AI methods using GISAXS scatter data as an example

Stephan Roth, DESY
Volker Skwarek, HAW Hamburg

The project is in the context of automated (pre-)classification and searchability of large amounts of experimental data with artificial intelligence methods. Two problems or research areas of DESY and HAW Hamburg are addressed:

1. Due to limited beam times, which are assigned by external reviewers, the experiment durations at DESY are very tightly timed, so that scientists generate very large data sets during a measurement campaign. For our project, these are recorded at PETRA III by several simultaneously used pixel detectors with up to 11 million pixels in the form of two-dimensional scatter images. During the campaign itself, these can only be checked for plausibility or classified into result categories. This results in loss of time and resources due to repeated campaigns as well as possible misinterpretations.
2. At HAW Hamburg, research is being conducted on person and object recognition using intrinsic identity features. However, due to the diversity of potential features, it is difficult to accurately predict the extent and expression of the features. Therefore, it makes sense for systems to use AI methods to learn through feedback mechanisms when objects are different so that they can learn additional features in addition to consciously trained ones, if necessary.

This research is currently being brought together using the example of surface structure analysis by small-angle X-ray scattering under grazing incidence (GISAXS) images in the form of a master's thesis by pre-analyzing and classifying them already during the measurement campaign and proposing them to researchers as interpretations.

The project extends this master thesis for a planned DFG application, where these identity learning methods will be generalized and implemented on an FPGA as a precursor to an artificial intelligence (AI) chip. Such an FPGA will then support AI-based feature extraction, measurement classification, and searchable identity formation using robust hashes at high speed. This project result is thus applicable to a wide range of measurement methods and campaigns at DESY.