Project: AI Analysis and Data Science
The data we work with ranges widely, from what experiments and measurements produce to what simulations generate. What it all has in common is that there is far too much of it for human eyes to follow.
What we work on
Our subjects run from measurement data — from the synchrotron facility NanoTerasu and from cryo-electron microscopy — to the output of simulations such as those of steam turbines. The methods are chosen to fit the nature of the data: graph-based feature extraction, classification with large language models (LLMs), and machine learning on time series. The problems range just as widely, from identifying molecular structures to detecting anomalies in machinery that cannot be measured directly while it runs.
Why it matters
As instruments and simulations improve, the data they produce grows faster than the analysis can keep up with. In fields where judgment still rests largely on expert experience, analysis becomes the factor that sets the pace of research. Making the analysis faster counts for as much as making the instruments and computers better.
Main activities
- Identifying molecular structures from three-dimensional electron diffraction data (graph-based feature extraction and LLM classification)
- Developing matching methods for electron microscopy images
- Anomaly detection and blade-wear prediction using digital twins of steam turbines
- Semantic extraction from synchrotron CT images (see the NanoTerasu Integration page)
- Predicting fault parameters with machine learning
Keywords
Machine learning / Large language models (LLMs) / Image and graph analysis / Digital twins / Measurement data analysis