Researchers at Pacific Northwest National Laboratory (PNNL) have developed a machine learning (ML) system that can detect subtle changes in thin film growth data in real time—faster than human experts—marking a key step toward fully autonomous materials research.
Machine learning programs can identify subtle changes in the structure of a thin film earlier than a human expert, enabling new opportunities to control thin film synthesis. Image Credit: Andrea Starr | Pacific Northwest National Laboratory
Thin films are essential to technologies ranging from cell phones and solar panels to quantum computers. But producing high-quality films requires precise control. These materials form slowly, atom by atom, over hours. Small variations in sensor readouts can signal defects—information that can be difficult for human observers to catch quickly. Detecting these issues early can allow researchers to intervene mid-process, potentially improving efficiency and reducing waste.
A new ML tool developed by the PNNL team, detailed in the Journal of Vacuum Science & Technology A, analyzes thin film growth data and flags changes as they emerge. It does so faster than a trained researcher, laying the groundwork for future systems capable of adjusting growth conditions automatically, without human input.
What we are doing is finding a way to make ML work for us. To make changes that can actually affect film growth, every second matters. Often, by the time we notice something’s going wrong, it’s too late to fix the film. As we train our ML program with more data, it should get even better at finding changes.
Tiffany Kaspar, Study Lead Investigator and Materials Scientist, Pacific Northwest National Laboratory
The project is part of the Adaptive Tunability for Synthesis and Control via Autonomous Learning on Edge (AT SCALE) Initiative, which supports cross-disciplinary efforts to integrate hardware, software, and domain expertise for autonomous materials research.
RHAAPsody in Films
To test the system, the team used titanium dioxide as a model material—chosen for its balance of simplicity and structural variability. Films were deposited atom by atom and monitored using electron beam diffraction imaging, captured once per second. These images contain streaks, spots, and other patterns that reflect the film’s crystalline structure and surface topography.
Traditionally, researchers would monitor these images manually, watching for deviations that indicate surface roughness or unintended structures. The new ML tool, named RHAAPsody, can perform this task automatically. It converts real-time instrument data into a format suitable for rapid analysis, compares data points second-by-second, and flags key changes known as “change points.”
Working with the film data proved quite challenging. We were surprised to find that there was so little data easily available in the community to train our machine learning models. We are planning to make our data accessible to others to hopefully accelerate more innovation.
Sarah Akers, Data Scientist, Pacific Northwest National Laboratory
In addition to flagging changes, the system includes visual tools that help researchers track film evolution more effectively, improving understanding of how structure develops during growth.
Autonomous Experimentation in the Laboratory of the Future
To validate the system, the researchers compared RHAAPsody’s performance against a human expert reviewing the same titanium dioxide growth data. The ML tool identified the same change points, but about a minute faster—an important improvement when real-time feedback is needed.
Kaspar added, “This improvement in detection time is huge for developing real-time feedback in our system.”
The ultimate goal is a fully autonomous film growth system. The team is working on enabling the instrument not just to detect structural shifts, but to respond to them—adjusting growth parameters automatically to steer the film back toward desired properties. This will rely on tight integration between the growth instrument, computer systems, and new predictive control algorithms.
“Before you can make decisions, you have to know when the branch points are,” stated Akers.
The team is working on the next step in the process, which will involve leveraging data and machine learning to modify growth conditions.
Kaspar noted, “The possibilities are endless. Imagine pairing an autonomous instrument with artificial intelligence-driven materials prediction that produces some sort of wild material we don’t know how to grow right now. The process isn’t perfect, but the opportunities are thrilling.”
RHAAPsody is a foundational step toward that vision. The team is now focused on the next phase: using ML outputs to guide active control of the film growth process.
In addition to Kaspar and Akers, the PNNL team included Henry Sprueill, Arman Ter-Petrosyan, Jenna Pope, Derek Hopkins, Ajay Harilal, and Jijo Christudasjustus. Collaborators also include Patrick Gemperline (Auburn University) and Ryan Comes (University of Delaware).
Using AI to grow next-generation materials
Pacific Northwest National Laboratory researchers are developing artificial intelligence tools to accelerate the pace of material growth experiments, and ultimately, the design of next-generation materials. Video Credit: Eric Francavilla | Pacific Northwest National Laboratory
Journal Reference:
Kaspar, T. C. et. al. (2025) Machine-learning-enabled on-the-fly analysis of RHEED patterns during thin film deposition by molecular beam epitaxy. Journal of Vacuum Science & Technology A. doi.org/10.1116/6.0004493