Adaptive process control for crystal growth using machine learning for high-speed prediction: application to SiC solution growth

Abstract

To design a time-dependent control recipe which can ensure consistently suitable growth conditions in an unsteady growth system with dynamic environmental changes, an adaptive control method based on high-speed machine learning prediction models was proposed and applied to the solution growth of SiC crystals. This approach comprised three parts, namely, a quasi-unsteady computational fluid dynamics (CFD) model for thermal and flow field simulation, machine learning models for approximating the simulation results and giving instant prediction, and an optimization algorithm for searching the optimal growth conditions. First, the evolution of the flow, temperature and carbon concentration fields over 50 h unsteady growth following an original recipe with fixed control parameters was analyzed by CFD simulation. Then, adaptive control was applied to design a time-dependent growth process with a 100-timestep sequence. The hybrid of machine learning models and CFD simulation accelerated the entire design and optimization process by 300 times, compared with CFD simulations alone. The adaptive control facilitated superior performance compared with the fixed recipe, where the single SiC crystal thickness increased by ∼30% and the growth interface was more uniform. Further, crucible dissolution and polycrystal precipitation were suppressed by ∼50%, enabling longer growth time and more stable growth. It is the first time that the importance of adaptive control during long-term SiC solution growth is discussed, and the method proposed in this study demonstrated the potential for real-time optimization in the future.