Machine-learning-enabled Synthesis & Characterization of Epitaxial Perovskite Oxide Films

Abstract

Perovskite oxides are a well-studied family of materials that possess a wide range of useful properties and novel properties. SrHfO3 (SHO) is a perovskite oxide with an orthorhombic Pnma structure and a large band gap of 6.1 eV. Its large band gap and insulating behavior have made it a candidate for next-generation gate dielectrics in MOSFET devices. Previous studies utilizing density functional theory modeling have suggested that a ferroelectric P4mm phase of SHO may be possible to synthesize under strained conditions. A ferroelectric phase of SHO would be an ideal candidate for a gate dielectric in ferroelectric random access memory (FRAM) devices. In order to investigate the possibility of a ferroelectric SHO phase, SHO thin film samples were synthesized on GdScO3 (GSO), TbScO3 (TSO), and SrTiO3 (STO) using hybrid molecular beam epitaxy (hMBE). Due to the refractory nature of Hf metal, it is highly difficult to synthesize thin films containing Hf using the effusion cells in traditional molecular beam epitaxy (MBE). hMBE has emerged as a leading alternative that blends the highly tunable precision of MBE with the ability of atomic layer deposition (ALD), to easily synthesize refractory metals by introducing Hf through the metal-organic precursor tetrakis(ethylmethylamino)hafnium(IV) (TEMAH).

In perovskite oxide heterostructures and devices, samples must be highly crystalline with sharp interfaces in order to exhibit the desired electrical properties and function correctly. Successfully synthesizing perovskite oxide samples usually requires significant numbers of growth attempts and detailed film characterization on each sample to find the optimal growth window of a material. The most common real-time in-situ diagnostic technique available during MBE synthesis is reflection high energy electron diffraction (RHEED). Conventional use of RHEED allows a highly experienced operator to make some qualitative observations during growth, such as recognizing when the sample has become amorphous or when large islands have formed on the surface. However, due to a lack of theoretical understanding of the diffraction patterns, finer, more precise levels of observation are challenging. To address these limitations, new programs were created that utilize modern data science techniques including principal component analysis (PCA) and k-means clustering to analyze the recordings of the RHEED patterns taken during growth.

Although hMBE growth allows for highly tunable control of film synthesis, the downside is that there is a very large parameter space that needs to be explored to find the optimal growth conditions. Common variables include the oxygen pressure, substrate temperature, substrate material, effusion cell temperatures, and metal-organic precursor pressures. As a consequence of having this many variables and the limitation of growing only a few samples a day, fully searching the parameter space is not feasible. As a result, growths need to be carefully monitored and adjusted during synthesis to obtain the desired results. However, the only real-time feedback that is available is RHEED; thus successfully finding the optimal conditions in a reasonable amount of time and growths requires the operator to possess significant expertise in film growth. Convolutional neural networks (CNNs) have been demonstrated to be capable of identifying patterns in data at a level that human operators are unable to. To investigate the ability of CNNs to make useful predictions that can provide more advanced and detailed information than traditional monitoring systems, a CNN was trained to predict film stoichiometry by learning from X-ray photoelectron spectroscopy (XPS) and RHEED data.