Tumor perfusion and vascular properties are important determinants of cancer response to therapy and thus various approaches for imaging perfusion are being explored. The ultimate goal of this research is to introduce and translate a bedside, non-invasive approach to characterize complex tumor vascular properties by considering the interconnected nature of voxels in 3D dynamic contrast-enhanced ultrasound (DCE-US).
Machine learning based approaches have proven very successful within certain problem settings, in particular in imaging. Lately also more and more successes for solving partial differential equations have been reported, e.g., by using pyhsics-informed neural networks, or to improve coarse-grid simulations by learned corrections. In this project we will investigate approaches to facilitate such techniques to improve performance of parallel-in-time methods.
Gradient-based methods for PDE-constrained optmization problems are extremely computationally expensive due to the need to fully solve the forward PDE and an adjoint equation every optimization iteration. We are therefore investigating the use of the PFASST algorithm to parallelize both forward and adjoint solution steps. Since PFASST is iterative, space-time initial conditions for each PDE solve can use the solution from the previous optimization iteration leading to additional computational savings.
In the BMBF funded project ExaOcean we aim to accelerate the computation of the ICON-O ocean model, moving toward ExaScale. For this, we will use parallel spectral deferred correction methods, and combine classical numerical methods with machine learning (ML) approaches to include effects of sub-mesoscale eddies, which are not resolved on the computational mesh.
Im Rahmen des ersten dreijährigen Förderzeitraums werden gemeinsam Greybox-Modelle zur Vorhersage des Werkzeugverschleißes bei der Drehbearbeitung von hochlegiertem Edelstahl mit TiAlN-beschichteten Werkzeugen entwickelt und qualifiziert. Neben der Entwicklung und Inbetriebnahme eines automatischen Verschleißversuchsstandes zur Generierung von Massendaten stehen vor allem Methoden im Fokus, welche maschinelles Lernen und domänenspezifisches Wissen in solcher Weise kombinieren, dass sich die Modelle des Werkzeugverschleißes über die Trainingsgrenzen hinweg extrapolieren lassen. Die zwingend erforderliche spanende Bearbeitung von Edelstählen, z. B. für maritime Anwendungen, gilt aufgrund der hohen Festigkeit und geringen Wärmeleitfähigkeit des Werkstoffes als schwer. Mit Hilfe der grundlagenorientierten Forschung im Rahmen dieses Projektes können durch eine genaue Vorhersage der Werkzeugstandzeit sowohl die Fertigungsmaschinen (Reduzierung Energieverbrauch) als auch die Ressource Werkzeug (Wolfram und Kobalt) besser ausgenutzt werden.
The aim of the project is the development of deep normalizing flow networks (NFs) to obtain probability densities of stochastic data and demonstrate their usage in precision measurements and searches for new interactions in particle collisions at highest energies, as well as for inversion problems. While deep neural networks are by now routinely used in most classification problems in particle physics, their usage is still limited since the understanding of the network response in terms of likelihoods is most of the time unknown. Likelihood-based estimation however is the standard working horse for both precision measurements of fundamental parameters (e.g. top quark mass) and for searches of rare processes (e.g. triple-Higgs interactions), where only a few measured events dominate the sensitivity. The NF approach can be a game changer towards interpretable neural networks for many applications. This however requires research for multiclass applications, uncertainty estimation, and inversion properties.