Taking the trial-and-error out of crystal growth

Just like snowflakes, two-dimensional crystals of graphene and transition metal dichalcogenides (TMDs) come in many shapes and sizes. While these materials have exotic properties that hold promise for novel technologies, reliable growth has proved something of an art. Crystals of 2D materials can be grown using chemical vapor phase deposition (CVD) but the vast range of process variables - from evaporation and reaction temperatures, reaction times, carrier gases, flow rates, and pressure to substrate materials – leads to a similarly vast range of diverse crystals shapes depending on individual systems and even individual experimenters.

To avoid this ‘trial-and-error’ approach to crystal growth, researchers from Rice and Pennsylvania Universities are calling for a systematic methodology based on large-scale databases and machine learning [Zhang et al., Materials Today (2020), https://doi.org/10.1016/j.mattod.2020.06.012].

“Many materials including single crystals, polycrystals, and amorphous materials can be synthesized by CVD,” explains Jun Lou. “[It] is particularly useful for 2D materials because it can produce crystals with large size and high quality at relatively low cost.”

Lou and his colleagues envisage a process inspired by Japanese researcher Ukichiro Nakaya’s detailed study of snowflakes, which he called a ‘letter from the sky’. Nakaya characterized snowflakes on the basis of temperature and moisture conditions. Similarly, the unique morphology of a 2D crystal could reveal its growth ‘history’.

“Given the multitude of complexity in the CVD process variables, it is very difficult to achieve robust growth of 2D crystals with consistent high qualities,” points out Lou.

Limited mechanistic understanding of the CVD growth process and the convoluted relation between system-specific and process variables are the main challenges to the controlled growth of 2D crystals. Instead of relying on trial-and-error or experimental experience, Lou and his colleagues advocate taking a systemic approach combining modelling, acquisition and processing of image data, and machine learning.

“We suggest that crystal shapes actually encode very rich and valuable information on the growth mechanisms and conditions of 2D materials. Using ‘reverse problem solving’, [this] information could be used to inform predictive modeling and guide experiments to grow desired 2D crystals,” says coauthor Ming Tang.

The approach, which follows the example of the Materials Genome Initiative (MGI), would create morphology diagrams from experimental data and use machine learning to extract useful information about the growth process from crystals.

“The development of morphology diagrams should be a community effort,” adds Vivek Shenoy. “We need different groups to contribute ‘homogeneous’ data [and] the constructed morphology diagrams need to be tested across different labs.”

The ability to predict the 2D growth process would be highly valuable in many ways. A grower could use crystal shapes to infer growth characteristics and determine optimum growth conditions with as few experimental runs as possible. This approach to rational growth could ultimately be applied to other materials, such as ultrathin materials and 3D crystals.