- How is this run the same or different from previous runs? Compare any run against your full process history using learned similarity embeddings, instead of parameter-by-parameter matching.
- How uniform is this run internally? Assess consistency within a single run, from whole-recipe sequences down to individual atomic layers and segment-to-segment variation.
Exploring Run Similarity
Comparing successful and failed runs is one of the most challenging tasks in process engineering. Run-to-run similarity uses learned embeddings to compare runs across their full process signatures, surfacing relationships that individual parameters wouldn’t reveal.How It Works
Select a workflow and metric
In the Explore Similarity
page, choose a metric, transformation, and time scale window for the similarity calculation.
These settings control what aspects of the process signature are used for comparison.
Explore the similarity map
A 2D scatter plot shows your data items positioned by similarity. Runs that were similar cluster
together, while runs that diverged are far apart. Points are colored by the selected metric
value.
When to Use This
Run-to-run similarity is useful when you need to diagnose an unexpected result. Instead of manually comparing numerous parameter logs, you can start from a problem run and immediately find its closest matches, both good and bad, to isolate what changed. It’s also valuable for process transfer: when bringing up a recipe on a new tool, similarity analysis shows you how closely the new tool’s behavior matches the original across the full process fingerprint rather than just a handful of setpoints.Real-Time Monitoring
During a live run, the Monitor page lets you watch a growth unfold in real time. This makes analysis actionable by following the process as it happens instead of waiting for the run to finish.Starting a Session
Start a stream
On the Monitor page, click
Monitor new growth and pick an entry point: capture a screen or
window from your browser, or stream programmatically via the Atomscale Python SDK.
Reading the Session
The live session view shows three synchronized panels:- RHEED Video: the live diffraction stream, with one panel per azimuth.
- Growth Metrics: time series of derived RHEED metrics as the growth progresses. This shows you the specific metrics that are changing.
- Tool State: instrument parameter logs during the growth, providing context for changes in the growth metrics.
Assess Within-Run Uniformity
Explore Similarity can be used with segments of a single run to assess how consistent it is internally, identifying variation between segments, layers, or across different regions of the process timeline. This catches issues that run-to-run analysis alone can miss: a growth that looks acceptable overall but has early-stage drift, mid-run excursions, or layer-to-layer inconsistency that will affect device performance.Visualize and Explore
Whenever you open a data item, physical sample, or project, Atomscale displays interactive charts for the associated data. These charts show both raw tool parameters and Atomscale’s derived metrics all synchronized on the same timeline.- Series Selection
- Multi-Chart Layout
- Live Streaming
Open chart settings to choose which series to display. Series are organized by data stream
(RHEED, optical, metrology, etc.), then by series type (e.g., specular intensity, lattice
spacing), then by individual data item. Toggle series individually or in bulk.
Next Steps
The analysis capabilities on this page (similarity scoring, uniformity assessment, and real-time monitoring) are the foundation for everything in the Act step. The same models that let you compare runs historically also power real-time anomaly detection and alerting.Act on Insights
Set up alerts, respond to anomalies, and close the loop with process control.
Workflow Reference
Technical details on how similarity, anomaly detection, and tool state workflows work under the hood.