The Institute of Structures and Design works on the development and optimization of materials and their processing and joining technologies, as well as new design approaches and the construction of full-scale demonstrators. Naturally, testing and validation in specialized test facilities and flight trials are also part of the daily work. The focus is on fiber-ceramic, polymer, and hybrid composite materials. The research of new multidisciplinary design tools and digital models forms the basis for hardware development.

What to expect

The research objectives of the Structural Integrity Department are to develop simulation and testing methods and to design aerospace structures in a way that ensures accidents or incidents in the crash and impact category are survivable for occupants. This includes material and structural testing as well as numerical analyses ranging from individual specimens to the entire system.

In the DLR Project “Agile hochratenfertigung und optimierte Inspektion für Kleinflugzeuge”, we aim to reduce the number of required material testing through a data-driven approach. Incorporating machine learning models could allow us to accelerate composite structure development while reducing associated costs and testing efforts.

However, machine learning models are often criticised as being “black boxes”, where the mapping of inputs to outputs is unclear. To address this issue, a new modelling technique called Constitutive Artificial Neural Networks (CANNs) has been shown to reliably incorporate constitutive laws into neural networks, allowing for better explainability. By integrating physical laws directly into the learning process, the model aims to enable reliable prediction of material properties. Thus, allowing the extraction of constitutive parameters from the model could provide interpretable intermediate outputs that correspond to known material properties. 

In our previous work, a CANN model was successfully developed to predict the stress-strain response of Fibre Reinforced Polymer (FRP) composites. The CANNs model has demonstrated that accurate predictions can be achieved with fewer experimental results, thus reducing the experimental burden typically required to characterise strain-rate dependent behaviour of FRP composite.

Your tasks

The primary objective of this work is to advance the previous implementation of the CANNs model, and you will reimplement the model to PyTorch and perform systematic analyses to ensure reproducibility and optimise network performance. Additionally, you will enhance the interpretability of the learned constitutive behaviour, thereby helping to demystify the model's decision-making process.

• implement an existing CANN model in PyTorch and compare the results with TensorFlow implementation. 
• optimise the CANN architecture to improve model accuracy with physical laws by choosing appropriate activation functions. 
• extract physical constitutive parameters by directly mapping activation function weights. 
• improve model stability to behave in accordance with physical laws.  

Your profile

• ongoing Master's degree in the field of aerospace engineering, mechanical engineering, computer science or comparable degree programmes
• very good programming skills in the Python
• experience in the field of machine learning and data analysis
• personal responsibility, structured way of working and commitment
• proficient in English (written and oral)
  

 We look forward to getting to know you!

If you have any questions about this position (Vacancy-ID 1894) please contact:

Nathalie Toso / Sanghyun Yoo
Tel.: +49 711 6862 564 / +49 711 6862-573