Lifetime prediction of high-end electronics

Deep learning and the challenge of data shortage

Deep learning makes it possible to learn the representation of data in multiple abstraction layers and use it for much more than classification or regression.
These methods have redefined the state of the art in many areas, such as computer vision or natural language processing. Deep learning, which is based on neural networks, requires a large amount of data as a basis for learning the representations, as the result of the parameters that can be learned is highly dimensional. However, building large data sets in materials science is difficult, as it requires a large number of samples and complex experiments. Learning precise and robust neural networks on the basis of small amounts of data therefore proves to be challenging.

Therefore, the goal of the MaWis-AI project is to develop, implement and test novel machine learning approaches for the lifetime prediction of solders made of tin, silver, copper and possibly other components (SAC or SAC+ solders). However, due to the small amount of data, standard deep learning methods cannot be used. Instead, model- and knowledge-based AI methods such as Neural Ordinary Differential Equations (Neural ODE) are being researched and their potential evaluated as part of the research project.

Contrastive learning for the learning of good representations

Contrastive learning is a machine learning method that aims to create useful representations of data. It works by creating pairs of data points that are either similar or different. These pairs are placed in an embedding space, with similar data points close to each other and different data points away from each other. Through this process, the model learns how to extract relevant characteristics that cause similar data points to be close to each other in a high-dimensional space and allows semantic information to be captured from the data. As part of the project, contrastive learning has proven to be useful for the targeted detection of cracks and defects in ultrasound microscopy images of solder connections.

While classic supervised learning involves clustering according to component and solder types in the embedding space, contrastive learning arranges the data points according to the degree of damage in the embedding space. This results in relevant characteristics being extracted by the model. (Illustration from the publication “Contrastive pretraining of regression tasks in reliability forecasting of automotive electronics”)

Project participants and funding

Project partners

Project sponsor

Bavarian Ministry of Economic Affairs, Regional Development and Energy

Funding program

BayVFP funding line digitization

Project management

VDI/VDE Innovation + Technik GmbH