Diamond Growth AI

The synthesis of single crystalline diamond (SCD) wafers and crystals is a highly non-linear process, making effective process controls complicated. This has a detrimental impact on wafer quality and yield, and in turn increases costs.

Fraunhofer USA Center Midwest (CMW) in collaboration with Fraunhofer USA Center Mid-Atlantic (CMA) developed a series of artificial intelligence (AI) algorithms that can predict the growth stage of these diamond wafers several hours into the future. This is achieved by equipping diamond growth tools with affordable optical monitoring tools, such as a camera and thermal imager. These images, alongside the reactor process telemetry, are used as input for the AI growth prediction algorithm.

These algorithms, when integrated as a process control system, will be able to fully automate synthesis of diamond wafers, while also significantly increasing crystalline quality and yield. The hardware needed for the AI growth prediction can be easily retrofitted to existing diamond growth tools

Segmentation Algorithms

Visualization of the geometric algorithm segmentation.

Visualization of the defect detection algorithm

Two AI algorithms are being used to segment input images into machine readable labels to (i) identify geometric features of the diamond and growth chamber, and (ii) identifies various crystal defects and undesirable growth patterns that are detrimental to the final diamond wafer.

The geometric segmentation algorithm labels and measuring the corresponding image regions of interest belonging to the top and sides of the diamond, as well as the recessed pocket holder used for the diamond growth.

The defect detection algorithm labels center defects, which are predominantly crystalline defect clusters, and edge defects, which are competing (100), (110) and (111) crystalline growth directions. Further, the defect detection algorithm can also identify and classify parasitic polycrystalline diamond (PCD) growth at the diamond itself, as well as on the holder. Extensive PCD growth on the holder has a detrimental impact on the growth of the SCD wafer itself.

Prediction Algorithm

Side by side comparision of predicted and actual growth state image for a diamond wafer 6 hours into the future based on an input time series of 40 minutes of growth

The growth state prediction algorithm uses a time series of labels from the geometric segmentation and the defect detection algorithms and reactor process conditions to predict future growth states. By using input series of 40 minutes of diamond growth it was possible to predict the growth state of diamond wafers up to 20 hours into the future. This prediction included accurate prediction that of parasitic PCD growth.