Helmholtz AI Consultant Team for Matter Research
Helmholtz AI aims to empower scientists to apply machine learning methods to their scientific problem domains. This goal will be achieved by fostering and stimulating collaborative interdisciplinary research projects; by leveraging similarities between data-driven solutions across domains; by integrating field-specific excellence and AI/ML prowess; by improving the quality, scalability and timely availability of emerging methods and tools and by training the current and next generation of scientists in using AI methods and tools.
These goals will be pursued beyond institute and center limits by regular funding opportunities and collaboration-as-a-service offerings as well as many more activites. At HZDR, a Young Investigator Gruppe led by Dr. Nico Hoffmann and a Helmholtz AI Consultant Team led by Dr. Peter Steinbach have been installed. Both build the Helmholtz AI Local Unit to support all scientists within the research field matter of the German Helmholtz Association. This page introduces the Helmholtz AI Consultant Team.
"The Helmholtz AI consultants team's mission at HZDR is to consult scientists primarily of the research field Matter in the application of automated data processing and knowledge extraction methods. We want to disseminate state-of-art best practises in ML and data science. With this, we hope to boost data understanding of our clients at the global academic scale in order to provide a competitive advantage. Within this mandate, we will try to advance methods or tooling in order to reduce the time investment on our as well as on our clients' side."
You, your data and us
It is our task to aid scientists with their needs to process small and big data. For this, we offer in-person consulting as well as collaborative projects, i.e. vouchers. A voucher is meant to guarantee a fair and uniform processing of projects at HZDR and other Helmholtz centers across Germany. Therefor, a voucher must comply to the following criteria:
- It should describe a feasible goal which can be achieved by state-of-the art machine learning methods and assets of artificial intelligence.
- It should describe a project, that can be concluded within a period of 2 weeks or upto 6 months.
- It should report and link to data, that can be used to train and use state-of-the art machine learning methods.
- It should define uncertainty bounds that current method obtain and a possible AI agent should improve on.
These vouchers are defined and created in collaboration with us. After that, they are submitted into a light-weight review process. The central administration of Helmholtz AI, other consultant teams as well as the central Helmholtz Office in Berlin will review the voucher and potentially approve it for action.
You have Questions or Ideas? Contact Us: firstname.lastname@example.org
- automated pipelines
- image processing
- reproducible environments
- denoising of (image) data with Deep Learning methods
- Deep Learning based inversion of inverse problems
- pattern recognition with Deep Learning
- uncertainties of Deep Learning networks
- fast inference of Deep Learning networks
Dendrite Tips in Radiograms
Hieram is interested in growing dendrites as viewed in radiograms. Data was taken at ESRF Grenoble (FR) by collaborator, Natalia Shevchenko (FWDH). In this imaging method, an x-ray beam is illuminating a flat glass cuvette containing partially solidified metal, which is recorded over time (GaIn, can freeze at room temperatures). The dendrite which builds is almost exlusively made from Indium and attenuates xrays before they reach the sCMOS imaging device (2048x2048 pixel). The Gallium-Indium material creates a structure that looks like a growing tree. Hieram wants to analyse the position and shape of the tip of the tree in the images at any given time.
The dendrite (Greek for tree-like) represents Indium in solid state and the environment is made from the binary InGa alloy. The curvature of the dendrite tip gives the characteristic length scale of the entire system. The entire system represents a model system to learn more about the solidification of metals.
Each radiogram scan takes 10-15 minutes over all tiles (called a mesh scan). The full image (full mesh scan) can be very large, e.g. a "whole sample" comes at 27760x19432 pixels of 16-bit grayscale intensities (20x14 tiles, each tile contains 1388x1388 pixels). This whole sample would occupy 1GB of disk space in an uncompressed fashion.
Currently, the noise can be reduced through time averaging inside a cropped frame around the growing dendrite edge.
|Expected Results (Goals):||
Denoise the data in order to facilitate a better/higher signal-to-noise ratio
Segment the tip of the dendrites
Orientation of Janus particles in Bright-Field Microscope Images
Yara is interested in extracting the orientation of Janus particles from her birgth field microscope images. These Janus particles are subject to her research as they can work as microscopic drug delivery vehicles. Yara is studying how these particles can be controlled by applying a magnetic field to the substrate they are in.
These Janus particles are almost perfect spheres, which have a completely black hemisphere and one completely white hemisphere. This bisectional structure stems from a metal cap deposited on one hemi-sphere which is susceptible to a magnet field in different ways. This heterogeniety is exploited when trying to control their movement with an external magnetic field. Yara is currently working with small clusters of 3 Janus particles in order to learn how to control the microscopic environment. The mid term goal of the group is to set up networks or structured clusters of these Janus particles. This work is hoped to bring them more close to a clinical application.
|Expected Results (Goals):||
Decreasing the Inference of a Beamline Ray Tracing Classifier
|Description:||Beamline raytracing is a powerful tool to understand X-ray beam propagation and to optimize beam properties given experimental requirements. However, at today’s synchrotrons and FEL beamlines one needs many components having altogether a couple of hundred parameters to fulfill these needs. This makes it impossible to map the full parameter space with traditional simulation tools. We approach this challenge with various deep learning methods. This allows, for predictions of X-ray footprints at specific positions and on the other hand, the determination of the current state of a beamline becomes accessible by simple diagnostics in combination with the neural network. Additionally, tuning the beamline to specific user demands can now be handled by the A.I. providing a high-dimensional solution in contrast to sequential beamline parameter tuning.|
|Expected Results (Goals):||
The overall goal of the project is to minimize the latency for running inference for one parameter set. Given the hardware at HZB, we want to achieve lower latencies than 0.02 seconds turn around time for such an execution. Lower latencies are always welcome.
Differentiable evaluation of high-degree multi-variate polynomials in PyTorch
The FWKT lab at HZDR develops and uses large-scale computationally expensive numerical simulations of laser- and plasma-wakefield particle accelerations in order to understand related acceleration techniques tested at light sources and medical facilities worldwide (digital twin).
Despite advanced computational hardware and software, these simulations are still very demanding. One line of research is thus to replace parts of the complex simulations with fast, yet precise ML-driven surrogate models. To this end, the FWKT actively investigates machine learning based surrogate models using the Physics Informed Neural Networks (PINN) method [^1]. This method provides explainable and precise surrogate models, which are used to learn representations of physical experiments. Besides computational speedup with respect to full simulations, these models can be adapted to new experimental setups using effective Transfer Learning methods, thereby facilitating fast deployment of digital twin models.
Initial tests of PINNs by the FWKT lab revealed that, while computationally efficient, PINNs still need large volumes of data in order to meet the lab's precision requirements. Based on recent progress in the field of multi-variate polynomial interpolation [^2], the FWKT lab started work on a very promising route to substantially decrease the data requirements of PINNs. This approach is based on a re-formulation of PINNs in terms of high-degree multi-variate Newton polynomials (Polynomial Neural Solver, poly-PINN).
Within this project, this new formulation of PINNs, for which experimental software code is available, shall be implemented in the Deep Learning framework PyTorch, which is used throughout FWKT's Machine Learning efforts as well as by HZDR's AI Consultants group (among other technologies). This implementation will be used to accelerate the future development as well as the controlled usage and deployment of this new method.
|Expected Results (Goals):||