Distortion removal in wafer surface interference fringe patterns based on large language models

The accuracy of wafer surface inspection via interference fringes is often degraded by optical aberrations, speckle noise, and unmodeled distortions. Existing approaches for distortion removal lack generalization across diverse distortions, limiting their practical deployment. This paper proposes a novel method that adopts a large language model (LLM) to robustly restore distorted fringe patterns. Our approach employs a cross-modal guidance strategy that deeply aligns fringe features with the semantic latent space of an LLM, guiding the LLM to better interpret fine-grained structures of dense fringe patterns. Furthermore, a fringe-fidelity constrained mechanism is incorporated to ensure high-quality restoration of distorted fringe patterns. To address data scarcity, we generated a large-scale synthetic dataset using physics-based modeling of optical aberrations and speckle noise, covering a wide range of distortion types and their combinations. Experiments demonstrate our method's superior performance in both controlled settings with various optical distortions and noise, and in challenging real-world scenarios. It shows strong generalization to unseen distortions, thereby offering a reliable preprocessing solution for high-precision wafer surface profilometry.