Perspectives for machine-learning applied to data-rich experiments on complex materials
Christoph Freysoldt, Max-Planck-Institut für Eisenforschung, Germany
The on-going hype of machine-learning promises that any insight can be gained by computers when you have enough data to learn from. Experiments that resolve the structural and chemical complexity of materials meet this demand: single experiments can provide gigabytes or terabytes of data, and the data is known to encode the underlying materials physics in a way that is somehow amenable to understanding. Such data is often noisy, but contains characteristic patterns in space or time that relate to the material’s performance. Reversely, machine-learning techniques are urgently needed to actually process the data in a way that extracts relevant information for our human understanding of materials. I will show examples from scanning transmission electron microscopy and atom probe tomography, and highlight how to shape and adapt existing algorithms to perform data evaluations that target useful materials science concepts.