Offers and Expertise
At HZDR, we aim to drive positive economic and social impact through our research and innovation efforts. Beyond advancing science, we invest in infrastructure and talent development, providing benefits that extend to the local industry.
Offer: Are you a company in need of AI expertise? Are you interested in a collaboration?
Please get in contact with:
Innovation Manager Open Hardware, Software & DataDr. Vladimir Voroshnin
Technology Transfer & Innovation
v.voroshninhzdr.de
Phone: +49 351 260 3944
Competencies at the HZDR – Methods with AI and ML
Data analysis and signal processing
Digital twins (surrogate models)
Pattern recognition and classification with ML methods
Our data sets in the field of AI
AI is a valuable tool that boosts scientific research and development. Some AI methods help save time and resources, while others make previously impossible tasks achievable. The choice of AI methods depends on the project's specific needs. At HZDR, we currently use the following AI-based methods to support our work:
Data analysis and signal processing
Animation HIBEF
Source: HZDR / Science Communication Lab
Examples:
- CASUS analysis on photons for the experiments at the Helmholtz International Beamline for Extreme Fields (HIBEF)
- Neural network as a cluster tool for data characterization (radiochemistry in Rossendorf: pH, temperature, concentration…), Contact: Dr. Andre Rossberg
- Denoising of data (images) for HIBEF and Institute of Fluid Dynamics, Department of Information Services and Computing (remove background noise to improve the contrast of images/videos/…)
- Anomaly detection in accelerator monitoring data (detector/beamline control), Institute of Radiation Physics and Department of Information Services and Computing
Digital twins (surrogate models)
A digital twin is a detailed digital representation of a complex real-world system. It is designed to monitor, diagnose, and predict physical processes within that system. To accomplish this, a digital twin integrates several key elements: data streams from various measurements and simulations, high-performance computing (HPC), mathematical and numerical models of subsystems across different time and length scales, data-driven surrogate models, machine learning methods for real-time analysis and prediction, and interactive, intelligent visualization techniques.
Surrogate models of neural networks in high energy physics
Source: CASUS/2021 CMS Collaboration
Examples:
- Materials Learning Algorithms (MALA): A data-driven framework to generate surrogate models of density functional theory calculations based on machine learning for materials simulations on length scales inaccessible with conventional electronic structure methods: Contact Dr. Attila Cangi from CASUS
- Digital twins based on deep process in metallurgy (Institute of Radiation Physics)
- Detector control (interpolate detector responses as AI can predict quicker than simulation can calculate)
- Detector response (time series forecasting with AI to predict where detector/machine next)
- Neural networks for phase retrieval at scattering experiments
- Neural networks for solving PDEs
- Electronic Density Response in Warm Dense Matter: Contact Dr. Tobias Dornheim from CASUS
Pattern recognition and classification with ML methods
Machine learning (ML) is a branch of artificial intelligence that focuses on developing algorithms and statistical models that enable computers to learn from data without explicit programming. ML is primarily concerned with identifying patterns and relationships in data, allowing machines to make predictions, make decisions, and gain valuable insights.
Electromagnetic fields excite oscillations in a magnetic vortex. The nonlinear interaction is similar to the interplay between neurons and synapses in the brain and can be used for pattern recognition.
Source: HZDR / H.Schultheiß
Examples:
- Exploration of mining facilities with drones
- Assistance with recycling
- Real-time object tracking with motor control
- Nonlinear Magnons for Reservoir Computing in Reciprocal Space, project: NIMFEIA: Contacts are Dr. Katrin Schultheiss, Dr. Attila Kakay and Dr. Helmut Schultheiss
- Computational Fluid Dynamics: Identification/localisation of bubbles for characterisation of multi-phase flows (Institute of Fluid Dynamics and Department of Information Services and Computing), Contact: Dr. Dirk Lucas
- ROFEX: Ultrafast electron beam X-ray computed tomographie for non-invasive analysis of multiphase flows, object tracking in real time
- Identification/localisation of tumours for radiotherapy (OncoRay and Department of Information Services and Computing)
- Characterisation/regression of image properties (Institute of Radiopharmaceutical Cancer Research)
Decision-making support
Symbolbild Where2Test
Source: CASUS / M. Bajda
Examples:
- OPTIMA research consortium for cancer research
- PIONEER: The European Network of Excellence for Big Data in Prostate Cancer
- Where2Test: Assistance in the coronavirus pandemic
- OncoRay: Prediction of treatment outcomes from multi-dimensional input
- Uncertainty quantification for AI predictions (Department of Information Services and Computing)
Deep learning related
A subfield of machine learning is deep learning, which uses artificial neural networks to solve complex problems. These networks are trained to automatically learn hierarchical representations of data and thereby extract complex patterns and features from the input data. The difference between machine learning and deep learning is that while the former encompasses a broader range of techniques, decision trees, support vector machines, and clustering algorithms, deep learning refers specifically to using deep neural networks with multiple layers. This allows complex, unstructured, and large data sets to be tapped for text and image recognition.
Snapshot of a deep learning simulation
Source: HZDR / CASUS
Examples:
- Uncertainty quantification for AI predictions
- Model benchmark comparisons in a statistically rigorous fashion
- Two-sample tests for quality assurance of generative models
Self-supervised learning
Self-supervised learning is another aspect of machine learning. This quasi-autonomous form of learning uses artificial neural networks trained with a simple preliminary task. Subsequently, the learning module can classify and solve similar but more complex problems. Self-supervised learning is characterized by its ability to act independently of previously annotated data. It is already widely used in image, video and audio processing.
Our data sets in the field of AI
- Multimodal imaging medicine
- CASUS: OPTIMA
- Institute of Radiopharmaceutical Cancer Research: PET scans
- PIONEER: The European Network of Ecxellence for Big Data in Prostate Cancer
Spread all across Europe, databases from clinical studies, public registries and electronic health records contain clinical data from thousands of prostate cancer patients. PIONEER collects, anonymizes and assembles these diverse data sets. CASUS takes over the task of providing a new centralized data and analytics platform for PIONEER. The cloud-based platform will provide data access and machine learning analytics capabilities for both academia and industry researchers. PIONEER operates both a central and federated model of data sharing. For the federated model, CASUS will take on the challenge of establishing a federated analytics network.
- Commodity exploration data
- Helmholtz Institute Freiberg for Resource Technology (HIF): drone recordings and processed data
3D image of rare earth minerals (red) in a carbonate rock
Source: Dr. da Assuncao Godinho, Jose Ricardo
Traditional techniques applied to raw material exploration are costly, time-consuming, and leave a significant environmental footprint. That is why, a particular focus of our research is to develop comprehensive and fast drill core mapping routines to extract the most information out of any sample taken. At the same time, we are investigating low environmental and societal impact exploration technologies to decrease the amount of drilling needed. Another cornerstone of our research is the development of models that take into consideration all relevant raw material characteristics in an effort to predict the outcomes of mining, mineral beneficiation and metallurgical extraction already at an early exploration stage. We do so by advancing the concept of predictive geometallurgy. We develop new characterization techniques and multi-sensor systems in combination with data processing routines for a detailed, fast, and cost-efficient identification of complex raw materials.
- Spectral analysis
- Material database
- Institute of Ion Beam Physics and Materials Research (FWI): AFLOW from Duke University
Any technology relies on a specific materials platform. As such, information technology is based on silicon and modern batteries are made of lithium compounds. The "Autonomous Materials Thermodynamics - AutoMaT" lab focuses on the self-guided, data-driven computational design of two-dimensional (2D) materials and high-entropy compounds for applications in information technology and the energy sector.
- Thermodynamic reference database
- Institute of Resource Ecology (FWO): THEREDA
THEREDA (Thermodynamic Reference Database) is a joint project dedicated to the creation of a comprehensive, internally consistent thermodynamic reference database, to be used with suitable codes for the geochemical modeling of aqueous electrolyte solutions up to high concentrations.
- Sorption database
- Institute of Resource Ecology (FWO): RES³T