Data‐Driven Multi‐Objective Optimization of Large‐Diameter Si Floating‐Zone Crystal Growth

Abstract Floating Zone (FZ) silicon crystal growth is essential for high‐power electronics and advanced detection systems. The increasing pressure to scale up the process is challenging due to competing objectives. This study presents a surrogate‐based optimization framework to address Multi‐Objective Optimization (MOO) in FZ growth, considering eight relevant objectives related to productivity, geometrical and growth parameters, and crystal quality. A Deep Ensemble (DE) of Neural Networks serves as a surrogate model, trained on numerical data from a Finite Element Model (FEM). Optimization is carried out using NSGA‐II and NSGA‐III, two variants of Genetic Algorithms that explore trade‐offs between competing objectives and identify high‐performing candidate solutions. Results show that NSGA‐II outperforms NSGA‐III. The optimal solutions correctly capture known trends, such as correlations between crystal size, pulling rate, and thermal stress. A subset of the more intricate solutions is validated through new simulations, showing excellent prediction performance. However, candidate solutions must still be verified by the FEM prior to experimental validation. This framework establishes a foundation for systematic, data‐driven process optimization in FZ growth and can be extended to accelerate improvements in other crystal growth methods.