Alerts and Anomaly Detection
Atomscale’s alerting system operates on our derived intelligence layer that holistically profiles growths, allowing them to catch problems that are invisible to conventional monitoring that operates on fixed combinations of sensors.Technical Dive: How Atomscale Alerts Differ
Technical Dive: How Atomscale Alerts Differ
Traditional alerting operates on tool state: temperature out of range, pressure spike, flow rate deviation. Atomscale alerts operate on materials state: the derived metrics that describe what’s actually happening to the film being deposited.
| Traditional Alerting | Atomscale Alerting |
|---|---|
| Temperature exceeds setpoint | Predicted composition drifting from target |
| Pressure outside range | Surface reconstruction deviating from expected sequence |
| Flow rate drop | Run trajectory diverging from process fingerprint library |
| Manual RHEED observation | Quantitative lattice spacing or spot count anomaly |
Atomscale’s anomaly detection has demonstrated greater than 96% accuracy and can flag issues 40 seconds earlier than manually monitoring the same process visually.
Types of Alerts
- Anomaly Detection
- Drift Tracking
- Threshold Alerts
Atomscale continuously scores each run against your process fingerprint library. When a run’s trajectory diverges from expected behavior, an alert fires, even if every individual tool parameter looks normal.This is particularly valuable for catching subtle defects like compositional drift or surface quality degradation that don’t manifest in any single sensor reading.
Setting Up Alerts
Choose an Alert Type
Select from anomaly detection (automatic, model-driven), drift tracking (cross-run trending), or
threshold-based (explicit limits on derived metrics).
Define Scope
Specify which data streams, runs, or recipes the alert applies to. Alerts can be scoped to a
specific tool, material system, or recipe family.
Set Severity and Notification
Assign a severity level (Info, Warning, or Critical) and configure how you’re
notified: in-app, email, or webhook for integration with external systems.
Respond to Anomalies
When an alert fires, the next step is deciding what to do. Atomscale supports a range of responses depending on the severity and your integration level.Operator-in-the-Loop
For most teams starting with Atomscale, the initial workflow is operator-assisted: Atomscale detects and surfaces the issue, and a human decides what to do.Investigate
Click through from an alert to the Analyze view. Compare the flagged run against reference runs
to understand exactly what diverged and when.
Annotate
Log observations and decisions directly on the run timeline. These annotations feed back into
your process library for future reference.
Adjust
Use the alert context to decide on recipe adjustments: modify setpoints, extend or shorten
growth stages, or compensate for detected drift.
Abort
For critical anomalies, decide to terminate a run early rather than wasting time and materials
on a doomed process.
Run Disposition
After a run completes, Atomscale’s analysis helps with go/no-go decisions before committing to downstream processing or characterization. Rather than waiting days or weeks for ex-situ measurement results, you can use in-process predictions to triage runs immediately.Integrating Process Control
For teams ready to move beyond operator-assisted monitoring, Atomscale supports progressively tighter integration with AI-driven process control.Integration Levels
Monitor and Alert
Atomscale observes and notifies. All decisions and actions are made by operators. This is where
most teams start.
Recommend
Atomscale suggests specific adjustments based on detected deviations. For example, recommending
a flux correction when predicted composition drifts from target. Operators review and execute.
Assist
Atomscale sends adjustment commands to the process controller via SECS/GEM or tool API, with
operator approval required before execution.
Agent-Based Process Control
At the most advanced integration level, Atomscale’s AI agent acts as a process copilot. The agent combines real-time derived metrics with the full process fingerprint library to make context-aware decisions. Key capabilities include:- Real-time recipe adjustment: Modify growth parameters mid-run to correct for detected deviations, such as adjusting source temperatures to maintain target composition.
- Natural language customization: Configure agent behavior and decision context using natural language instructions, adapting the scope and sensitivity of automated control without code changes.
- Audit trail: Every automated action is logged with full context: what was detected, what the agent decided, and what command was sent. This provides complete traceability for process validation.
Track and Review
Acting on process data is only valuable if you can measure whether those actions improved outcomes. Atomscale provides tools to close the learning loop.Run Reports
Each run generates a summary that includes:Derived metrics timeline with key events and transitions highlighted
Comparison to target values and reference baselines
Any anomalies detected and actions taken (manual or automated)
Annotations and operator notes
Process Improvement Tracking
Over time, track how interventions, whether manual adjustments informed by Atomscale or automated corrections, affect yield and process stability. This historical record is what transforms individual run insights into systematic process improvement.Next Steps
You’ve completed the getting started guide! Continue exploring:Detect Anomalies
Detailed guide on configuring anomaly detection for active runs.
Diagnose Run Differences
Compare runs and understand what changed when results differ.
Characterization Workflows
Set up monitoring and analysis for specific characterization methods.
Python SDK
Programmatic access to Atomscale’s analysis and control capabilities.