Computer vision assisted decomposition analysis of atom probe tomography data

Abstract

Abstract While atom probe tomography (APT) offers 3D spatially-resolved compositional characterization at the nanometer scale, individual inspection of large datasets is time demanding and the interpretation thereof is highly operator-dependent. Recently, the decomposition behavior of (V,Al)N thin films was investigated by APT. The conventional analysis, employing standard tools from commercial software, is laborious and only a small fraction (below 5%) of the total acquired data has been used. In the present work, the analysis workflow is automated, taking into account more than 90% of the total acquired data. As a first step, the decomposition products, including matrix and enriched phases, were identified via deep learning-based semantic image segmentation. Performance of the ‘matrix’ and the decomposition product ‘aluminum-rich’ was very good, achieving F1-scores of 0.994 and 0.943, respectively, while performance for ‘vanadium-rich’ was with 0.854 slightly lower. The segmentation enables the investigation of phase formation evolution over temperature. As a next step, early stages of the spinodal decomposition were probed on the ‘matrix’ phase by a neighborhood analysis. While results were consistent with prior work, the data analysis in this work is statistically more robust by leveraging an order of magnitude larger dataset. In the end, the transferability of the presented phase segmentation workflow of the (V,Al)N-based model is shown and discussed for the decomposition of isostructural (Ti,Al)N.