A Known Problem — Without the Bandwidth to Solve It

An Atomscale customer was processing thin-film oxides within a very narrow process window, using a highly multivariate active process control system. The process engineering team had been fighting that window for the better part of a year. They were not starting from zero — within the first few months they already suspected the cause: a dynamic lag between a tool setpoint change and the wafer's true response. But pinning the lag down rigorously meant pulling scarce engineering time away from a long list of competing priorities, and the evidence was effectively locked away in large, noisy tool-state logs and in-situ metrology data the team had no clean way to extract and interpret. So the problem held: understood in outline, unsolved in practice.

When they brought it to Atomscale, the question was direct: could we isolate the impact of individual controls on the wafer using only the data they already had, without running additional wafers?

Measuring the True Material Response

Atomscale's physics-aware foundation models extract high-resolution representations of the material and process state directly from metrology and log data, making it possible to automatically analyze the downstream effect of every setpoint change over multiple runs and aggregate the response.

The aggregate tool-material trajectories generated by Atomscale demonstrated that the customer's process was overdriven (Fig. 1, blue curves) and the true material response lag was 75% larger than what the control system assumed.

With a practical way to measure the tool-material response lag, the control system could be properly tuned to deliver smoother transitions to the target state (Fig. 1, black), reducing control excursions and defects. Importantly, the result came from a reusable, automated workflow run over data the customer already had — no additional wafer runs, no lengthy design-of-experiments campaign.

Analysis in the Hands of Those Who Know the Process

Atomscale closes the gap between knowing what's wrong and having the bandwidth and tooling to fix it. Instead of routing the data to a separate analytics team with no feel for the materials, Atomscale puts the analysis directly in the hands of the people who understand the process. Working side-by-side with the customer's own materials science team, we shaped the analysis and delivered a measured result in a matter of days. The control system's assumed lag was corrected against the measured value, and the workflow was handed off as a reusable capability the customer's engineers run themselves.

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 setpoint change response analysis 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

Isolating the material response is difficult due to many competing and convolved physical changes in the growth environment. In-situ metrology generates large volumes of abstract data where very subtle changes need to be observed with consistency to build insight into the system dynamics.

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. Because that representation stays consistent from run to run, small changes remain legible against the noise, and the response to a single setpoint can be separated from everything else moving at the same time. It is the same foundation behind our real-time monitoring and automation tools — and here, it is what made the material's response time measurable at all.