Machine learning model for Nanosheet FETs that predicts device performance (ION, IOFF, Cgg) in milliseconds, enabling faster design and optimization replacing traditional TCAD tools.
I worked on building a faster alternative to TCAD simulations for nanosheet FET devices using machine learning. Instead of running full physics-based simulations every time (which can take a long time), the idea was to train a model that can predict device behavior almost instantly.
How I approached it
Since I didn’t have access to real TCAD data, I created a synthetic dataset that follows basic device physics trends. I varied parameters like:
- corner radius
- gate length
- oxide thickness
- doping
- bias conditions (Vgs, Vds)
and generated corresponding outputs like current and capacitance.
I trained an XGBoost regression model to learn the relationship between these inputs and outputs.
The model predicts:
- ION (drive current)
- IOFF (leakage current)
- Cgg (gate capacitance)
The goal was to capture the nonlinear behavior between geometry and electrical performance.
Once trained, the model can predict results in milliseconds instead of minutes like TCAD. This makes it much easier to explore different device configurations quickly.
I added a simple optimization loop to find the best device parameters. For example, I looked at maximizing the ratio of ION to Cgg, which reflects a common tradeoff in device design.
I built a small Streamlit app where you can:
- change device parameters
- instantly see how performance changes
This makes it easier to interact with the model instead of just running scripts.
The main idea is to speed up the design process. Instead of running multiple expensive simulations, you can use a trained model to quickly:
- estimate performance
- compare designs
- narrow down good configurations