Supervised learning study of Raman spectroscopy features characteristic of MOCVD–grown transition metal dichalcogenides

Analyzing the vibrational modes in Raman spectra acquired of transition metal dichalcogenide (TMD) thin films can provide critical sample information on defects, sample thickness, and coverage. Identifying the synthesis parameters that will achieve the optimal values for these sample characteristics is nontrivial; however, machine learning techniques provide a possible path towards recognizing patterns between synthesis parameters and the resulting Raman spectra features. Leveraging data from 300 growth trials, we use statistical analysis and supervised learning technologies, including tree–based algorithms, to study the relationships between gas chalcogen precursor MOCVD synthesis parameters of MoS₂ thin films and features within Raman spectra of the resulting samples. Using this approach, we identify a growth recipe that minimizes the A_1g and E_2g mode peak distance, corresponding to maximizing the monolayer coverage in the sample. We also trained a model on spectra acquired at both the center and the edge of the samples to identify a recipe that minimizes the difference between the two spectra, improving deposition uniformity. This machine learning investigation of synthesis–structure relationships can be directed towards additional features of interest within Raman spectra. Furthermore, this approach can be applied to data describing the synthesis and Raman characterization of additional TMDs, such as WS₂ and WSe₂.