Probabilistic Data Analysis and Active Learning (DAAL)

The DAAL group of the NMPP division develops and applies modern data analysis methods (ML/AI-based) and modelling tools within a Bayesian framework to investigate and understand non-equilibrium and plasma-driven processes.

The DAAL group pursues two main lines of research:

  • The first research area - 'Active Learning' (AL) - is located at the intersection of model-based inference ('Artificial Intelligence' (AI)) and data-based reasoning ('Machine Learning' (ML)) and
  • 'Plasma-material interactions' : the understanding and modelling of the interaction of energetic and/or reactive species with surfaces

Active Learning:

Until recently data analysis has been considered as a kind of post-processing step: data had been collected somewhere and had to be evaluated in the best possible manner. However, nowadays the picture is strikingly different [1]: To maximize information gain and information compression ('learning') an active interplay with the system of interest instead of passive data recording is essential. Within this 'active learning' approach the research interest is the design of optimal measurement strategies (Bayesian experimental design) for computer- and physics experiments in closed- and open-loop settings. This encompasses not only modern concepts of uncertainty quantification (UQ) of complex computer codes (e.g. plasma-wall simulations) but also learning systems, which dynamically decide which action (e.g. measurement of a specific spectral line) might yield the most informative future data based on the results from previous actions. This is adressed with physics-informed AI/ML techniques, e.g. Hidden Markov Models (HMM), reinforcement learning, neutral networks or bayesian acyclic graphs and complemented by numerical methods like Markov Chain Monte Carlo (MCMC), sequential optimization or polynomial chaos expansion. In collaboration with the geometry group of NMPP physics-informed reduced complexity models (surrogate models) for complex processes are being developed.

Plasma material interactions:

The plasma inside a fusion power plant is never perfectly confined. Some of it interacts with the walls, modifying both the plasma and the wall materials. Thus, the interaction of energetic particles (e.g. hydrogen or helium) with plasma-facing surfaces like tungsten  is of crucial importance for the design and development of a fusion power plant. Phenomena like sputtering and erosion of the surface, but also implantation and layer growth may occur and typically exhibit a pronounced non-linear dynamics. In addition most of the relevant processes are far from thermodynamic equilibrium. The studies of the DAAL group target a quantitative understanding of hydrogen retention as well as defect-forming and erosion processes on an atomistic and mesoscopic scale. The latter processes are important to define the plasma boundary conditions, ie. particle and energy influx. Here a close collaboration with other (experimentalist) groups at IPP exist, ie. PMI, PBP and PCI. For its research the group develops and operates a large set of simulation codes, like molecular dynamics (MD), diffusion-reaction models, kinetic Monte-Carlo (KMC) and MCMC codes, tools for inverse problems [2] and uncertainty quantification (UQ) [3] and rate-equation models.


Details to lectures on the research topics of DAAL are given here .


Selected publications:

von Toussaint, U.: Bayesian inference in physics. Review of Modern Physics 83, pp. 943 - 999 (2011)
Linden, W. v. d.; Dose, V.; Toussaint, U. v.: Bayesian Probability Theory: Applications in the Physical Sciences. Cambridge University Press, Cambridge (2014), 649 pp.
Smith, R. C.
Uncertainty Quantification. Theory, Implementation, and Applications
SIAM, Computational Science and Engineering Series, CS12 (2014)
Go to Editor View