Surface roughness is a key measure of defectivity in the semiconductor and display industries because it can degrade device electrical performance and lifetime. The causes of surface roughness process failures can vary, but the lack of an effective model to comprehensively evaluate and analyze surface roughness metrics has made process improvement a huge challenge. This project aims to develop a JMP analysis package that links surface roughness metrics to non-normal distribution models for efficient root cause analysis and process adjustments. The JMP Random Simulation platform was used to generate roughness Z-profile data and classify them into six non-normal distribution models: (1) normal; (2) uniform; (3) heavy tail; (4) right skewed; (5) bimodal; (6) outliers (3%). Five surface roughness metrics (Ra, Rz, Rp, Rv, Rk) and eight statistical parameters describing the distribution were calculated. We compared different clustering methods, including Hierarchical Cluster, K-means Cluster, Normal Mixtures, Cluster Variables, Multivariate Correlation, Principle Components Analysis, and Model Driven Control Charts , to see if they can effectively classify six non-normal distribution models into light-tailed clusters and discrete points, and classify 13 variables into peak clusters, asymmetric clusters, and light-tailed clusters. We effectively distinguished and predicted the relationship between surface roughness measurement indicators and non-normal distribution models. The JMP Surface Roughness Data Mining project can effectively detect the cause of process failure, thereby shortening the process debugging time. In addition, we have also built a strong cross-departmental JMP surface roughness analysis team within the company to establish a database (roughness, original Z- profile, roughness measurement, process adjustment) and develop prediction models for the entire Applied Materials company .
