Increasingly miniaturized device interconnections constantly compress the space in chip packages, forcing the control of wire loop profiles of great significance. A machine learning (ML) wire profile prediction model combining the finite element (FE) simulation and ML algorithms is developed in this study. First, a flat-looping pattern with three kinks is designed to model the wire profile. Thereafter, the relevant database of the wire profile is established by utilizing FE simulation results. After correlation analysis and features redundant of the data, the relationship between the wire bond profile and the three input parameters are modeled, namely, the first reverse angle theta(1), high-span difference TK, and the span K. Subsequently, the ML model trained and optimized through Support Vector Regression (SVR) and Support Vector Machine (SVM) algorithms. The output parameters of the model are the coordinate values of three kinks, among which K1h, K2l, K2h, K3l, and K3h are trained through the SVR regression model, and K1l is trained through the SVM classification model. After hyperparameter optimization, the final high-accuracy prediction model is obtained. This model demonstration proves a viable approach to predicate wire profiles with various spans and heights. The accuracy of model prediction is verified with four sets of actual wires. The mean absolute percentage error of each predicted wire is 6.3 , 6.4 , 6.3 and 7.1 , respectively. Compared with traditional FE models that take hours for prediction, this ML model can predict a wire profile with high accuracy in a few seconds. The proposed ML model exhibits great potential for wire profile prediction that significantly reduces the wire profile prediction period.
展开▼
