PaA – Partnership Area

Project

PaA-5

AI Based Simulation of Transient Physical Systems – From Benchmarks to Hybrid Solutions

Project Heads

Martin Eigel, Hanno Gottschalk, Alexander Klawonn

Project Members

Onur Tanil Doganay

Project Duration

01.01.2024 − 31.12.2025

Located at

TU Berlin

Description

AI tools have entered the application area of industrial physics-based simulations, but there is still little knowledge on the actual strengths and weaknesses of the different methods. We develop a systematic benchmark of various recent AI techniques with
regard to their industrial applicability and develop hybrids that combine their respective strengths.

Toward this end, we apply three different AI techniques to two reference problems that sufficiently capture the difficulties and complexities of applications from the industry partner Siemens Energy. The two reference problems are explained in detail below. As for the three AI techniques to benchmark we investigate

• Physics-Informed Neural Networks (PINNs) since they have become a widely popular technique to solve complex partial differential equations (PDEs),
• Graph Neural Networks (GNNs) which take the approach of learning physical behavior on a local scale (the communication
  between nodes across edges) and generalize to the global problem setting at hand,
• and Generative AI, where we consider Generative Adversarial Networks (GANs).

For this project we aim to successfully tackle the two reference problems with each of the three AI techniques and benchmark their performances. Furthermore, with these insights we intend to develop hybrid approaches, i.e., combinations of the three techniques.

There is one central problem of industrial relevance in this project, which serves as a reference to evaluate the performance of the different methods.  It describes heat transfer phenomena which, in general, are encountered in many different technical applications where systems or components are heating up during operation.

The applications we have in mind here are assets of the electrical grid, such as high-power switch gears or transformers. In these kinds of components, one major heat source is the electrical current’s power loss in the conductors, which might have a cross sectional diameter of up to ten centimeters or more. Different cooling mechanisms are utilized to keep the system’s temperature below the respective design limits. One of the central contributors in this context is convection, a mechanism which is often difficult to simulate in complex settings (especially in its evolution over time) due to being a fluid-solid interaction. The  reference problem we consider is therefore both describing primarily convection phenomena.

The general reasoning behind chosen problem was to include the physical phenomena which make their simulation challenging while keeping the problem complexity as low as possible, to avoid high costs for training data generation as well as effort for data handling. This is also the reason why both the problem is defined in 2D instead of 3D.

The difficulties in describing the physics accurately lies in the fact that there is a slow transient global warming process combined with unsteady flow features of convective heat transfer. This requires long roll-out times and hig model stability beyond next state snapshot prediction.

The Benchmark Problem – Natural Convection

This bechmark problem describes two conductors in air with gravity (g) acting downwards, see the figure below. The two aluminum conductors are positioned vertically aligned with some distance between them, so that a convective air stream from the conductor below will meet the conductor above and influence its temperature. A respective heat source is defined on both conductors, so that they will heat up with time. The two conductors are enclosed in a box which has a fixed ambient temperature at its boundary, allowing the generated heat from the conductors to leave the domain. Consequently, the system cannot increase its average temperature indefinitely but will converge to a stationary state. However, due to turbulence there will still be small flow field fluctuations in this quasi-static converged state. Note the authors are aware that turbulence is a 3D-phenomenon while the considered problem is only in 2D. But even if physically not fully accurate, the chaotic behavior is also encountered in a 2D setup.

Problems like this are encountered for example in high-power, gas-insulated switch gears. Here, not two but three conductors (three-phase current) are enclosed in a metal hull containing an insulating gas (i.e., pressurized air).

Project Webpages

Related Publications

When do World Models successfully Learn Dynamical Systems?

In a fist paper we introduced World Models into operator learning. We achieved

• One of the first papers to pursue operator learning with world models.
• Laying system theoretical foundations for the question, when a WM is capable to learn a given dynamical systems
• An extensive data set published on ZENODO, containing wave, heat, Kuramoto-Sivashinski and Navier Stokes data.
• Training of various linear and non linear WM.
• Numerical analysis of observability as a foundation of successfull learning.
• Benchmarking with other operator learning methods like DeepONet and FNO with clear superiority of WM on long roll-out times

Edmund Ross, Claudia Drygala, Leonhard Schwarz, Samir Kaiser, Francesca di Mare, Tobias Breiten, Hanno Gottschalk, When do World Models successfully Learn Dynamical Systems?, arXiv:2507.04898 (https://arxiv.org/abs/2507.04898).

Learning Transient Convective Heat Transfer with Geometry Aware World Models

The main publication of this project utilizes world models to learn the transient physics of convective heating in a substation. We achieved

• The production of a challenging data set for convective heat transfer published open source on Zenodo.
• The design of a new world model architecture which is geometry and boundary value aware (i.e. able to freely place the conductor in the space)
• Training of the world model giving fast reproduction of the flow physics
• Development of detailed metrics for physics based evaluation
• Extensive benchmarking against FNO and DeepOnet operator learning approaches and PINN, with our WM approach the clear winner
• A proof that also transient physics can be learned along with turbulent flow

Onur T Doganay, Alexander Klawonn, Martin Eigel, Hanno Gottschalk, Learning Transient Convective Heat Transfer with Geometry Aware World Models, arXiv: 2601.22086 (2026).https://https://arxiv.org/pdf/2601.22086

Related Picture