ASML is a high-tech semiconductor company that manufactures complex lithography machines. Thousands of ASML lithography machines are operating 24/7 throughout the world at chip factories, to produce integrated circuits (IC). Due to the extreme accuracy of these machines, problems or failures in their operations occur regularly and eventually this would cause downtime. Fast diagnosis of failures is a critical step in maintenance of an ASML equipment, as downtime in high-end chip factories is extremely costly and it can cost $20 per second of unscheduled downtime. When a machine has a failure, a service engineer must be able to identify the source of the problem. That is why this is crucial to help steer engineers in the correct direction. However, this sometimes can be a time consuming and knowledge intensive task. Therefore, it is necessary to research and discover a methodology that addresses those issues and helps engineers discover the issue in a more efficient manner.
This project explores probabilistic reasoning as a method for modeling and interpreting uncertainty in complex systems. Using Bayesian Networks, it demonstrates how probabilistic graphical models can capture dependencies among variables and support informed decision-making, particularly in domains like semiconductor manufacturing, where complex diagnostics are common. The main focus is on applying and analyzing probabilistic inference in the context of machine troubleshooting workflows, modeling how engineers identify faults and decide on the next diagnostic steps.
Link to the thesis:
- Master Thesis paper master_thesis_Deyna_Baeva.pdf
- Python – Core language for modeling and experimentation
- pgmpy – Bayesian network construction and inference
- pandas, NumPy – Data preprocessing and structure
- matplotlib, seaborn – Visualizing networks and inference outcomes
- Jupyter Notebook – For step-by-step development and explanation
- Bayesian Network Construction – Manually defined network structure based on domain knowledge of troubleshooting processes.
- Inference Algorithms – Applied exact inference methods (e.g., variable elimination) to simulate reasoning under partial evidence.
- Scenario Simulation – Modeled realistic troubleshooting sequences to evaluate how inference supports step-by-step fault diagnosis.
- CPT Definition – Conditional probabilities were specified manually to reflect realistic causal relationships in the network.
- Uncertainty Handling – Demonstrated how probabilistic models deal with incomplete or ambiguous evidence during diagnosis.
- Engineer Troubleshooting Workflow Modeling
- This case study models the diagnostic reasoning process of engineers fixing semiconductor equipment. A Bayesian Network was manually designed to represent the logical sequence of symptoms, faults, and diagnostic questions. Inference was used to simulate real-world troubleshooting scenarios and identify the most probable causes and next best actions.
- Demonstrate how Bayesian Networks can support diagnostic reasoning under uncertainty
- Simulate realistic engineering scenarios to evaluate model behavior
- Create interpretable models that reflect domain expert logic and decision flow