Learning Calibrated Visual Representations for Transformer Layout Hotspot Prediction via DINOv2 Adapters

Abstract

Hotspot predictors for advanced-node transformer layout verification should be sensitive to weak lithographic risk but not unstable under process variation. This paper introduces a DINOv2-Adapter model with uncertainty-calibrated feature fusion for transformer hotspot prediction in dense mask layouts. The method freezes a self-supervised visual backbone, adds lightweight domain adapters, and integrates shallow edge-density cues with semantic patch tokens via a temperature-scaled uncertainty gate. A curated layout benchmark of 62,400 clips, four process-window corners and six metal-pattern density bands was used for the experiments. Compared with the ResNet-50 baseline, the proposed model increased the F1-score from 91.4% to 96.2%, improved hotspot recalls from 90.8% to 97.1%, and reduced the expected calibration error from 8.7% to 2.9%. Under 10% label noise, the model achieved a 94.3% F1 score; the strongest non-calibrated transformer variant dropped to 91.8%. Ablation results show that adapter tuning increased F1 by 2.1 percentage points, uncertainty-gated fusion added 1.4 points, and calibration loss reduced overconfident false alarms by 31.6%. Based on the above results, transferrable representation learning can be combined with calibrated fusion to construct a practical engineering predictor for transformer-oriented hotspot screening. The main contribution is the integration of adapter-based foundation-model transfer, branch-wise uncertainty estimation, and calibrated hotspot ranking within one deployable prediction pipeline.