Identify potential causes.
Predict outcomes before processing.
Combine equipment, process, inspection and design data. Analyze potential causes through causal relationships, with supporting evidence and improvement proposals.
Beyond correlation
Correlation alone cannot establish causation
Many variables change together. Correlation does not determine which is a cause and which is an effect.
Root-cause analysis relies on expert experience
Process tuning and diagnosis often depend on the experience and trial and error of a small number of specialists.
The effect of a change is hard to predict
A changed process condition often has to be run before its effect can be assessed.
Connect the data.
Trace the process issues.
Collect and integrate manufacturing data, then use a causal graph to identify issue locations and potential causes.
Understand tool state
Estimate from causal relationships
Inform parameter adjustments
Before the wafer enters the tool,
see the likely outcome.
Virtual metrology uses equipment sensor data to predict outcomes and reduce physical measurements. Causal relationships provide evidence for the prediction.
Explore the effect of a change. Find a better condition.
Counterfactual analysis estimates outcomes under changed process conditions. Integration with a local LLM explains improvement proposals in language your engineers can use.
Recommendations remain under engineering control.
Proposed process changes are presented with their supporting evidence for engineer review. Decisions to implement changes follow the site’s existing approval and change-control process. An AI recommendation does not, by itself, authorize a change to production settings.
A pipeline from analysis to action.
Ten stages describe the path from manufacturing data to evidence-backed recommendations, engineering review and future refinement. Analysis and validation support the decision; changes to production remain subject to site approval. The exact workflow is adapted to each application.
Data ingestion
Integrate equipment, process, inspection and design data.
Factor analysis
Identify variables that may influence the target outcome.
Causal graph
Infer causal relationships between relevant variables.
Effect estimation
Estimate how individual factors affect the outcome.
Counterfactual analysis
Estimate outcomes under alternative process conditions.
Recommendation
Propose candidate process improvements for review.
Validation & refutation
Challenge causal assumptions and assess the robustness of findings.
Explainable evidence
Present the data, assumptions and estimated effects supporting the proposal.
Engineer review
Review proposed actions through the site’s approval workflow.
Learning & refinement
Use confirmed outcomes to inform future analysis and model updates.
The stages describe the overall workflow. Validation methods, human decision points and model-update procedures are defined for each deployment.
Applications
Yield improvement
Identify potential process causes of yield loss.
New-process ramp-up
Reduce iterations in process tuning.
Pre-processing quality decisions
Use outcome predictions to support processing decisions and parameter adjustments.
Feedback to design
Provide manufacturing causal insights to design teams.
Measured in BFAI PoCs
PoC using causal AI and predictive models for a new process at an advanced node.
PoC using equipment sensor data for outcome prediction through virtual metrology.
Based on BFAI PoC results. Contact us for detailed evaluation conditions.
Run it in your environment.
Connect it to your workflow.
Choose a deployment model that fits your site, and combine the platform with local LLM explanations and MLOps for operation after deployment.
Deployment
Cloud, on-premises and hybrid configurations are available, depending on your environment and requirements.
Local LLM integration
Explain causal findings and improvement proposals in language your engineers can use.
Enterprise MLOps
Support model monitoring, evaluation and controlled updates as confirmed outcomes feed back into analysis.