Predictive Doping and Thickness Analysis of a Multi-Wafer SiC Warm-Wall Epi Reactor for Improved Layer Cpks

Rapid progress in the growth of 4H-SiC epitaxial layers allow device scientists/engineers to tighten the specifications of doping and thickness uniformities of SiC epitaxial films. Further, reducing the cost of SiC epitaxial layers is a continuing goal. A compelling approach is to choose a multi-wafer warm-wall epi reactor which has been shown to have very high wafer throughput. The precursors decompose upon heating by passing over hot reactor components, however, the precursor molecules crack before reaching the substrate and can form parasitic SiC coatings. Such coatings change the emissivity of reactor parts, changing their temperatures. The allowed vapor pressure in the gas phase is also a function of the chemical composition of these deposits. Consequently, the effective Si/C ratio at the wafer varies the nitrogen incorporation efficiency on the SiC epitaxial wafer. In this paper, we have reported an approach on how to minimize the effect of changing Si/C ratio on absolute layer doping and thickness over the full campaign. We analyzed the data, identified the pattern, and have used it to make predictions or decisions to keep the deviation within control limits. The nitrogen incorporation was analyzed as a function of cumulative coating on the reactor parts. The derived models were used to make the decisions for predictive doping by adjusting the flow rates of nitrogen precursors during upcoming campaigns at specific cumulative thickness of reactor parts coating. The same approach was also used for the adjustment of growth time to obtain the targeted epi layer thickness as a function of cumulative coating. Consequently, the predictive doping control resulted in the improvement of doping Cpk from 0.37 to >1.67 and the predictive thickness control resulted in the improvement of thickness Cpk from 0.75 to 1.61. This implies that the process is six sigma qualified and expected overall nonconformance was 0.001% for doping. Moreover, the average 200 source contrast projected 5×5 mm 2 chip yield using a Lasertec system 88-HIT and the machine learning based PLDLZ recipe was >94% by considering the Particle, Bump, Micropipe, ComplexSF, Polytype Inclusion, Particle Inclusion, and ScratchTrace as device killer defects. The average BPDs were <25 on 150mm wafers using a 1µm thick buffer layer. Initial results on 200 mm wafers are also presented.