Working student for Agile high-rate manufacturing and optimized inspection for small aircraf (f / m / x)

What to expect

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

Machine learning has shown promise in accurately predicting complex material properties from data. However, in safety-sensitive domains such as healthcare and the aerospace industry, the reliability of these models is crucial for decision-making process. These industries require trustworthy ML predictions and high confidence.

Neural networks (NNs) naturally carry epistemic (or model) uncertainty arising from limited data and the chosen model architecture. Unlike classical NNs, Bayesian neural networks (BNNs) treat their weights as random variables and learn a posterior distribution over these weights from the training data. Then, the output of BNNs is a distribution over predictions, reflecting uncertainty in the model.

This project aims to ensure the reliability of ML-based predictions regarding material properties by implementing deep learning methods. To achieve this, you will implement BNNs and state-of-the-art inference methods. Additionally, you will leverage domain knowledge to generate physically consistent predictions and reduce spurious uncertainty. Finally, we will establish evaluation protocols for measuring uncertainty.

Your profile

• Sytematic literature review on Bayesian statistics for reliability of NNs.
• Implement Bayesian neural networks (BNNs) for material property prediction and  compare inference techniques to approximate the posterior over model weights.
• Embed domain knowledge into the Bayesian framework to assess whether physics-informed priors improve calibration and reduce epistemic uncertainty.
• Establish evaluation protocols for model reliability by computing statistical metrics to quantify model uncertainty.

Your profile