Invariance Constraints for Computational Lithography

Injecting expert invariances into deep models (CNNs) with exact-fit constraints on critical samples.

During my internship at TSMC (Computational Lithography), I worked on bringing domain expertise into deep learning models.

Problem

In many semiconductor ML settings, there are rare-but-critical regimes where you want the model to respect known structure (physics, invariances, constraints), even if the full dataset is messy.

Approach

I developed a method to impose expert knowledge in the form of invariances by enforcing exact fit on a designated subset of samples, while retaining high performance on the broader training distribution.

What I did

  • Developed and implemented the training method.
  • Built custom training and ablation tooling to check constraint satisfaction and overall predictive performance.