analysis of musculoskeletal systems. Conventional modeling approaches like the finite element (FE) method are computationally costly when dealing with such models, limiting their usability for real-time simulation or deployment on low-end hardware, if the model at hand cannot be simplified without losing its expressiveness. Non-traditional approaches such as surrogate modeling using data-driven model order reduction are used to make complex high-fidelity models more widely available regardless. They often involve a dimensionality reduction step, in which the high-dimensional system state is transformed onto a low-dimensional subspace or manifold, and a regression approach to capture the reduced system behavior. While most publications focus on one dimensionality reduction, such as principal component analysis (PCA) (linear) or autoencoder (nonlinear), we consider and compare PCA, kernel PCA, autoencoders, as well as variational autoencoders for the approximation of a continuum-mechanical system. In detail, we demonstrate the benefits of the surrogate modeling approach on a complex musculoskeletal system of a human upper-arm with severe nonlinearities and physiological geometry. We consider both, the model’s deformation and the internal stress as the two main quantities of interest in a FE context. By doing so we are able to create computationally low-cost surrogate models which capture the system behavior with high approximation quality and fast evaluations.
Highly nonlinear dynamic finite element simulations using explicit time integration are particularly valuable tools for structural analysis in fields like automotive, aerospace, and civil engineering, or in the study of injury biomechanics. However, such state-of-the-art simulation models demand significant computational resources. Conventional data-driven surrogate modeling approaches address this by evolving the dynamics on low-dimensional embeddings, yet the majority of them operate directly on high-resolution data obtained from numerical discretizations, making them costly and unsuitable for adaptive resolutions or for handling information flow over large spatial distances. We therefore propose a multi-hierarchical framework for the structured creation of a series of surrogate models at different resolutions. Macroscale features are captured on coarse surrogates, while microscale effects are resolved on finer ones, while leveraging transfer learning to pass information between scales. The objective of this study is to develop efficient surrogates for a kart frame model in a frontal impact scenario. To achieve this, its mesh is simplified to obtain multi-resolution representations of the kart. Subsequently, a graph-convolutional neural network-based surrogate learns parameter-dependent low-dimensional latent dynamics on the coarsest representation. Following surrogates are trained on residuals using finer resolutions, allowing for multiple surrogates with varying hardware requirements and increasing accuracy.
Generative models are transforming science and engineering by enabling efficient synthetization and exploration of new scenarios for complex physical phenomena with minimal cost. Although they provide uncertainty-aware predictions to support decision making, they typically lack physical consistency, which is the backbone of computational science. Hence, we propose VENI, VINDy, VICI - a novel physical generative framework that integrates data-driven system identification into a probabilistic modeling approach to construct physically consistent and efficient reduced-order models with uncertainty quantification. First, VENI (Variational Encoding of Noisy Inputs) employs variational autoencoders to identify reduced coordinates from high-dimensional, noisy measurements. Simultaneously, VINDy (Variational Identification of Nonlinear Dynamics) extends sparse system identification methods by embedding probabilistic modeling into the discovery process. Last, VICI (Variational Inference with Credibility Intervals) enables efficient generation of full-time solutions and provides uncertainty quantification for unseen parameters and initial conditions. We demonstrate the performance of the framework across chaotic and high-dimensional nonlinear systems.
27 November 2025 in Computational Science and Engineering
Conventional physics-based modeling techniques involve high effort, e.g., time and expert knowledge, while data-driven methods often lack interpretability, structure, and sometimes reliability. To mitigate this, we present a data-driven system identification framework that derives models in the port-Hamiltonian (pH) formulation. This formulation is suitable for multi-physical systems while guaranteeing the useful system theoretical properties of passivity and stability. Our framework combines linear and nonlinear reduction with structured, physics-motivated system identification. In this process, high-dimensional state data obtained from possibly nonlinear systems serves as input for an autoencoder, which then performs two tasks: (i) nonlinearly transforming and (ii) reducing this data onto a low-dimensional latent space. In this space, a linear pH system that satisfies the pH properties per construction is parameterized by the weights of a neural network. The mathematical requirements are met by defining the pH matrices through Cholesky factorizations. The neural networks that define the coordinate transformation and the pH system are identified in a joint optimization process to match the dynamics observed in the data while defining a linear pH system in the latent space. The learned, low-dimensional pH system can describe even nonlinear systems and is rapidly computable due to its small size. The method is exemplified by a parametric mass-spring-damper and a nonlinear pendulum example, as well as the high-dimensional model of a disc brake with linear thermoelastic behavior.
28 September 2023 in Proceedings in Applied Mathematics and Mechanics
The challenges in automation system development are driven by short development cycles and individualization along with resource‐constraints. State of the art solutions do not provide the necessary digital tools to apply model‐based methods in automation engineering to achieve higher performing systems. To overcome these issues this paper presents a novel approach to address some of the current challenges in automation systems development using digital twins with extended dynamic behaviour. The study underscores how dynamic models can be imported through standardised interfaces into virtual commissioning (VC) tools, improving the development process by effectively utilising domain‐specific expertise. The paper highlights how these digital twins enhance not only the VC process but can also be applied to model‐based control methods. Initial experiments showcase the utility of digital twins in calculating dynamic acceleration limits during trajectory planning of CNC control and enhancing feed‐forward control. Further, the importance of parameter identification in achieving accurate system models is stressed. The initial results are promising, and future work aims to combine these methods in an industrial application involving a newly developed, individual lightweight robot, demonstrating the potential for enhanced design, accelerated development, and resource efficiency in automation systems.