Andrew Messecar1,Steve Durbin1,Robert Makin1
Western Michigan University1
Andrew Messecar1,Steve Durbin1,Robert Makin1
Western Michigan University1
Materials informatics has demonstrated immense utility in the design and development of synthesis routes for a wide spectrum of materials, ranging from aluminum alloys for additive manufacturing to lead titanate used in aqueous photocathodes for dye-sensitized solar cells. One synthesis route that can benefit greatly from a materials informatics approach is molecular beam epitaxy (MBE). Molecular beam epitaxy offers purity and control in the synthesis of thin-film materials and devices, but its high purity source materials and ultra–high vacuum environment can make finding the optimal growth conditions for a given material costly and time-intensive. A materials informatics approach to MBE can aid in the determination of the optimal synthesis parameters and help reduce the number of runs needed to achieve high quality samples. Recently, machine learning has been used to assist the design of perovskite oxides to be synthesized via molecular beam epitaxy as well as superconducting titanium nitride. Here, we report on initial results that build upon these reported studies by applying statistical learning to plasma–assisted molecular beam epitaxy (PAMBE) growth data to investigate optimal synthesis parameters of PAMBE-grown nitride thin films.<br/><br/>Utilizing the detailed records of over 600 PAMBE nitride thin films grown in a single Perkin-Elmer 430 system, key synthesis parameters such as substrate temperature, substrate lattice mismatch, growth duration, and RF power applied to the plasma source, are associated with the resulting crystallinity (single crystalline, polycrystalline, or amorphous) of the sample as determined by analyzing in-situ acquired reflection high-energy electron diffraction (RHEED) images. Taking the set of synthesis parameters as the predictors and the crystallinity as the categorical response, machine learning algorithms for classification and inference have been implemented to uncover insights regarding the relationship between synthesis parameters and the quality of the PAMBE–synthesized film. Algorithms were implemented using the R open-source programming language, specifically including the “tree”, “rpart” and “randomForest” packages.<br/><br/>The initial results indicate that the temperature of the Knudsen cell with the cation (gallium or indium) source material was the most significant predictor in whether an epitaxial film developed to be single crystalline, which is consistent with conventional wisdom regarding the importance of growing under metal-rich conditions. The complete data set was also divided into separate, focused data sets for gallium nitride and indium nitride, on which a random forest, classification tree and logistical regression were modeled. For gallium nitride films, agreement was found between the results of the classification tree and the random forest, which both found the gallium Knudsen cell temperature to have the greatest impact on the sample's crystallinity, followed by substrate temperature and initial nitrogen pressure. For indium nitride, the logistical regression and random forest model both indicated that the greatest impact on sample crystallinity was the initial nitrogen pressure, followed by substrate temperature and the forward power on the RF plasma source.<br/><br/>These initial results identify key synthesis parameters for obtaining single crystalline nitride thin films, however the differing results among the various statistical algorithms points to an underlying complexity that is not fully captured by these approaches. These results are helping to inform the development of a neural network architecture to be used in predicting crystallinity of PAMBE grown nitride semiconductors based on provided synthesis parameters, while the size of the data set is also being increased by incorporating synthesis and crystallinity data from the literature.<br/><br/>This work was supported in part by the National Science Foundation (grant number DMR-2003581).