Temporal-enhanced semiconductor yield prediction by combining equipment usage data and accumulated cycle time data

In the semiconductor industry, yield serves as a critical performance indicator for assessing fabrication efficiency. Accurate yield prediction enables proactive quality control, cost estimation for fabrication, and risk mitigation arising from yield reductions. Initially, each piece of equipment’s characteristics were analyzed independently to determine its impact on yield for prediction purposes. Recent advancements in fabrication data storage have enabled data-driven yield prediction approaches based on analysis of equipment usage data. However, these yield prediction approaches demonstrate limitations in capturing temporal aspects that influence yield through cumulative environmental exposure during fabrication. The present research established a temporal-enhanced yield prediction approach that combines equipment usage data and Accumulated Cycle Time (ACT) data extracted from production log data. ACT data provides essential temporal information based on cycle times, reflecting cumulative chemical alterations across process steps. After concatenating these two distinct datasets, regression models are applied following model-agnostic principles to quantify relationships between equipment usage data and ACT data for yield prediction. Experimental validation using actual production log data from a leading manufacturer demonstrated significantly improved yield prediction performance, averaging a 15.38% improvement over conventional approaches that relied solely on equipment usage data. Furthermore, SHapley Additive exPlanations (SHAP) analysis revealed critical temporal factors that influence yield prediction, particularly in wet-etching processes. The improvements highlighted the advantages of combining equipment condition aspects and temporal aspects to enhance proactive quality control and reduce fabrication costs.