Access Official Technical PDF

Share Summary:LinkedInEmail

Lithography simulation is essential to semiconductor process development and design verification, but detailed modeling can demand significant computation. As feature sizes shrink and process complexity grows, teams face pressure to accelerate simulation without losing confidence in predicted wafer patterns. LithoGAN explores a different approach: use machine learning to translate a mask pattern directly into a predicted resist pattern, rather than modeling each intermediate stage of the conventional flow.

Why lithography simulation is computationally demanding

In optical lithography, a designed mask pattern is transferred into photoresist on a wafer. A conventional simulation flow first calculates an aerial image using an optical model, then applies a resist model to determine how exposure and development shape the printed pattern. Further processing can be needed to derive contours or critical dimensions.

Rigorous simulation accounts for interacting properties of the optical system, resist materials, mask geometry, and process conditions. That physical detail can make it computationally expensive. Compact models offer faster alternatives, but may trade away some accuracy. Their calibration can also be demanding, particularly at advanced technology nodes.

Machine-learning methods have been used to improve resist modeling. Some predict resist height; others estimate slicing thresholds within an aerial image. These approaches can improve a portion of the workflow, but still depend on optical simulation and additional contour processing. LithoGAN’s central idea is to learn the relationship between the mask and the resulting resist pattern as a whole.

How LithoGAN approaches lithography modeling

LithoGAN formulates lithography modeling as a paired image-translation problem. During training, the model receives examples consisting of a mask image and its corresponding resist pattern, with the latter generated by rigorous simulation. Once trained, it can take a mask pattern as input and predict a resist pattern directly.

The image translation model is a conditional generative adversarial network, or CGAN. It includes two neural networks:

  • Generator: produces a candidate resist-pattern image from the input mask.
  • Discriminator: learns to distinguish a simulated reference pattern from a generated one, helping guide the generator during training.

A pixel-wise loss also encourages the generated image to resemble its paired reference. Together, these training objectives are intended to produce patterns that are both visually plausible and aligned with the corresponding mask input.

Separating shape prediction from center localization

Predicting a pattern’s shape is not sufficient if its position is wrong. The study found that the CGAN could reproduce resist shape well while placing the shape’s center inaccurately. LithoGAN addresses this by dividing the task into two parts: the CGAN predicts the resist shape, while a separate convolutional neural network (CNN) predicts the pattern center. The predicted center is then used to reposition the generated shape in the final output.

This design reflects a practical modeling consideration: geometric fidelity includes both contour shape and location. A useful prediction needs to capture where the pattern is as well as what its boundaries look like.

What the reported evaluation shows

The evaluation used benchmark mask clips at the 10 nm and 7 nm technology nodes. The reference, or “golden,” resist patterns were produced using rigorous lithography simulation. The task focused on the resist pattern of a target contact within each clip, rather than every pattern in a full layout.

To assess similarity to those references, the authors reported several measures. Edge displacement error (EDE) measures the separation between corresponding edges of predicted and reference contours. Pixel accuracy, class accuracy, and mean intersection over union (IoU) assess agreement between the predicted and reference images at the pixel level.

On these test sets, LithoGAN’s mean EDE was 1.08 nm for the 10 nm benchmark and 0.88 nm for the 7 nm benchmark. The reported pixel and class-based metrics also indicated strong image agreement. The comparison method based on optical simulation and machine-learning threshold prediction achieved slightly better accuracy, while LithoGAN’s results were described as within an acceptable range for the evaluated contact patterns.

Runtime was a key outcome. Predicting resist patterns for an entire benchmark dataset with CGAN or LithoGAN took less than 30 seconds on the reported platform. The paper reports approximately 1,800 times lower runtime than rigorous simulation and approximately 190 times lower runtime than the compared threshold-prediction flow. These ratios apply to the study’s benchmarks and implementation; they should not be treated as guaranteed results for different data, tools, or production environments.

Practical considerations for engineering teams

Because LithoGAN learns from paired mask and resist-pattern examples, its usefulness depends on having suitable training data for the modeling task. In the reported work, the reference patterns came from rigorous simulation, so the model learns to reproduce those simulation results. The evaluation therefore measures similarity to the simulated reference, not direct agreement with manufactured wafer data.

The benchmark scope also matters. The experiments covered selected contact clips at two technology nodes, with the target contact centered in each training image. Applying the method to other pattern types, process conditions, or layout contexts would require validation with relevant data. Teams would need to assess whether the model’s output accuracy meets their own requirements and how it fits into existing verification and process-development workflows.

For engineering organizations, the broader lesson is that machine learning can potentially reduce repeated computation by approximating a costly modeling flow after training. Such a model is best considered alongside, rather than automatically in place of, established simulation: validation against appropriate references remains essential, especially when decisions depend on contour placement or critical dimensions.

Conclusion

LithoGAN demonstrates an end-to-end approach to lithography simulation that maps mask patterns directly to predicted resist patterns. Its combination of a CGAN for shape generation and a CNN for center localization addresses two distinct aspects of pattern fidelity. The reported benchmarks show substantial runtime reductions with accuracy close to the rigorous-simulation references. For semiconductor modeling teams, the practical implication is promising but specific: direct prediction may accelerate selected workflows, provided the model is trained and validated for the patterns and conditions where it will be used.

Access Official Technical PDF

About The Author