[01099] Physics-based and data-driven modeling for digital twins
Session Date & Time :
01099 (1/3) : 2C (Aug.22, 13:20-15:00)
01099 (2/3) : 2D (Aug.22, 15:30-17:10)
01099 (3/3) : 2E (Aug.22, 17:40-19:20)
Type : Proposal of Minisymposium
Abstract : Digital twins have emerged in recent years as a paradigm for the lifetime operation of physical assets. A digital twin is an exact virtual representation of a physical asset that uses real-time data. The construction of a digital twin requires the use and sometimes union of different modeling methods that include physics-based modeling, data assimilation, data-driven modeling, and model reduction. This minisymposium has the goal to bring together researchers working on the theory and practice of modeling in the context of digital twins with a particular focus on industrial applications.
Shen Yin (Norwegian University of Science and Technology (NTNU))
Dimitri Goutaudier (Federal Institute of Technology Lausanne (EPFL))
Peter Benner (MPI Magdeburg)
Steven Brunton (University of Washington)
Art Pelling (TU Berlin)
Jochen Cremer (TU Delft)
Olga Fink (Federal Institute of Technology Lausanne (EPFL))
Sebastian Hartmann (Siemens AG/TU Munich)
Karim Cherifi (TU Berlin)
Zoltan Horvath (Széchenyi István University)
Talks in Minisymposium :
[01503] Hierarchical modeling of electrical machines in the context of digital twins
Author(s) :
Karim Cherifi (TU Berlin)
Volker Mehrmann (TU Berlin)
Philipp Schulze (TU Berlin)
Abstract : Digital twins of electrical machines require mathematical models that are accurate enough for the design and fast enough to be used for condition monitoring. This leads to a hierarchy of models that are used within the digital twin. These models must in addition incorporate the physical coupling between electrical, mechanical, and thermal phenomena for more accurate computations. In this talk, we present how one can construct this model hierarchy by incorporating physics-based and data-driven modeling.
[01631] Weakly supervised learning for power grid state estimation
Author(s) :
Jochen Lorenz Cremer (TU Delft)
Elvin Isufi (TU Delft)
Benjamin Habib (TU Delft)
Abstract : In this talk, I present a novel approach for Distribution System State Estimation (DSSE) called the Deep Statistical Solver for Distribution System State Estimation (DSS^2). This approach, based on graph neural networks (GNNs) and weakly-supervised learning, addresses the challenges of lack of observability and high density in the distribution system. DSS$^2$ uses hypergraphs to represent the heterogeneous components of the distribution system and updates their latent representations via a node-centric message-passing scheme. Our approach allows for the training of DSS$^2$ using noisy and corrupted measurements, alleviating the need for ideal labelled data. The results of our experiments on various sizes of power networks showed DSS2 outperforms the conventional Weighted Least Squares algorithm in terms of accuracy, convergence, and computational time and is more robust to noisy, erroneous, and missing measurements. Our approach demonstrates the potential of weakly-supervised learning in DSSE and the ability to respect the physical constraints of the distribution system while learning from noisy measurements.
[02075] Hamiltonian structure-preserving non-intrusive operator inference for predictive digital twins
Author(s) :
Anthony Gruber (Sandia National Laboratories)
Irina Tezaur (Sandia National Laboratories)
Max Gunzburger (University of Texas at Austin)
Abstract : To serve as reliable predictive tools, digital twins require dimensionality-reduction techniques that preserve key properties of the underlying equations. This talk presents a novel non-intrusive structure-preserving model reduction technique for canonical and non-canonical Hamiltonian systems based on operator inference. The method reduces to a straightforward linear solve given snapshot data and “gray-box” knowledge of the underlying problem. We demonstrate that, unlike traditional reduction methods, the proposed approach delivers stable, accurate, energy-conserving and robust reduced-order models.
[02141] Machine Learning for Scientific Discovery, with Examples in Fluid Mechanics
Author(s) :
Steven Brunton (University of Washington)
Abstract : This work describes how machine learning may be used to develop accurate and efficient nonlinear dynamical systems models for complex natural and engineered systems. We explore the sparse identification of nonlinear dynamics (SINDy) algorithm, which identifies a minimal dynamical system model that balances model complexity with accuracy, avoiding overfitting. This approach tends to promote models that are interpretable and generalizable, capturing the essential “physics” of the system.
[03242] Towards smart city digital twins
Author(s) :
Francisco Chinesta (ENSAM)
Daniele Di Lorenzo (ESI)
Victor Champaney (ENSAM)
Angelo Pasquale (ENSAM)
Amine Ammar (ENSAM)
Elias Cueto (I3A)
Dominique Baillargeat (CNRS@CREATE)
Abstract : Smart cities are composed of a number of coupled complex system of systems. Modelling them needs enhancing the traditional physics-based modelling approaches as well as speeding-up the predictions. Artificial intelligence and data-driven models obtained by using physics-informed and physics-augmented learning represents a valuable approach where accuracy and rapidity met. In this presentation advanced methodologies will be used for addressing complex scenarios, enabling real-time diagnosis and prognosis.
[03365] Exploring security challenges in enhancing Digital Twins capabilities with ChatGPT
Author(s) :
Xingheng Liu (NTNU)
Shen Yin (NTNU)
Jie Liu (NTNU)
Jørn Vatn (NTNU)
Asmae Bni (NTNU)
Abstract : Digital twins offer valuable insights for monitoring and managing physical systems. With the upsurge of ChatGPT, integrating it with digital twins during the design or operation phase could unlock new capabilities. However, this integration may introduce new security challenges and vulnerabilities. In this talk, we will briefly discuss the potential of enhancing digital twins with ChatGPT and the associated security concerns, emphasizing the importance of addressing these issues to ensure robust, secure DTs.
[03438] Data-driven Balancing for Acoustical Systems
Author(s) :
Art J.R. Pelling (TU Berlin)
Ennes Sarradj (TU Berlin)
Abstract : Constructively modelling acoustical systems is difficult due to unknown material and domain properties and complexity of dynamics. Although measurement data is abundantly available, reduced order modelling is not well-established in the field.
We showcase recent system identification methods from the mathematical community and analyze their aptitude and performance in real applications that involve high-dimensional measurement data. Amongst others, we consider head-related transfer functions that are used for auralization in virtual reality applications.
[04060] From physics to machine learning and back: Applications to fault diagnostics and prognostics
Author(s) :
Olga Fink (EPFL)
Abstract : Deep learning requires representative data, but condition monitoring data for complex systems lack labels and representativeness. Integrating physics can help to overcome this.
The talk will give some insights into various techniques that combine physics-based and deep learning algorithms, as well as incorporate structural inductive bias for fault diagnostics and prognostics. The focus will be in particular on calibration-based hybrid approaches, physics-enhanced graph neural networks and on transformer-based architectures combined with transfer learning.
[04485] Constrained Optimal Sensing for Nuclear Digital Twins
Author(s) :
Krithika Manohar (University of Washington)
Abstract : We develop a constrained optimization for sensor placement in nuclear digital twins where sensing capability may be severely constrained or limited. These constraints may arise in certain areas of a reactor due to hostile operating conditions, accessibility issues, and physical limitations on sensing capability. Our data-driven method optimizes sensor placement with constraints for full flow field reconstruction, leveraging reduced order models of flow physics. We demonstrate the technique is near optimal using empirical and theoretical validation and provide uncertainty analyses for noisy sensor measurements. The method is demonstrated on a nuclear fuel rod prototype which is heated to mimic the neutronics effect of nuclear fuel within the Transient Reactor Test facility (TREAT) at Idaho National Laboratory.
[04687] Comparison of physics-based and data-driven surrogate models of a gas-bearings supported rotor
Author(s) :
Dimitri Goutaudier (EPFL)
Jürg Schiffmann (EPFL)
Fabio Nobile (EPFL)
Abstract : Gas bearings use pressurized gas as a lubricant to support and guide rotating machinery. These bearings have several advantages over traditional lubricated bearings but they more complex to operate and exhibit nonlinear behaviors. In this contribution, we present physics-based and data-driven frameworks to compute the dynamics of a gas-bearings supported rotor operating at very high rotation speeds. We compare the numerical performances of the two approaches, and we propose research directions to improve the models.
[04925] Reduced order modelling for large-scale CFD
Author(s) :
Zoltán Horváth (Széchenyi István University)
Mátyás Yves Constans (Széchenyi István University)
Abstract : The RedSim in-house reduced-order modeling (ROM) software for the 3D compressible Euler and Navier-Stokes equations is introduced. RedSim’s core consists of a finite volume code running on GPUs and the proper orthogonal decomposition-based ROM-module. The application of RedSim to digital twinning for urban air pollution (in HiDALGO2 EuroHPC Centre of Excellence) and an acoustics problem in the automotive industry are presented. Numerical experiments raise mathematical challenges, which will be presented and some of them solved.
[05203] Digital twins for green carbon processes
Author(s) :
Peter Benner (MPI for Dynamics of Complex Technical Systems, Magdeburg)
Ion Victor Gosea (Max Planck Institute for Dynamics of Complex Technical Systems)
Abstract : Recent advances in scientific computing have made digital twins (DTs) increasingly popular in various fields,
including process engineering (ProcEng). DTs have the potential to transform ProcEng by enabling real-time
monitoring or process optimization. However, their full potential in this area has yet to be realized. We provide an
overview of computational tools required for developing DTs in ProcEng. It characterizes models used to develop
DTs and discusses the challenges and requirements associated with their implementation. We particularly focus on
the development of sustainable chemical production processes in the context of a green carbon society.
