Dimensional Stacking for Machine Learning in ToF‐SIMS Analysis of Heterostructures. Issue 3 (10th December 2020)

• Record Type: Journal Article
• Authors: Abbasi, Kevin; Smith, Hugh; Hoffman, Matthew; Farghadany, Elahe; Bruckman, Laura S.; Sehirlioglu, Alp
• Is Part Of: Advanced materials interfaces. Volume 8:Issue 3(2021)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2020-12-10
• Subjects: dimensional stacking -- heterostructures -- non‐negative matrix factorization -- principal component analysis -- time‐of‐flight secondary ion mass spectroscopy; Materials science -- Periodicals; 620.11
• DOI: 10.1002/admi.202001648
• Languages: English
• ISSNs: 2196-7350

Abstract

Abstract: Output from multidimensional datasets obtained from spectroscopic imaging techniques provides large data suitable for machine learning techniques to elucidate physical and chemical attributes that define the maximum variance in the specimens. Here, a recently proposed technique of dimensional stacking is applied to obtain a cumulative depth over several LaAlO3 /SrTiO3 heterostructures with varying thicknesses. Through dimensional reduction techniques via non‐negative matrix factorization (NMF) and principal component analysis (PCA), it is shown that dimensional stacking provides much more robust statistics and consensus while still being able to separate different specimens of varying parameters. The results of stacked and unstacked samples as well as the dimensional reduction techniques are compared. Applied to four LaAlO3 /SrTiO3 heterostructures with varying thicknesses, NMF is able to separate 1) surface and film termination; 2) film; 3) interface position; and 4) substrate attributes from each other with near perfect consensus. However, PCA results in the loss of data related to the substrate. Abstract : Dimensional stacking of datasets provides larger datasets that improve classification and elucidate chemical attributes of many heterostructures with varying parameters when applied to mass spectroscopy data. Dimensional reduction on stacked data shows with near excellent consensus the chemical attributes of the surface, film, interface, and substrate for films with varying thicknesses stacked into one dataset.