We develop data analysis concepts based on machine learning. In this context, we pursue several scientific and software methodological goals:
- Development of a universally valid experimental data
- Generalisation of post-processing steps such as binarization, spatial and temporal filtering of data,
- Generation of synthetic training data from experimental data using novel methods like domain randomisation.
-
- Development of our own data generation framework that can assemble a complete dataset (~10000 images) from a database of interested objects with proper annotation standards.
- Enhancement of the framework by parallelisation and addition of different types of data generation capabilities.
-
- Coupling of conventional methods with AI
-
- We incorporate AI into our existing data analysis framework:
- Image segmentation into signal and background via Instance Segmentation. Here we use, among others, state-of-the-art Deep Neural Networks such as Mask-RCNN and SparseInst.
- Optical flow detection on PIV data. We try to adapt general use SOTA models (like RAFT) to our PIV measurements to accurately infer velocity fields and displacements.
- We incorporate AI into our existing data analysis framework:
-
- Automation of the AI model updating and hyperparameter tuning
-
- Incorporation of reinforcement learning to automatically correct false detections and self-optimise over time.
- Enhancement of training process by automatically preselecting the most relevant parts of the training data by context.
-
- Development of a complete AI assisted data analysis suite
-
- All these separate AI tasks are brought together under one common roof in the form of an industrial standard software suite.
-
We develop/reuse state-of-the-art instance segmentation models to detect soot from PIV data.
Development of segmentation models that can produce refined masks for different sprays. These models are validated using reliable methods.
We aim to detect displacements and velocity gradients from our experimental data using state-of-the-art optical flow detection models. These models provide adequate generalisation with sufficient accuracy.
More details will be added soon.
More details will be added soon.
Our AI models
Our models are available in DaRUS and can be accessed from the below link:
Additional information / Get involved
If you are interested in our project, have further questions, or would like to support us through student work, internships, or thesis projects, we would be delighted to hear from you via email or phone. Contact details can be found below.
Related Publications
- Jose, B., & Hampp, F. (2024). Code for training and using the spray segmentation models. https://doi.org/10.18419/darus-4147
- Jose, B., & Hampp, F. (2024). Machine learning based spray process quantification. International Journal of Multiphase Flow, 172, 104702. https://doi.org/10.1016/j.ijmultiphaseflow.2023.104702
- Jose, B., & Hampp, F. (2023). DNN model development for experimental characterisation of multi-phase flows. In Statusseminar SimTech.
- Jose, B., & Hampp, F. (2023). Machine learning based spray process quantification. 32nd ILASS Europe Conference.
- Jose, B., & Hampp, F. (2023). Training computer vision models without labeled data using physics-informed domain randomisation. In SimTech Conference 2023.
- Jose, B., Geigle, K. P., & Hampp, F. (2023). Physics-informed domain randomisation for synthetic training data generation. In 9th WAW Workshop Machine Learning.
- Jose, B., Geigle, K. P., & Hampp, F. (2022). Supervised learning without labelling: Applied domain randomization on scientific images. In 8th WAW Workshop Machine Learning.
Contact
Fabian Hampp
Dr.Junior Research Group Leader