制造良率学习与失效分析通常从少量经扫描电子显微镜(Scanning Electron Microscope, SEM)或电性分析确认的问题位置开始。为扩大物理复查范围,提出由已知缺陷驱动的多实例版图检索系统 R2-Layout。该系统结合结构感知的检测器-描述子学习、旋转与镜像增强、模板引导候选热力图及局部几何验证,在大版图中检索方向变化或局部裁剪后的相似结构。在包含50张拼接版图和400个真值实例的可控基准上,系统取得了0.962 6的中心命中精度、0.965 0的中心命中召回率和0.970 0的交并比(Intersection over Union, IoU)阈值 0.1(IoU@0.1)召回率。进一步,在6个OpenROAD完整布线设计(routed design)版图上随机构造300个查询,中心命中召回率和IoU@0.1召回率分别达到 0.970 0 和0.973 3,表明模型在复杂完整布线背景下仍能够稳定定位查询结构。真实案例进一步从一处 SEM 确认的桥连缺陷出发,在同一晶圆不同裸片的500 μm版图窗口中发现3个额外待复查位置,后续SEM在其中一处观察到形态相似的桥连缺陷。结果表明,该方法可支持“缺陷定位-相似结构扩展-物理复查”的工程闭环,但版图相似性本身不等同于缺陷判定或成因解释。

Yield learning and failure analysis often begin with a small set of sites confirmed by scanning electron microscopy (SEM) or electrical analysis. This work presents R2-Layout, a known-defect-driven multi-instance retrieval system that expands physical review from one confirmed site to structurally similar layout regions. R2-Layout combines structure-aware detector descriptor learning, rotation and reflection augmentation, a template-guided candidate heatmap, and local geometric verification. On a controlled benchmark comprising 50 stitched layouts and 400 ground-truth instances, it achieves a center-hit precision of 0.962 6, a center-hit recall of 0.965 0, and an IoU@0.1 recall of 0.970 0. Furthermore, on 300 random queries across six OpenROAD realistic routed-design layouts, center-hit recall and IoU@0.1 recall reach 0.970 0 and 0.973 3 respectively, demonstrating stable query repositioning under complex routed back- grounds. In a real failure analysis case, a layout context around one SEM confirmed bridge defect is used to search 500 μm windows from different dies on the same wafer. Three additional sites are returned for review, and subsequent SEM examination identifies a morphologically similar bridge defect at one site. The case demonstrates an engineering loop from defect localization to similar structure expansion and physical confirmation. Layout similarity is used only for candidate discovery and is not interpreted as a defect label or causal conclusion.