• 목록

Semiconductor manufacturing is a multistage process consisting of hundreds of steps, in which precise process control is essential. Metrology is performed for this purpose; however, measuring every wafer is impractical due to time and cost constraints. Based on a literature review, we find that existing VM models have generally focused on fitting algorithms to data, without exploiting the common structural characteristics inherent in the data itself. In particular, two challenges remain underexplored: lot-level variation and input uncertainty. These factors cause models to rely on unreliable information, degrading both generalization and interpretability. To mitigate these issues, we propose a new VM model that combines an attention mechanism with an efficient ensemble. Each input variable is represented as a learnable token, allowing attention to suppress contaminated inputs and capture variable dependencies, while the ensemble reduces prediction variance induced by lot-level variation. Experiments on real-world semiconductor data demonstrate that the proposed model attain robust prediction, achieving the smallest gap between training and test performance among the baselines and suggesting its potential for production deployment.