Moving to autonomous experiments: Growing thin films with machine learning

From mobile phones to solar panels to quantum computers, thin films are essential for current and emerging technologies. However, it requires control to create functional thin films. During a few hours of process, the film forms atoms by atoms. Small changes in data reading can tell researchers if something is not working. Detecting defects as quickly as possible can help scientists fix movies while they are growing up, saving time and money.

Researchers at Pacific Northwest National Laboratory (PNNL) use machine learning (ML) to identify subtle changes in cultivated films that humans don’t know about as detailed in a paper published in the Journal of Vacuum Science & Technology A.

“What we’re doing is finding ways to make ML work for us,” said Tiffany Casper, materials scientist at PNNL and lead researcher for the project. “It’s important every second to make changes that could actually affect the growth of the film. Often, it’s too late to fix the film before you realize something goes wrong. As you train your ML program with more data, it should be even better to find the changes.”

This work is a collaboration between materials and data scientists. The initiative brings together scientists from various fields to jointly develop the hardware, software and equipment knowledge needed to achieve autonomous materials science.

Rhaapsody of the movie

The project began with a growing film. The team chose titanium dioxide as their model system because of the right blend of simplicity and complexity. Materials can form several different potential structures depending on the growth conditions.

The growing film was deposited by atoms and was too thin for naked eye. Instead, the team collected images of electron beam diffraction captured every second to visualize the structure of the growing film. The captured images show stripes, speckles, and other patterns corresponding to the film’s crystal structure and surface topography properties.

Traditionally, humans monitored these patterns to track the progression of the film towards ideal smooth surfaces, problematic rough surfaces, or completely unintended structures. Now, using the ML method, the computer can automatically perform these tasks.

“Working with film data proved to be extremely challenging,” said Sarah Akers, who led the ML development work. “We were surprised at how little data is easily available in the community to train machine learning models. We are planning to make the data accessible to others to encourage more innovation.”

The ML process, known by scientists as rhaapsody, begins by converting instrumental measurements into data formats into complex, yet easier to use for rapid analysis. The algorithm looks for the point where things start to change and compares the next data from one second. These “changes” are programmatically flagged. The team worked tirelessly to determine how to clearly articulate the data to demonstrate change.

In addition to flagging changes, additional graphics-based analysis will help researchers visualize the evolution of the film, and will give them a deeper understanding of the film’s growth process itself.

Autonomous experiments in future labs

To test Rhaapsody, the team monitored the same titanium dioxide deposition data with film growth researchers and identified them when they saw the changes in the images. Rhaapsody not only spotted changes like the experts, but also flagged them about a minute faster.

“This improvement in detection time is a major factor in developing real-time feedback in the system,” Kaspar said.

The ultimate goal of this project is to create a fully autonomous film growth system. In the next stage, the instrument identifies that the film’s structure is beginning to move in the wrong direction, adapting growth conditions to counter the problem. This constant surveillance and aggressive defect mitigation is built from the connections between the equipment, computer hardware and software developed by the team. The connected system incorporates a new predictive control algorithm, a key component of the final autonomous experiment.

Rhaapsody represents an important step towards this autonomous device. “Before you make a decision, you need to know when there is a branch point,” Akers said.

The team is developing the next part of the process to change growth conditions using data and ML processes:

“The possibilities are endless,” Kasper said. “Imagine combining it with an prediction of autonomous equipment and artificial intelligence-driven materials that produce some kind of wild material that you don’t know how to grow right now. The process is not perfect, but the opportunity is thrilling.”