Using Image Based Artificial Intelligence Growth Prediction to Improve SCD Wafer Dimensions and Growth Yield

Single crystal diamond (SCD) is an attractive wide bandgap semiconductor material for a variety of applications ranging from advanced optics, solid state electronics to thermal management solutions. While lab-scale prototype demonstrations have demonstrated superior performance, utilizing diamond has been prohibitively expensive compared to more mature, commercialized semiconductor materials. This is attributed to the fact the SCD wafer size of suitable quality is smaller than 1 inch, and often as little as small as 3 mm x 3mm. For comparison, the commercially available wafer size for single crystalline Silicon Carbide is beyond 6 inches. A direct result is a lack of process scaling when processing individual diamond applications. The only way to overcome this is by controlling the size and quality of SCD wafers during chemical vapor deposition (CVD) growth to enable realization of wafers 2 inches in size and beyond either. This type of process control can be applied either for epitaxial layer outgrowth, or for tiled wafer growth. We are envisioning to address this challenge by developing an artificial intelligence (AI) based algorithm to predict SCD growth states (size and quality) through use of in-situ RGB images. Once established, this AI growth prediction can be incorporated as control system to increase SCD wafer size and quality, by adjusting process conditions before they achieve critical turnover points.<br/><br/>We are reporting on our efforts on AI algorithm development and validation. First, we installed a full-frame mirrorless interchangeable lens camera equipped with a macro lens to a CVD diamond reactor. Then a cumulative image-based AI pipeline consisting of three inter-connected thrusts was developed to model diamond growth. These are 1) Feature extraction pipeline, for extraction of geometrical features in the recorded imaged, 2) Defect detection pipeline, to extract macroscopic defect features in the recorded images, and 3) Frame prediction pipeline utilizing features, defects and reactor telemetry, to predict future image states 6, 8 and 12 hours into the future.<br/><br/>The objective for the feature extraction pipeline was to isolate and classify accurate pixel masks of geometric features like diamond, pocket holder and background, and their translation into geometrical shapes without the need of human-generated input. Our approach was enhanced to deliver results with high precision within the constraints of being limited to low-volume high-feature-complexity training dataset environments, given that data procurement, requiring physical SCD growth, is extremely time-consuming and expensive. Our best performing DL-based model achieved excellent accuracy metrics of &gt;98% for the pocket holder, diamond top and diamond side features. Similarly, our DL-based defect detection pipeline achieved excellent accuracy metrics (&gt;95%) for detecting center, polycrystalline and edge defects.<br/><br/>Prediction accuracies of ~99.9999% were obtained with minimal information loss between predicted and actual outputs. This constitutes a never-before obtained result of spatiotemporal (AI) prediction of diamond shape from in-situ obtained growth data obtained based on few inputs from data collected within an hour apart. This demonstrates the potential of these algorithms as a machine-intelligence-enabled solution for automated optimization and control of the diamond growth process.