Algorithms for Semiconductor Modeling and Classification Based on Machine Learning

In this dissertation, we propose two algorithms for semiconductor modeling and classification to utilize machine learning techniques. The first algorithm we propose is an automatic approach that significantly reduces the effort and time compared with the manual approach for the pattern matching method for capacitance extraction. Our approach consists of the following steps: 1) generate sample geometries that cover the problem space, 2) for each sample geometry, call a field solver to compute the capacitance value, 3) use unsupervised learning with the field solver values and foundry values as guidance to cluster geometries into pat- terns and derive the capacitance formula, and 4) use supervised learning to train a neural network that will be used to classify any layout geometry into a pattern in the library, which will be used at the extraction time. Experiment results show the proposed approach achieves satisfactory accuracy well above the foundry requirements and runs as fast as traditional pattern matching software. Although our approach is only demonstrated for 2D, it is the first framework using the automatic procedure and machine learning in capacitance extraction. As the interconnect structures and foundry requirements become more complex, our approach may be attractive to EDA vendors and foundries. The second algorithm we propose is an approach that can significantly reduce the workload of manually labelling defects and the heavy computation burden of expanding the ADC, at the same time, maintain high-level classification accuracy and recall rate. Our approach consists of following steps: 1) Analyze the defect dataset to look for simple features that can easily separating each class, 2) build a decision tree to divide the problem based on these features, 3) for each sub-tree, train a simple DNN as the decision rule, 4) for each DNN, tune its classification threshold value for the best classification accuracy and recall rate. Experiment results show that the proposed approach achieves 100% classification accuracy with 93% recall rate well above the foundry requirements. Its learning and updating efficiency are at least sixty times better than the most innovative ADC, single neural network ADC