講演情報
[23a-P06-16]Application of machine learning to thin film fabrication
〇YUNLONG SUN1, Haotong Liang3, Mikk Lippmaa2 (1.GSFS UTokyo, 2.ISSP UTokyo, 3.Maryland Univ)
キーワード：
machine learning、thin film
We have developed a high-speed film fabrication method to collect structure data for machine learning quickly. In this method, up to 12 thin film samples are grown sequentially on a single 15-mm-wide substrate while analyzing the film surface structure by in-situ real-time electron diffraction (RHEED). After each film growth, a neural-network-based image segmentation algorithm developed by Liang is used to analyze the RHEED patterns and derive a quality metric (such as diffraction intensity or feature sharpness). Bayesian optimization is then used to predict the process conditions for growing the next film on the same substrate. The process is repeated until no better point can be found in the parameter space. We have used this method to optimize the deposition conditions of ReFeO3 (Re=Tb, Eu, Yb) films. Several process parameter mapping experiments were conducted to validate the Bayesian predictions and to find phase boundaries of ReFeO3 films in the available process parameter space.
The optimization experiments all converged, successfully finding the optimal deposition parameters for TbFeO3, EuFeO3, and YbFeO3. As for the EuFeO3, the Bayesian optimizer obtained films with sharp diffraction streaks, indicating high crystallinity and surface flatness. Sharp diffraction spots and even bulk-like Kikuchi lines were obtained for TbFeO3 and YbFeO3 films. The correctness of the Bayesian predictions was validated for TbFeO3 in mapping experiments. Surface morphology analysis showed that there is a clear correlation between the growth oxygen pressure and temperature.
The optimization experiments all converged, successfully finding the optimal deposition parameters for TbFeO3, EuFeO3, and YbFeO3. As for the EuFeO3, the Bayesian optimizer obtained films with sharp diffraction streaks, indicating high crystallinity and surface flatness. Sharp diffraction spots and even bulk-like Kikuchi lines were obtained for TbFeO3 and YbFeO3 films. The correctness of the Bayesian predictions was validated for TbFeO3 in mapping experiments. Surface morphology analysis showed that there is a clear correlation between the growth oxygen pressure and temperature.