Pinter Pinter

Please refer to the former doctoral project:
"Quantitative description of the inner structure and interfacial properties of CoDiCoFRP" [link]

Scientific Start-up: 1st Generation - "Machine Learning in Material Science"

The research focus of this project is the application of machine learning methods for material science. These methods are particularly suitable for extremely complicated problems, which are very difficult or impossible to solve using analytical approaches. In materials science, these methods are interesting, for example, for the evaluation of noisy image data, which is ultimately a question derived from the experience during the doctorate of the junior research group leader. When analyzing volumetric image data, pores often cannot be separated from the surrounding material using conventional methods because the contrast in the noisy image is too low. However, since the human eye can separate these pores very well from the environment, this is a promising application for these methods. However, the application of machine learning methods is not limited to this alone, but is manifold.

While machine learning is usually relatively easy to use because many algorithms are freely available through open source software, it is often difficult to make a statement about the trustworthiness of the results. Some approaches, such as those based on Gaussian processes, offer the possibility of deriving a probability with which a data point belongs to a certain class. On the one hand, these procedures are very transparent, but offer only little flexibility, since the characteristic space in which learning is to take place must be defined manually. A newer method is Deep Learning, which is based on neural networks and does not require a fixed feature space, i.e. the algorithms can be trained directly without selecting specific features a priori. Deep learning is therefore very well suited for the application on image data, sound signals or other large amounts of data obtained from mechanical experiments, but require a large amount of learning data in order to be trained sensibly.

In general, machine learning procedures can also be divided into supervised and unsupervised procedures. Supervised learning is when the output is already known for each input data set. For example, speech recognition can be trained by specifying specific words and using them as speech files as inputs. Unsupervised learning does not require knowledge of the outcome. It tries to find recurring matches in signals and to combine them into clusters. In terms of speech recognition, for example, recurring words could be recognized but not assigned. A hybrid of the two methods is semi-supervised learning, in which the output only needs to be known for a few data records.