Needles in a haystack

First place in IAEA’s "Materials for Fusion" challenge / Computer-simulated wall load in a fusion power plant

October 05, 2018

With many entries submitted in the competition, first place in the world-wide "IAEA Challenge on Materials for Fusion" has been awarded to scientist Dr. Udo von Toussaint and Humboldt scholarship holder Javier Dominguez from Max Planck Institute for Plasma Physics along with Dr. Markus Rampp and Dr. Michele Compostella from Max Planck Computing and Data Facility (MPCDF) at Garching.

The problem featured by the International Atomic Energy Organisation (IAEA) was tricky: The subject was the wall material for the plasma vessel in a future fusion power plant, e.g. tungsten or steel.  The material must withstand high temperatures and impact of fast neutrons and plasma particles. Since appropriate experiments are difficult and expensive, one of the reasons for this being that a burning fusion plasma does not yet exist, the material behaviour expected is being investigated instead with molecular dynamics simulations.

For the competition IAEA prescribed datasets describing the positions of every tungsten or iron atom 40 picoseconds after impact of a fast particle into the originally perfect metal crystal. The task then is to filter from the huge quantity of data what changes impact has caused. The calculations necessary thus have to track the motion and interaction of many millions of particles in order to identify in this haystack the few defects or “needles” of interest, i.e. the particles that are no longer located at the correct site after the motion cascade is triggered. “After the first wild shakeup of the atoms and after many changes of site the crystal structure can relax again”, states Dr. Udo von Toussaint, “so that ultimately just a handful of atoms are affected”. Even with high-performance computers such a complex problem can only be overcome with highly refined computational strategy.


For one of the test problems posed by the IAEA Challenge, viz. a tungsten atom with an energy of 50 kilo electron volts impacting on a tungsten monocrystal, the new method automatically identifies and characterises from more than a million atoms the approx. 50 defects caused by impact. In the coloured coding, red, for example, denotes an atom alongside a defect in the lattice of the tungsten crystal, green-coloured atoms indicate surplus atoms in the lattice, and blue clouds show low atomic density; mixed colours indicate transitions between these "pure" properties.

Who is wrong?

For one of the test problems posed by the IAEA Challenge, viz. a tungsten atom with an energy of 50 kilo electron volts impacting on a tungsten monocrystal, the new method automatically identifies and characterises from more than a million atoms the approx. 50 defects caused by impact. In the coloured coding, red, for example, denotes an atom alongside a defect in the lattice of the tungsten crystal, green-coloured atoms indicate surplus atoms in the lattice, and blue clouds show low atomic density; mixed colours indicate transitions between these "pure" properties.


The winning team’s method: they described every atom by the fingerprint of its surrounding, by means of the distances and relative orientations of its neighbouring particles on spherical surfaces in the crystal, expressing this as vectors in a 50-dimensonal space of properties. This allowed the algorithm developed by the team to compare the fingerprints of an unimpaired crystal with the dataset specified by the IAEA and identify all atoms present in a perturbed environment after 40 picoseconds.

In a second step these defects could then be classified more exactly: Because the fingerprints of known types of defects can be calculated in advance, it was possible to recognise and sort out the appropriate atoms. By a special computation method a few hitherto unknown types of defects were then also localised and classified semi-automatically. Udo von Toussaint: “By our method we can not only determine where the few needles are located in a large haystack, but also identify their type.” The large quantities of data were then visualised in close cooperation with the MPCDF High-Performance Computing, Application Support and Visualisation group, headed by Dr. Markus Rampp.

Of all the entries submitted this method was awarded first place by IAEA: “In addressing the detection and classification of unknown defect structures in a novel way, the software has clear scientific potential and affords immediate application to ongoing research”. This is in fact the case at IPP. Meanwhile the method is being applied in daily research, e.g. to investigate the influence of hydrogen on formation of defects in various materials.

The prize, endowed with 5,000 euros, was presented to the winning team on 20 September 2018 at IAEA’s headquarters in Vienna.

Go to Editor View