Efficient optimization of crystal-growth conditions is often hindered by variations in furnace design, which require reoptimization whenever the configuration changes. We propose a method to transfer an optimized growth environment from one furnace design to others by representing complex thermal-flow distributions through a variational autoencoder. In this approach, the optimized reference environment is represented as a target point in the latent space constructed by the VAE, and the growth conditions are optimized to reproduce an environment close to this target point. A surrogate model of computational fluid dynamics simulations and a genetic-algorithm-based multiobjective optimization are combined to rapidly identify suitable process conditions. The method was applied to the top-seeded solution growth of 200 mm-diameter 4H-SiC, successfully reproducing and further improving the optimized environment obtained in the reference furnace. The results demonstrate that the proposed framework enables efficient transfer and enhancement of growth conditions across different furnace systems.
Sakamoto et al. (Tue,) studied this question.