Neutron and X-ray scattering experiments have numerous applications. How can they be made more precise? Find out more in today’s Helmholtz AI project showcase ‘Artificial intelligence for neutron and X-ray scattering (AINX)’, a collaboration between the Jülich Centre for Neutron Science (JCNS) at Forschungszentrum Jülich (FZJ) and Helmholtz-Zentrum Dresden-Rossendorf (HZDR).

Could you introduce yourself, giving your affiliation, area of work, and of course, the project title?

My name is Marina Ganeva, head of the JCNS Neutron SimLab at FZJ. My main area of interest is the development of AI-assisted methods for neutron data acquisition, reduction and analysis. I am the PI on this Helmholtz AI project along with Thomas Kluge from the Laser Particle Acceleration division at HZDR. He works on large scale computer simulations for laser-solid interaction and development/analysis of respective X-ray scattering probing experiments. Our project title is Artificial Intelligence for Neutron and X-ray scattering (AINX).

In simple words, what specifically is your project about? And, how and why do you think it is a high risk, high gain endeavour?

Focus of the present project is developing AI-assisted data acquisition, reduction and analysis techniques for neutron and X-ray scattering experiments. We aim to optimise the beam time usage and accelerate the data analysis. Application of AI-assisted methods in the field of X-ray and neutron scattering is a new research field, where not a lot of experience has been gained by the community so far. Although there are a number of well established machine learning methods available, there is no evidence of their good performance for our (generally, reciprocal space) data. Here is always a risk, that development of novel approaches will be required and it can slow down our AINX project. However, our first experiences show that even imperfect methods could be a high gain for the neutron and X-ray scattering community. For example, a neural network can reconstruct desired parameters from a single Grazing-Incidence Small-Angle Scattering (GISAS) pattern in microseconds, while a human scientist usually needs days at least. Thus, 10⁴-10⁵ GISAS patterns resulting from a single experiment can be reconstructed in minutes and even live reconstruction during the experiment becomes possible. Presently, analysis of such an amount of experimental data takes years. The other example is implementation of machine learning methods for autonomous experiments at triple-axis spectrometers. Here, a GPR or RL algorithm optimises beamtime use by intelligent selection of the measurement points in the 4-dimensional (momentum transfer, energy transfer)-space considering a lot of parameters including the time needed for moving of the instrument axes. Thus, a measurement of the region of interest can be performed with better statistics in a shorter time. Moreover, such a measurement can be performed autonomously, without any need of interacting with beamline users or instrument scientists.

How important has the Helmholtz AI funding and platform been to carry out this project?

Thanks to the Helmholtz AI funding, we could hire two people (one for FZJ and one for HZDR) presently working on this project full-time. This staffing level  allows for fast progress in the methods’ development and implementation. The funding also helps in creation of interdisciplinary research teams and this is crucial for such a project. The Helmholtz AI platform allows for networking, outreach to other institutes that face similar challenges and could profit from and assist in the development of the proposed tools, and access to various resources. One important example is that we were able to organise with the AINX founding a workshop that brought together the community in this quickly emerging science field.

Figure:Illustration of the autonomous experiments of the AINX project. Postdoctoral researcher Mario Teixeira Parente (photo top left corner) works together with the instrument team (A. Schneidewind, C. Franz) and in close cooperation with G. Brandl from instrument control on autonomous data acquisition for the triple-axis spectrometer PANDA (Proton ANtiproton Detector Array; schematics left, photo right). Preliminary results are given in the middle, showing which intensities are measured in AI-supported and simple grid measurement.