A Transfer Learning‐Based Method for Facilitating the Prediction of Unsteady Crystal Growth

Abstract

Real-time prediction and dynamic control systems that can adapt to an unsteady environment are necessary for material fabrication processes, especially crystal growth. Recent studies have demonstrated the effectiveness of machine learning in predicting an unsteady crystal growth process, but its wider application is hindered by the large amount of training data required for sufficient accuracy. To address this problem, this study investigates the capability of transfer learning to predict geometric evolution in an unsteady silicon carbide (SiC) solution growth system based on a small amount of data. The performance of transferred models is discussed regarding the effect of the transfer learning method, training data amount, and time step length. The transfer learning strategy yields the same accuracy as that of training from scratch but requires only 20% of the training data. The accuracy is stably inherited through successive time steps, which demonstrates the effectiveness of transfer learning in reducing the required amount of training data for predicting evolution in an unsteady crystal growth process. Moreover, the transferred models trained with relatively more data (no more than 100%) further improve the accuracy inherited from the source model through multiple time steps, which broadens the application scope of transfer learning.

Recent Advances in the Application of Machine Learning to Crystal Behavior and Crystallization Process Control

“…With the trend of big data, migration learning is a good prospect for expansion. Dang et al 132 demonstrated the ability of migration learning for efficient prediction of 133 Modelfree control strategies are based on expert intuition and heuristics. These mainly include supersaturation control (SSC) and direct nucleation control (DNC), 134 where control is achieved through proportion integral differential (PID) algorithms.…”

“…With the trend of big data, migration learning is a good prospect for expansion. Dang et al demonstrated the ability of migration learning for efficient prediction of nonconstant crystal growth. With migration learning, the amount of required training data can be reduced by up to 80%, showing great potential for application in crystal growth engineering.…”