Research On Integrated Circuit Hotspot Detection Method Based On Transfer Learning

Although the integrated circuit technology has been developed below 45 nm,the mainstream lithography manufacturing wavelength is still at 193 nm.The gap between them will deform the lithography pattern and affect the performance of the chip.Even though the various resolution enhancement techniques have been proposed to improve lithography accuracy,there are still many hotspot designs that are difficult to manufacture,so it is very important to detect hotspots before chip production.This paper focuses on the following research work based hotspot detection:1.To solve the problem that the hotspot detection algorithm based on machine learning needs to manually extract features,this paper proposes a hotspot detection method based on transfer learning.Hotspot detection is achieved by transferring a pre-trained deep network,which can obtain excellent detection results under a limited sample size.Three classic convolutional neural network models are transferred and compared,and the most suitable network architecture for hotspot detection is VGG16..2.In order to improve the results of hotspot detection and enhance the generalization ability of the network,the analysis and optimization are mainly made from three aspects of hyperparameters,dataset and network models:The principle and function of the three commonly used hyperparameters of learning rate,momentum and dropout are analyzed,and through the optimization of these three hyperparameters,the network training process is effectively optimized and the generalization ability is improved.To solve the problem of class imbalance in the layout of integrated circuits dataset,this paper compares the detection results of the weighting and sampling methods and scans the class weight ratio,and finds that the introduction of class weighting in the cost function can effectively improve the recall of hotspot detection.Besides when the weight ratio is set to 3:1,the optimal result under the balance of recall and precision can be obtained.Due to the large difference between the ImageNet image which is used in the pre-training network and the layout design image.In order to improve the reduction of the transfer results which caused by this difference,this paper tunes the transfer depth of the model and searches for the optimal transfer depth.When the first 2 to 4 convolutional layers of network are transferred,the optimal detection result can be achieved.3.Based on ICCAD 2012 contest,the original dataset of this article is obtained.At this time,the method in this paper can obtain a recall of 98.1%,a precision of 0.5136,a harmonic mean(F1)of 0.6765 and a F2 value of 0.8313,which can effectively improve the harmonic mean of the existing detection methods.In addition,the method can still obtain good detection results when reducing the training set samples.When there are only 1/4 of the hotspot samples and 1/2 of the non-hotspot samples in the original training set,the recall of 97.62% and 0.6116 F1-score are obtained.4.This paper finds that the original dataset exist the problem that there is a different distribution of training set and test set.It is verified by analyzing the overall distribution of the dataset and the specific analysis of samples.Finally,it is verified that the method of combining the training set and the test set with random sampling can solve the problem of different distributions.