AIML and Data Analytics for Semiconductor Manufacturing

Conventional analytics has limited knowledge of cross-modal process interactions, which contributes to low accuracy in fault identification. This is due to their incapacity to handle the massive and diverse datasets produced by contemporary semiconductor fabrication facilities. In advanced manufacturing, yield, dependability, and production costs are directly affected by both undiscovered defects and tool abnormalities. The purpose of this research was to develop a unified multimodal framework for accurate defect and anomaly detection. An AI framework that could combine wafer pictures, time-series telemetry, and analyze sensor data. We evaluated baseline, hybrid, and a unique Perceiver-GraphFusion (PGF) architecture in an experimental investigation using publicly accessible wafer-map, SECOM sensor, and UCR time-series datasets. The PGF model allows for joint spatial-temporal-structural representation learning across heterogeneous modalities by combining graph-based lot-tool reasoning with cross-modal perceiver attention. The proposed PGF model significantly outperformed unimodal and hybrid baselines, achieving a Macro-F1 of 0.90, AUROC of 0.96, and AUPR of 0.89, compared to 0.78 (ResNet-18) and 0.65 (XGBoost). Case analysis revealed increased prediction confidence following telemetry cleaning (0.54 → 0.81), and ablation studies validated the significance of each modality. These results show that multimodal, graph-aware learning is a strong way to improve semiconductor analytics by making it easier to spot defects and helping people make better judgments about how to govern processes.

Перевод пока недоступен