Master thesis (f/m/x): Loss balancing algorithms for Physics-Informed Neural Networks

Master Thesis – Physics‑Informed Neural Networks (PINNs) (m/f/d)

The DLR Institute of Maintenance and Modification is dedicated to shaping the future of aviation through research and technology transfer. With its vision of “We Maintain Mobility for a Sustainable Future,” the institute focuses on life cycle analysis, maintenance technologies, and the optimization of digital processes to improve technical operations in aviation.

What To Expect

A dynamic team of scientists awaits you in the Process Optimization and Digitalization department at the DLR Institute of Maintenance and Modification. We are researching new concepts for digitalized maintenance. This includes digital twins, methods for data analysis for condition monitoring and prediction, new approaches to automated interaction between networked stakeholders, AI methods, and new methods for dynamic process modeling and simulation.

Your tasks

We conduct research in hybrid models for ultrasonic inspection optimization incorporating prior physical knowledge about structural wave propagation into the training process of neural networks. Physics‑informed neural networks (PINNs) are deep learning models that exploit physical laws by incorporating partial differential equations as penalty terms in their loss function. PINNs rely less on labeled data because the underlying physical equations act as an intrinsic form of supervision. However, the accurate weighting of the combination of multiple loss functions for the training of PINNs represents a significant challenge in the process of generating predictions with low uncertainty. Different algorithms exist for the adjustment of weighting factors of the loss function. As part of the master’s thesis, these algorithms are to be adapted and analysed using PINNs for the two‑dimensional acoustic wave equation.

Thesis includes

• literature research regarding existing loss balancing algorithms for Physics‑Informed deep learning models
• implementation and adaptation of existing loss balancing algorithms and implementation of grid search algorithms and convergence analyses with continuous optimisation of the model hyperparameters
• evaluation of results as well as statistical data evaluation and error analysis and documentation and presentation of the results

Your profile

• ongoing scientific studies in the field of aerospace engineering, computer science, mathematics, physics or comparable degree programmes and methodical, autonomous and precise approach to work
• good programming skills and experience in Python, additional experience in development environments e.g. Docker, Gitlab
• basic knowledge of machine learning algorithms e.g. Tensorflow and advanced statistical data analysis

We look forward to getting to know you!

If you have any questions about this position (Vacancy‑ID 3070) please contact: Ann‑Kathrin Koschlik, Tel.: . Location: Aachen, North Rhine‑Westphalia, Germany.