Yujia Wang1,Guoyan Li1,Xiaoning Jin1,Swastik Kar1
Northeastern University1
Yujia Wang1,Guoyan Li1,Xiaoning Jin1,Swastik Kar1
Northeastern University1
Abstract<br/>When it comes to the predictable synthesis of 2D transition metal dichalcogenides using chemical vapor deposition, the large number of design of experiments (DoE) parameters (such as temperature, pressure, flowrate, quantity and initial positioning of precursors, pressure etc.) and their and complex interdependencies is a well-recognized experimental challenge. Often, small changes in one of these parameters can lead to large variabilities in size, yield, and quality. Although literature is replete with information regarding growth conditions, they vary from system to system. Hence, a method that can reduce the inevitable risk of “trial and error” approaches and guide experimentalists systematically towards the most suitable outcome is important for accelerating research. Here, in order to increase the efficacy and accelerate development, we present a machine learning approach to experiment design providing an alternative data-enabled materials synthesis design to establish desired process-and-quality linkage within the target material quality space. The proposed approach uses a batch Bayesian optimization-based method that employs a batch acquisition strategy to adaptively guide the selection of the experimental parameters and iteratively find the most rapid path for reaching the optimized design of experiments for a target material quality. Using the synthesis of molybdenum disulfide (MoS2) by thermal chemical vapor deposition as an example, we extracted five physically correlated experimental parameters (position of the sulfur precursor, position of MO2 precursor, Ar gas flow rate, temperature ramp-up rate, maximum temperature holding time) and three physically correlated observational parameters (Raman peak height, Excitonic linewidth, Shape). We combine an efficient and predictive surrogate model for monolayer prediction with a batch Bayesian Optimization algorithm to adaptively explore and exploit the design space. We show the success of implementing this method in MoS2 synthesis in which MoS2 with target observational parameters is achieved within 3 batches of experiments (46 samples) and the optimal region of synthesis parameters need to make targeted materials are identified.<br/><br/>Acknowledgements<br/>This research was supported by the Crystal Systems Innovations Fellowship to support Y.W. and partially supported by the Northeastern University TIER1 Seed Grant.<br/>Y.W. and G.L. contributed equally