Atomscale's physics-aware foundation models deliver a real-time composition metric from two reference runs: a 10x accuracy improvement over the customer's best existing metric.
A Year Without a Reliable Growth Metric for Control
An Atomscale customer is building a thin-film materials stack for photonics and quantum computing, where film stoichiometry sets downstream device performance. Holding composition inside the window means measuring it during the run and feeding that measurement back into process control, rather than waiting on ex-situ characterization after the fact.
The process team had spent more than a year building a composition metric out of in-situ diffraction-based metrology. The metric existed, but it was not accurate enough to control against — it tracked composition in parts of the process window and lost it in others. They brought the question to Atomscale: could we produce a more robust measure of composition from the in-situ data they were already collecting?
An Accurate, Robust Composition Metric from Only Two Reference Runs
Atomscale's physics-aware foundation models extract high-resolution representations of the material and process state directly from metrology, and do it in real time for in-line use. That representation is a comprehensive fingerprint of the material, sensitive enough to resolve the subtle changes that correspond to differences in composition. Turning it into a composition metric tailored to this customer's application took two reference runs — where classical ML approaches typically need hundreds or thousands.
The result is a 10x improvement in accuracy over the customer's best existing composition metric: an RMSE of 0.1 against XPS ground truth, against 1.0 for the original (Fig. 1). Just as important, the Atomscale metric holds that accuracy across the entire process window instead of only part of it. That is what makes it usable for active process control, and what gives the team a real chance of catching and correcting excursions consistently.
Real-Time Feedback as a Critical Requirement for Materials Growth
As materials requirements become more demanding and applications more complex, the need for in-situ, feedback-driven process control will rise dramatically. The ability to monitor, study, and respond to the complete material state in real time is becoming critical to delivering high-yield, high-performance materials. This tailored composition metric is one of many AI-driven workflows that can be deployed either asynchronously or in real time from our library of tools.
Atomscale's Strength: Finding Subtle Signals in Noisy In-Situ Data
Physics-rich in-situ metrology has historically been difficult to exploit, because of the methods and tools available to analyze it. It generates large volumes of abstract data in which very subtle changes have to be observed consistently before they can drive a reliable feedback mechanism.
Atomscale is built for materials engineers working at that scale. Our models construct a high-resolution fingerprint of the material in process and of how it evolves in time, creating a comprehensive data layer that manages noise and deconvolves variables in real time. The result is high consistency, high sensitivity to subtle signals, and low-latency availability. It is the same foundation behind our real-time monitoring and automation tools — and here, it is what turned a metric the team could not control against into one they can.