Viktor Grimm successfully defended his dissertation

On July 10, Viktor successfully defended his dissertation at the University of Cologne. Viktor's research is situated at the confluence of machine learning and Computational Fluid Dynamics (CFD). His primary focus has been the development of surrogate models for the behavior of incompressible fluids within diverse geometries.

Specifically, he has concentrated on the utilization of machine learning methods, particularly convolutional neural networks (CNNs). Such networks can be trained on large datasets of reference data, often high-fidelity CFD simulations. However, these datasets can be exceedingly large and exceedingly expensive to produce, sometimes reaching prohibitive levels. Moreover, data-centric training methodologies are susceptible to overfitting, leading to the generation of physically implausible predictions.

In response to these challenges, Viktor investigated a novel approach to training CNNs as surrogate models for fluid behavior—one that exclusively relies on the inherent physics of the system. To achieve this, he has employed finite difference techniques to calculate the residuals of the governing equations, known as the Navier-Stokes equations. Building on these residuals, Viktor has formulated a loss function that permits CNN training without dependence on reference data.

Viktor's innovative technique has been applied to successfully train surrogate models for fluid flow in various two and three-dimensional geometries. These encompass simple rectangular shapes to more intricate configurations, such as intracranial arteries with aneurysms. Furthermore, he has extended this approach to accommodate varying boundary conditions, enhancing the versatility and applicability of his physics-aware surrogate modeling approach.