Phase control of heterogeneous Hf x Zr (1−x) O 2 thin films by machine learning

Polymorphic Hf x Zr (1− x ) O 2 thin films have been widely used as dielectric layers in the semiconductor industry for their high- k , ferroelectric, and antiferroelectric properties in the metastable non-monoclinic phases. To maximize the non-monoclinic components, we optimize the composition depth profile of 20 nm PVD Hf x Zr (1− x ) O 2 through closed-loop experiments by using parallel Bayesian optimization (BO) with the advanced noisy expected improvement acquisition function. Within 40 data points, the ratio of non-monoclinic phases is improved from ∼30% in pure 20 nm HfO 2 and ZrO 2 to nearly 100%. The optimal sample has a 5 nm Hf 0.06 Zr 0.94 O 2 capping layer over 15 nm Hf 0.91 Zr 0.09 O 2 . The composition and thickness effect of the capping layer has been spontaneously explored by BO. We prove that machine-learning-guided fine-tuning of composition depth profile has the potential to improve film performance beyond uniform or laminated pure crystals and lead to the discovery of novel phenomena.