Gen-Fab: A Variation-Aware Generative Model for Predicting Fabrication Variations in Nanophotonic Devices

Abstract

Silicon photonic devices often exhibit fabrication-induced variations such as over-etching, underetching, and corner rounding, which can significantly alter device performance. These variations are non-uniform and are influenced by feature size and shape. Accurate digital twins are therefore needed to predict the range of possible fabricated outcomes for a given design. In this paper, we introduce Gen-Fab, a conditional generative adversarial network (cGAN) based on Pix2Pix to predict and model uncertainty in photonic fabrication outcomes. The proposed method takes a design layout (in GDS format) as input and produces diverse high-resolution predictions similar to scanning electron microscope (SEM) images of fabricated devices, capturing the range of process variations at the nanometer scale. To enable one-to-many mapping, we inject a latent noise vector at the model bottleneck. We compare Gen-Fab against three baselines: (1) a deterministic U-Net predictor, (2) an inference-time Monte Carlo Dropout U-Net, and (3) an ensemble of varied U-Nets. Evaluations on an out-of-distribution dataset of fabricated photonic test structures demonstrate that Gen-Fab outperforms all baselines in both accuracy and uncertainty modeling. An additional distribution shift analysis further confirms its strong generalization to unseen fabrication geometries. Gen-Fab achieves the highest intersection-over-union (IoU) score of 89.8%, outperforming the deterministic U-Net (85.3%), the MC-Dropout U-Net (83.4%), and varying U-Nets (85.8%). It also better aligns with the distribution of real fabrication outcomes, achieving lower Kullback-Leibler divergence and Wasserstein distance.

Figures (12)

• Figure 1: Overlay of SEM images from repeated fabrications of the same photonic cross design. Although each fabricated structure used the same GDS file layout, visible differences emerge in edge sharpness and arm geometry, indicated in red.
• Figure 2: IoU similarity heatmap of SEM images from four fabricated crosses. Each cell shows the IoU score between pairs of SEM images from four devices fabricated from the same layout.
• Figure 3: Overview of the Gen-Fab architecture with a noise-injected generator.Top: Generator. The model receives a GDS layout and a latent noise vector injected into the bottleneck, enabling it to produce diverse SEM outputs from the same input. Bottom: Discriminator. The PatchGAN discriminator takes the input GDS and corresponding SEM (real or generated) and evaluates $70\times70$ pixel patches to ensure outputs are locally indistinguishable from real SEMs.
• Figure 4: Overview of Gen-Fab’s generation and evaluation pipeline.Top: Generation Process. The Gen-Fab generator takes a GDS input and multiple noise vectors to produce a diverse SEM predictions, while the discriminator is discarded after training. Bottom: Evaluation Process. Generated SEMs are compared with real SEMs using pixel-level metrics (IoU, variance map) and distribution-level metrics (KL-D and W-D).
• Figure 5: Paired examples of GDS layouts and their fabricated SEM used for training. (a) Input GDS design (ground-truth). (b) SEM of the fabricated structures, showing deviations due to the fabrication process.