Proximity Effect Corrections in Electron Beam Lithography Using a Neural Network

NeurIPS 1990

TL;DR

Electron beam lithography suffers from proximity effects that blur fine features. This paper introduces a neural network approach to correct these effects, achieving comparable accuracy to slower iterative methods but with a massive speedup. This opens the door to practical, real-time proximity effect correction for complex chip designs.

What Problem Does This Paper Address?

Electron beam lithography is a powerful technique for creating extremely small features on semiconductor wafers. However, when the electron beam hits the material, electrons scatter, causing the exposed area to be larger than intended. This "proximity effect" blurs fine details, especially when features are close together (interproximity effects). Traditional methods to correct this, like iterative self-consistent calculations, are highly accurate but computationally prohibitive, taking hours or even years to process a full chip. Analytical methods offer speed but sacrifice accuracy. The authors aim to bridge this gap by developing a fast and accurate proximity effect correction method.

Methods

The core of this paper is the application of a neural network to learn the complex mapping required for proximity effect correction. Instead of relying on analytical formulas or slow iterative computations, they train a neural network to perform the correction.

  • Iterative Training Data Generation: They first use a known iterative method to generate a "ground truth" corrected image for a specific test pattern. This corrected image serves as the target output for training the neural network.
  • Neural Network Architecture: The network's architecture is designed to mimic the basic structure of an analytical method (specifically, a weighted sum of neighboring pixel contributions), but the weights are learned adaptively. The network takes a neighborhood of pixels around a target pixel as input and outputs the corrected dose for that pixel.
  • Analog and Digital Implementations: The authors explore both analog and digital hardware implementations of their trained neural network. The analog network uses resistors to represent weights, while the digital version uses a floating-point array processor.
  • Adaptive Learning: The network's weights are adjusted using the delta rule, a common supervised learning algorithm, to minimize the difference between the network's output and the desired output from the iterative method.

Datasets

The primary dataset used is a specific test pattern designed to include challenging features for proximity effect correction. This pattern is 180x180 pixels (22.5 µm x 22.5 µm) with a minimum feature size of 0.25 µm. This pattern was used to generate the training data by applying the iterative correction method. The authors also mention experimental verification on a larger test image (1 mm² with 11,165 shapes and 6.7x10⁷ pixels) under different exposure conditions.

Key Insights and Ideas

  • Neural Networks as Fast Approximators: The paper demonstrates that neural networks can effectively learn and approximate complex, non-linear functions like proximity effect correction, offering a significant speed advantage over traditional methods.
  • Bridging Accuracy and Speed: The neural network approach successfully combines the accuracy of iterative methods with the computational simplicity of analytical methods, a long-sought goal in this field.
  • Hardware Implementation Feasibility: The successful implementation of both analog and digital hardware versions of the neural network highlights the practical viability of this approach for real-time lithography systems.
  • Generalization Capability: The neural network showed good performance on patterns not included in its training set, indicating it learned the underlying physics rather than just memorizing specific examples.

Key Results

  • Accuracy: Both analog and digital neural network implementations achieved an average error of 0.5% compared to the iterative solution, with a maximum error of 9% on individual pixels. This accuracy was maintained on untrained images.
  • Speedup:
    • The iterative method took 4 hours to correct the example pattern on a 386 computer.
  • The software (digital) neural network achieved the same result in 6.5 minutes.
  • The analog hardware network performed the correction in just 13.5 seconds.
  • For a larger experimental test (1 mm²), the digital network took 18 hours, a significant improvement over the estimated "6 years/mm²" for the iterative method.
  • Experimental Verification: SEM micrographs showed significantly improved pattern resolution and reduced feature distortion in corrected patterns compared to uncorrected ones, validating the approach.

Limitations and Future Work

  • I/O Bottleneck: For the analog hardware, the primary limitation was the speed of input/output operations between the network and the computer, rather than the computation speed of the network itself.
  • Generalization to Extreme Conditions: While the network generalized well, the experimental verification on a larger test with significantly greater electron scattering range required a much larger network, suggesting that network size might need to scale with exposure conditions.
  • Future Work: The authors suggest that improved I/O hardware for analog networks could lead to even faster operation. They also note that the flexibility and generality of the digital approach are considerable advantages, and further optimization with coprocessors could lead to even greater speedups, making full chip correction practical.

Summary updated

BibTeX
@inproceedings{1990proximity,
  title = {{Proximity Effect Corrections in Electron Beam Lithography Using a Neural Network}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1990},
}
Figure