TL;DR
Researchers have developed a way to drastically speed up semiconductor "nanometrology" by combining physics-based simulations with deep learning. By training a neural network on a small set of expensive Monte Carlo simulations, they can predict electron microscope scan profiles in milliseconds rather than minutes, turning a three-month computational task into a one-minute job without sacrificing accuracy.
What Problem Does This Paper Address?
In the world of microchips, "Critical Dimension" (CD) refers to the smallest features on a circuit. Measuring these features accurately is vital for quality control. The industry standard tool is the CD-SEM (Critical Dimension Scanning Electron Microscope), which uses an electron beam to scan the surface.
However, there is a catch: the "edge effect." When the electron beam hits the edge of a nanostructure, it produces a burst of secondary electrons, causing the edges to look "bloomed" or blurry in the scan. You can't just look at the image and point to the exact edge.
To solve this, researchers use a Model-Based Library (MBL) approach. They simulate millions of possible scan profiles using physics (Monte Carlo methods) and match the real-world scan to the closest simulated one. The problem? These simulations are incredibly slow and computationally "heavy," often requiring supercomputers and weeks of processing time to build a comprehensive library.
Methods
The authors propose a hybrid approach:
- Monte Carlo (MC) Simulation: They use a high-fidelity physical model (incorporating Mott’s scattering cross-sections and Penn’s dielectric function) to generate "ground truth" scan profiles for specific geometries.
- Neural Network (NN) Acceleration: Instead of simulating every possible geometry, they train a fully connected neural network (built on the Keras platform) to learn the relationship between a structure's dimensions (input) and its scan profile (output).
- Architecture: They utilized a 7-layer network. For a single-layer structure, the input is 3 neurons (Top CD, height, sidewall angle). For a double-layer structure, it expands to 6 neurons (adding corner radii and additional angles). The output is a vector representing the intensity at 200–300 scanning points.
Datasets
The study used two primary simulated datasets to validate the model:
- Au (Gold) Single-layer Trapezoid: A dataset of 1,197 profiles. 65% were used for training, and 35% for testing.
- Si (Silicon) Double-layer Trapezoid: A much larger dataset of 22,500 profiles. Only 5,000 (~22%) were used for training, while the remaining 17,500 were used to test the network's ability to generalize.
Key Insights and Ideas
- Physics-to-Pattern Mapping: The neural network effectively "memorizes" the complex physics of electron scattering. Once trained, it can interpolate between simulated points with much higher accuracy than traditional linear interpolation.
- Sparse Training, Dense Results: You don't need a massive library of simulations to start. You can build a "sparse" library with Monte Carlo, then use the neural network to "fill in the gaps" to create a "dense" library for practical use.
- Generalization vs. Extrapolation: The network is incredibly accurate within the range of its training data. While accuracy dips when predicting structures far outside the training range (e.g., much wider lines than it has ever "seen"), it still maintains a respectable level of accuracy for near-range extrapolation.
Key Results
- Massive Speedup: Generating 17,500 scan profiles via Monte Carlo took approximately 3 months on a supercomputer. The trained neural network predicted the same 17,500 profiles in about 1 minute on a standard laptop.
- High Accuracy:
- For the Gold structures, 94.1% of predictions had a Root Mean Square Error (RMSE) of less than 3%.
- For the Silicon structures (within the training range), 94.85% of cases had an RMSE of less than 0.07%.
- Storage Efficiency: A full MBL database can be massive, but a trained neural network model file is less than 20MB, making it much easier to deploy in industrial settings.
Limitations and Future Work
- Out-of-Distribution Accuracy: The authors noted that when predicting dimensions (like Top CD) outside the training range, the error increases. For example, at $T=75$ nm (outside the 30-65 nm training range), the RMSE rose to about 11.5%.
- Complexity Scaling: While the model handled double-layer trapezoids well, real-world semiconductor features can be even more complex (3D gratings, rough surfaces).
- Future Direction: The authors suggest integrating sophisticated image recognition algorithms to infer dimensions directly from measured SEM images by matching them against this AI-accelerated library.
Summary updated