Abstract

Numerical simulations have a long history of reshaping the way we model and understand physical phenomena and help to optimize processes, predict behavior and discover new principles. In the past decade machine learning, especially deep learning, has become a popular new tool that industries in many domains are starting to pick up not only to exploit data patterns but as a complementary approach to these classical numerical simulations. Physics-based deep learning is a field that combines neural networks, typically supervised or reinforcement learning models, alongside numerical PDE solvers to tackle problems in domains including climate modeling, predictive maintenance, computer graphics, plasma or accelerator physics and many others. This field has a huge potential to influence basically all computational methods in the upcoming decades. 

This talk will give an overview of some of the hot and quickly evolving topics related to deep learning in the context of physics-based simulations, including examples from standard supervised learning, physical loss constraints, more tightly coupled learning algorithms with differentiable simulations, training algorithms tailored to physics problems, as well as reinforcement learning and uncertainty modeling. The lecture is based on the book "Physics-based Deep Learning" by N. Thuerey et. al. 

Bio

Peter is a PhD student in particle accelerator physics, studying the nonlinear dynamics and instabilities of colliding beams at the proposed Future Circular Collider (FCC-ee). Based at CERN, he is part of a team developing a general purpose beam dynamics simulation software. He also holds a BSc in physics and an MSc in applied and engineering physics with a specialization in scientific computing. Besides his studies, he broadened his experience in the frame of side projects and internships, both in industry and academia.