Real-time Autonomous Optimization of Thin Film Growth

We have developed an autonomous experimentation pipeline to fine-tune key synthesis parameters (oxygen pressure, temperature, laser pulse rate, energy) of pulsed laser deposition with the aim to speed-up the exploration process by a factor of 10. We designed a deposition scheme that allows multiple samples to be deposited on a single substrate, improving efficiency. This work focus on optimizing growth conditions for stabilizing hexagonal LnFeO3 (Ln = lanthanides) on hexagonal substrates. These materials, as type II multiferroics, exhibit properties including non-collinear antiferromagnetic ordering, spin canting, and weak coupling between ferromagnetism and ferroelectricity, making them promising for spintronics. While prior studies show it’s possible to grow these films on YSZ (111) substrates using pulse laser deposition (PLD), a more in-depth investigation on optimal growth conditions is needed. During the deposition, a convolutional neural network would analyze the RHEED data to monitor epitaxial growth and crystallinity. Using an active learning approach with a Gaussian process regression model, and the real-time closed-loop operation runs at 15 minutes per cycle. We identified non-obvious parameter combinations that produced high-quality films of TbFeO3, YbFeO3, and EuFeO3.