Minimum Max-Rank Method for High-Sigma Yield Analysis

In the field of high-sigma yield analysis, it is common to utilize machine learning-based surrogate models. However, current methodologies struggle with balancing the accuracy of the models with the computational cost. Our findings indicate that the conventional definition of accuracy is inadequate for high-sigma yield analysis, with the order or ranking of predictions being more crucial than relative error. To address this, we introduce the concept of Max-Rank as a Figure of Merit(FOM) for evaluating different models, and present the Minimum Max-Rank method as a means of producing models that result in faster, more accurate yield analysis. Finally, we compared our algorithm with commercial tools such as Cadence, WICkeD, varman, and others, and obtained satisfactory accuracy.