Data Analytics & Machine Learning for Manufacturing
Define
In high-precision industries such as semiconductor manufacturing, system reliability and component accuracy are mission-critical.
-
Manufacturing processes generate massive sensor data, but distinguishing random variability from actionable deviations is challenging
-
Small misalignments, such as overlay errors in lithography, can cascade into costly system failures
-
Run-to-Run (R2R) control loops monitor and correct deviations, yet traditional methods struggle to detect subtle systemic drifts
​
Process engineers requested a predictive, data-driven solution to anticipate deviations, improve yield, and enhance system reliability.
Goal
Develop a data-driven framework that:
-
Monitors and predicts overlay errors in R2R control loops
-
Transforms raw sensor data into actionable insight
-
Enables proactive process interventions to reduce variability and prevent failure
​
Mission: Create a predictive, closed-loop control strategy that ensures optimal precision and performance in high-stakes manufacturing environments.
Approach
Empathize
-
Studied complex manufacturing processes and R2R control workflows
-
Analyzed high-frequency sensor outputs, including thermal and vibration telemetry
-
Identified challenges in separating noise from meaningful system deviations
​
Define Core Engineering Problem
Predict and correct overlay errors in a Run-to-Run loop by leveraging statistical analysis and machine learning to extract actionable signals from complex, high-dimensional sensor data.
​
Investigation Goals / Design Requirements
-
Detect subtle temporal and correlated deviations that traditional SPC methods miss
-
Enable predictive intervention without destructive testing
-
Integrate seamlessly into R2R control loops to improve real-time process performance
-
Quantify and reduce expected yield loss and metrology variability
​
Ideate / Hypothesis Generation
-
Explored statistical and machine learning approaches to separate noise from actionable deviations
-
Evaluated univariate vs. multivariate control strategies
-
Considered physics-informed models for overlay prediction
​
Prototype / Experimental Implementation
-
Ingested, preprocessed, and analyzed sensor data using Python and MATLAB
-
Applied signal processing and spectrogram analysis to detect subtle patterns
-
Implemented univariate control charts (X-bar & Range) to monitor individual process variables
-
Applied PCA for multivariate monitoring:
-
EWMA Control Charts to track correlated sensor variations
-
Variational EWMA Charts for low-amplitude drift detection
-
-
Developed probabilistic failure models to quantify yield and metrology reliability
-
Applied virtual metrology via machine learning (Ridge regression with 5-fold cross-validation) to predict critical overlay metrics without destructive inspection
-
Modeled time-series ARMA processes to forecast systematic drift in process inputs
​
Test
-
Integrated predictive outputs into R2R loop for closed-loop correction
-
Evaluated reduction in overlay errors and process variability
-
Validated predictive accuracy and early detection capabilities against historical data
​
Iterate
-
Tuned hyperparameters for optimal predictive performance
-
Refined model integration into R2R workflow
-
Optimized visualization and alerting to guide proactive adjustments
Impact
-
Early detection of overlay drift enabled preemptive corrections and prevented downstream defects
-
Improved prediction of critical metrology outputs using physics-informed machine learning
-
Reduced process variability, increased yield, and enhanced system reliability
-
Demonstrated a scalable, data-driven strategy applicable to high-stakes, precision engineering and aerospace reliability systems
Key Takeaways
-
Shows ability to combine classical statistics with AI/ML to turn complex sensor data into actionable insight
-
Demonstrates predictive, proactive control strategies that improve reliability, yield, and precision
-
Highlights skill in integrating data-driven models into real-time, high-stakes control systems
-
Emphasizes evidence-based decision-making to prevent failures before they occur
​
This project reflects my approach to engineering: analyze complex data deeply, engineer predictive solutions, and integrate insights into real-world control systems for measurable impact.
​