Computational methods and machine learning algorithms for automatic information extraction are crucial to enable data-driven materials science. These approaches are changing materials characterization and analytics, which often require a user-specified threshold to e.g. detect structure or symmetries in structures with defects. Here, we present a machine learning-based approach that directly works on the original periodic arrangements of atoms based on a three-dimensional convolutional neural network without any transformation of descriptors. Our approach shows a high classification accuracy and tolerance to the presence of random displacements and missing atoms. Experimentally, we successfully reconstruct the ordered L12 precipitates extracted from atom probe tomography data, consistent with segmentation based on isocomposition surfaces. The convolutional layers are essential for the simultaneous identification of compositional and structural information, which also give rise to its high tolerance. Our work advances machine learning-based crystal structure identification for incomplete crystal structural data.

This work was supported by Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - Project-ID 405553726 – TRR 270 and the funding from the DFG contract GA 2450/2–1. Y. Li acknowledges the research fellowship provided by the Alexander von Humboldt Foundation. This research received funding from BiGmax, the Max Planck Society's Research Network on Big-Data-Driven Materials Science.