What Gets Detected
Anomaly detection covers three general categories:- Point and Drift
- Cross-Source
- Predictive
Unusual patterns in individual timeseries. The system encodes timeseries windows into a learned representation using a timeseries model fine-tuned on process data. Outliers in this latent space correspond to patterns that deviate from what the model has learned as normal for your process.This captures both sudden anomalies and gradual drift. Because detection operates on learned embeddings rather than fixed thresholds, it catches subtle structural changes that raw value comparisons miss.Recipe-data drift is also detected for metrology sources by comparing measured values against expected recipe setpoints.
Viewing Anomalies
On Timeseries Charts
Anomalies appear directly on timeseries plots:| Anomaly Type | Visual Indicator |
|---|---|
| Point anomaly | Marker at the detected timestamp |
| Drift anomaly | Highlighted region spanning the detection window |
| Correlated group | Highlighted region spanning the detection window |
| Forecast anomaly | Dashed region in the projected portion of the timeseries |
Anomaly Summary
Each run includes an anomaly summary with:- Counts by type and severity
- Growth quality score for the run
- Anomaly classification labels
- Correlated anomaly groups
Filtering and Querying
You can filter anomalies by:| Filter | Options |
|---|---|
| Data source | Any data source (e.g. characterization or tool state) |
| Anomaly type | Point anomaly, drift, recipe-data drift, predicted anomaly |
| Classification | Root cause category |
| Severity | Warning or error |
| Correlated groups | Only anomalies correlated across sources |
| Time range | Before or after a given time |
| Property | Specific timeseries property (e.g., pyrometer channel, RHEED spot spacing) |
| Quality score impact | How much the anomaly affected the growth quality score |
Real-Time Alerts
During streaming runs, anomaly detections are delivered in real time to any of the following channels:Platform
Alerts appear in the monitoring dashboard and notification panel during active runs.
Email notifications including anomaly details and a link to the
affected growth session.
Slack
Route alerts to create messages in Slack channels.
Custom
Configure custom alerting channels.
Per-Project Thresholds
Each project can override detection thresholds. Tighten sensitivity for critical parameters or relax thresholds for parameters with known noise. Configuration is available under project settings. Default thresholds are learned baselines calibrated from your organization’s historical data. Per-project overrides build on top of these defaults.Responding to Anomalies
During an Active Run
When an anomaly is detected during a live run, Atomscale surfaces it alongside a recommended corrective action.Review the alert and recommended action
Check which parameter triggered, its severity, classification, and detection window. The system
shows a recommended corrective action
with the expected effect and confidence level.
Check the safety envelope
Review the safety envelope: the hard constraints for affected parameters that no corrective
action can exceed. The recommended action is always within these bounds.
Approve, modify, or dismiss
Accept the recommended action, modify its parameters, or dismiss it. Approved or modified
actions trigger effects depending on integration level (e.g. assist or control). Dismissed anomalies are logged
as acknowledged with no intervention.
After Historical Data Upload
For historical data, anomaly detection provides diagnostic value. Review the anomaly timeline against your process log and ex-situ characterization to see whether detected anomalies correlate with quality outcomes. Confirmed and dismissed anomalies feed back into the classification model, and anomaly-outcome pairs contribute to growth quality score training.How Anomaly Detection Works
Detection Process
As timeseries data arrives, characterization workflows extract features and embed them into the model’s latent space. Outlier detection scores each embedding against learned baselines, then correlates related anomalies across instruments. A classification model assigns root cause categories, forecasts predicted anomalies, and routes to alert channels based on severity.For more details, see the Anomaly Detection - Technical Reference.
Batch Processing
When you upload historical data, anomaly detection runs automatically after the analysis pipeline completes. Each timeseries is evaluated independently, then cross-source correlation runs after all sources are processed. Anomalies appear in the UI once processing finishes.Live Streaming
During a live run, anomaly detection runs incrementally with each incoming data chunk. The system maintains a rolling detection context across chunks to detect both sudden anomalies and emerging trends. Forecasts and the growth quality score update with each chunk. Error-level anomalies trigger real-time alerts.Severity Levels
| Severity | Meaning | Action |
|---|---|---|
| Warning | Score elevated but within tolerance, or minor quality score impact. May self-correct. | Review after the run or if warnings accumulate. |
| Error | Score exceeds the critical threshold, or quality score has dropped significantly. | Investigate immediately. Triggers real-time alerts in streaming runs. |
Thresholds and Sensitivity
The system uses learned baselines calibrated from your organization’s historical data, adapted to the patterns of each data source, recipe, and instrument combination. Per-project threshold overrides let you adjust sensitivity: relax thresholds for parameters with natural variation, or tighten them for critical parameters.Next Steps
Diagnose Why a Run Differs
When an anomaly is detected, compare the run against references to understand the deviation.
Identify Uniformity Issues
Find consistency problems within a single run that may not appear as time-domain anomalies.
Anomaly Detection Reference
Technical details on the embedding pipeline, outlier detection, and forecasting algorithms.
Python SDK
Programmatic access to anomaly data, alerts, and control actions.