Access Official Technical PDF

Share Summary:LinkedInEmail

Wafer SEM image generation offers a way to address a persistent challenge in semiconductor manufacturing: obtaining enough labeled scanning electron microscope images to develop and evaluate defect-analysis methods. SEM imaging and annotation can be costly and time-consuming, while image noise and variation in manufacturing processes make representative datasets difficult to assemble. A conditional generative adversarial network (cGAN) approach described by Du and Shi uses circuit layout images to generate corresponding wafer SEM images, with the goal of supplementing limited image collections.

For automation and process teams, the practical interest is not simply whether a generated image looks plausible. Synthetic data must reproduce useful characteristics of real images if it is to support work such as training defect classifiers. The proposed method addresses image detail and distribution matching through gradient information, two discriminators operating at different resolutions, and a combination of training losses.

Why wafer SEM image data is difficult to scale

Integrated circuit manufacturing involves many process steps, and defects can affect device function, reliability and production yield. SEM images are used to inspect and assess these defects, but relying on manual review can become labor-intensive when production generates large volumes of data. Machine-learning approaches can help automate pattern recognition, yet they typically need substantial image datasets for training.

That creates a data bottleneck. Collecting SEM images and labeling them takes time and resources, and the images themselves may contain noise or reflect changes in manufacturing conditions. Conventional image simulators have also been used, but the paper notes that tools not designed specifically to generate SEM samples may produce images of limited quality. Deep-learning generators provide another option, although standard GANs can suffer from mode collapse and produce problems such as blurred circuit edges or uneven background noise.

How the conditional GAN generates wafer SEM images

The method builds on pix2pix, a supervised image-to-image translation framework. Its input is a circuit layout image, together with random noise, and its output is a synthetic wafer SEM image. The paired layout and SEM data give the model a target relationship to learn: how a particular layout corresponds to the appearance of an SEM image.

The generator uses a U-Net architecture. Its skip connections pass lower-level information across the network, helping retain details from the input while producing the translated image. The model’s design also includes two discriminators. Each assesses whether images are real or generated, but they operate at different resolutions: one receives the image at its original scale, while the other receives a down-sampled version. This gives the training process feedback about image characteristics at more than one scale.

Using edges and loss functions to guide training

A Sobel operator calculates gradient information from the images. In practical terms, gradients emphasize changes in intensity, including edges and boundaries that can be important in circuit imagery. The gradient image is supplied as part of the discriminator’s input, providing an additional signal for assessing whether generated structures resemble those in real SEM images.

The training objective combines adversarial losses with smooth L1 loss. The adversarial component encourages the generator to produce images that the discriminators consider similar to real samples. Smooth L1 loss contributes a direct comparison between generated and target images. The authors also use Wasserstein distance in the discriminator training to address unstable GAN convergence. Together, these choices are intended to improve the match between generated and real image distributions while retaining image-level similarity.

Dataset and evaluation approach

The experiments used circuit layouts derived from GDS files and three wafer SEM datasets collected from a fabrication facility. The layout images were RGB, with three layers, and measured 256 by 256 pixels. The SEM images were grayscale at the same resolution, with 8,000 images in each dataset. For each dataset, 80% of the images were assigned to training and the remainder to testing.

The researchers compared their method with pix2pix, pix2pixHD and DualGAN. They used peak signal-to-noise ratio (PSNR) to compare pixel values and a 1-nearest-neighbor (1-NN) classifier to assess how distinguishable generated images were from real ones. In the reported results across the three datasets, the proposed method recorded PSNR values from 29.53 to 29.84 dB. Its 1-NN scores ranged from 0.623 to 0.668, while the comparison models were reported at 1.0.

In the paper’s interpretation, higher PSNR indicates closer pixel-level similarity, while a 1-NN score nearer 0.5 indicates that the classifier has more difficulty separating real and generated images. These metrics offer different views of image quality; neither should be treated on its own as proof that synthetic images will improve defect-detection performance in a production setting. The reported results support the authors’ conclusion that their model better matched the tested SEM image data than the compared methods.

What the approach means for semiconductor operations

For teams developing automated inspection or defect-classification systems, conditional SEM image synthesis may help explore ways to supplement scarce training data. The layout-conditioned approach is especially relevant where the desired output is linked to a known circuit design rather than an unconstrained image sample. Multi-resolution discrimination and edge guidance also reflect an important engineering consideration: both overall image appearance and fine structural details matter when generating inspection imagery.

However, synthetic images should be treated as a complement to, not a replacement for, real SEM data. The experiments evaluate image similarity on specific datasets; they do not establish performance for every fab, process variation or defect category. The authors identify generating images with specified defects and improving the match between generated and real distributions as areas for future work.

Conclusion

Wafer SEM image generation with a conditional GAN offers a data-synthesis approach for semiconductor defect-analysis workflows. By translating circuit layouts into SEM-like images and combining Sobel gradient guidance, discriminators at multiple resolutions, Wasserstein distance and smooth L1 loss, the method targets both structural detail and distribution similarity. Its reported results show promise on the evaluated datasets, while deployment decisions still require validation against the real images and defect-analysis requirements of the intended application.

Access Official Technical PDF

About The Author