Yield Modeling and Patterns Prediction Using Data Mining: A Preventive Approach
François Bergeret, PhD, Statistician – Ippon Innovation
Vincent Barec; Alexandre Couvrat, Yield and Defectivity Engineers – Soitec
For complex process manufacturing, yield and defects depend on a huge number of factors. Yield enhancement tools already in production are helping Soitec to detect most process issues (quick reactive approach). This paper presents two new data mining tools created with JMP Scripting Language (JSL) and JMP that are now used for a preventive approach. Yield Model detects the main factors related to yield. These factors are then included in a model to provide accurate predictions. Simulations — based on manufacturing volume, process tools loading and throughput — are available for the user to predict yield and defects variations. Yield Model is a very interactive tool, and its ultimate JSL capabilities will be presented live during the presentation. For the next SOI (Silicon on Insulator) generation, within part thickness uniformity is the main yield driver. Pattern Prediction uses the JMP 9 Neural platform to analyze and predict thickness uniformity with very promising results. The user trains the neural net on a learning sample, and the neural net automatically predicts the pattern for future parts with a good accuracy, assessed on a validation sample. This paper includes advanced JSL features used for the Yield Model and the discovery of the power of neural nets to accurately predict complex patterns. The two tools presented in the paper are preventive tools, helping Soitec to save time and to improve quality.